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 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 fio_unused 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 static void sig_ill(int fio_unused sig)
 129 {
 130         if (!threads)
 131                 return;
 132
 133         log_err("fio: illegal instruction. your cpu does not support "
 134                 "the sse4.2 instruction for crc32c\n");
 135         terminate_threads(TERMINATE_ALL);
 136         exit(4);
 137 }
 138
 139 static void set_sig_handlers(void)
 140 {
 141         struct sigaction act;
 142
 143         memset(&act, 0, sizeof(act));
 144         act.sa_handler = sig_alrm;
 145         act.sa_flags = SA_RESTART;
 146         sigaction(SIGALRM, &act, NULL);
 147
 148         memset(&act, 0, sizeof(act));
 149         act.sa_handler = sig_int;
 150         act.sa_flags = SA_RESTART;
 151         sigaction(SIGINT, &act, NULL);
 152
 153         memset(&act, 0, sizeof(act));
 154         act.sa_handler = sig_ill;
 155         act.sa_flags = SA_RESTART;
 156         sigaction(SIGILL, &act, NULL);
 157 }
 158
 159 /*
 160  * Check if we are above the minimum rate given.
 161  */
 162 static int check_min_rate(struct thread_data *td, struct timeval *now)
 163 {
 164         unsigned long long bytes = 0;
 165         unsigned long iops = 0;
 166         unsigned long spent;
 167         unsigned long rate;
 168
 169         /*
 170          * No minimum rate set, always ok
 171          */
 172         if (!td->o.ratemin && !td->o.rate_iops_min)
 173                 return 0;
 174
 175         /*
 176          * allow a 2 second settle period in the beginning
 177          */
 178         if (mtime_since(&td->start, now) < 2000)
 179                 return 0;
 180
 181         if (td_read(td)) {
 182                 iops += td->io_blocks[DDIR_READ];
 183                 bytes += td->this_io_bytes[DDIR_READ];
 184         }
 185         if (td_write(td)) {
 186                 iops += td->io_blocks[DDIR_WRITE];
 187                 bytes += td->this_io_bytes[DDIR_WRITE];
 188         }
 189
 190         /*
 191          * if rate blocks is set, sample is running
 192          */
 193         if (td->rate_bytes || td->rate_blocks) {
 194                 spent = mtime_since(&td->lastrate, now);
 195                 if (spent < td->o.ratecycle)
 196                         return 0;
 197
 198                 if (td->o.rate) {
 199                         /*
 200                          * check bandwidth specified rate
 201                          */
 202                         if (bytes < td->rate_bytes) {
 203                                 log_err("%s: min rate %u not met\n", td->o.name,
 204                                                                 td->o.ratemin);
 205                                 return 1;
 206                         } else {
 207                                 rate = (bytes - td->rate_bytes) / spent;
 208                                 if (rate < td->o.ratemin ||
 209                                     bytes < td->rate_bytes) {
 210                                         log_err("%s: min rate %u not met, got"
 211                                                 " %luKiB/sec\n", td->o.name,
 212                                                         td->o.ratemin, rate);
 213                                         return 1;
 214                                 }
 215                         }
 216                 } else {
 217                         /*
 218                          * checks iops specified rate
 219                          */
 220                         if (iops < td->o.rate_iops) {
 221                                 log_err("%s: min iops rate %u not met\n",
 222                                                 td->o.name, td->o.rate_iops);
 223                                 return 1;
 224                         } else {
 225                                 rate = (iops - td->rate_blocks) / spent;
 226                                 if (rate < td->o.rate_iops_min ||
 227                                     iops < td->rate_blocks) {
 228                                         log_err("%s: min iops rate %u not met,"
 229                                                 " got %lu\n", td->o.name,
 230                                                         td->o.rate_iops_min,
 231                                                         rate);
 232                                 }
 233                         }
 234                 }
 235         }
 236
 237         td->rate_bytes = bytes;
 238         td->rate_blocks = iops;
 239         memcpy(&td->lastrate, now, sizeof(*now));
 240         return 0;
 241 }
 242
 243 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 244 {
 245         if (!td->o.timeout)
 246                 return 0;
 247         if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
 248                 return 1;
 249
 250         return 0;
 251 }
 252
 253 /*
 254  * When job exits, we can cancel the in-flight IO if we are using async
 255  * io. Attempt to do so.
 256  */
 257 static void cleanup_pending_aio(struct thread_data *td)
 258 {
 259         struct flist_head *entry, *n;
 260         struct io_u *io_u;
 261         int r;
 262
 263         /*
 264          * get immediately available events, if any
 265          */
 266         r = io_u_queued_complete(td, 0);
 267         if (r < 0)
 268                 return;
 269
 270         /*
 271          * now cancel remaining active events
 272          */
 273         if (td->io_ops->cancel) {
 274                 flist_for_each_safe(entry, n, &td->io_u_busylist) {
 275                         io_u = flist_entry(entry, struct io_u, list);
 276
 277                         /*
 278                          * if the io_u isn't in flight, then that generally
 279                          * means someone leaked an io_u. complain but fix
 280                          * it up, so we don't stall here.
 281                          */
 282                         if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
 283                                 log_err("fio: non-busy IO on busy list\n");
 284                                 put_io_u(td, io_u);
 285                         } else {
 286                                 r = td->io_ops->cancel(td, io_u);
 287                                 if (!r)
 288                                         put_io_u(td, io_u);
 289                         }
 290                 }
 291         }
 292
 293         if (td->cur_depth)
 294                 r = io_u_queued_complete(td, td->cur_depth);
 295 }
 296
 297 /*
 298  * Helper to handle the final sync of a file. Works just like the normal
 299  * io path, just does everything sync.
 300  */
 301 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 302 {
 303         struct io_u *io_u = __get_io_u(td);
 304         int ret;
 305
 306         if (!io_u)
 307                 return 1;
 308
 309         io_u->ddir = DDIR_SYNC;
 310         io_u->file = f;
 311
 312         if (td_io_prep(td, io_u)) {
 313                 put_io_u(td, io_u);
 314                 return 1;
 315         }
 316
 317 requeue:
 318         ret = td_io_queue(td, io_u);
 319         if (ret < 0) {
 320                 td_verror(td, io_u->error, "td_io_queue");
 321                 put_io_u(td, io_u);
 322                 return 1;
 323         } else if (ret == FIO_Q_QUEUED) {
 324                 if (io_u_queued_complete(td, 1) < 0)
 325                         return 1;
 326         } else if (ret == FIO_Q_COMPLETED) {
 327                 if (io_u->error) {
 328                         td_verror(td, io_u->error, "td_io_queue");
 329                         return 1;
 330                 }
 331
 332                 if (io_u_sync_complete(td, io_u) < 0)
 333                         return 1;
 334         } else if (ret == FIO_Q_BUSY) {
 335                 if (td_io_commit(td))
 336                         return 1;
 337                 goto requeue;
 338         }
 339
 340         return 0;
 341 }
 342
 343 /*
 344  * The main verify engine. Runs over the writes we previously submitted,
 345  * reads the blocks back in, and checks the crc/md5 of the data.
 346  */
 347 static void do_verify(struct thread_data *td)
 348 {
 349         struct fio_file *f;
 350         struct io_u *io_u;
 351         int ret, min_events;
 352         unsigned int i;
 353
 354         /*
 355          * sync io first and invalidate cache, to make sure we really
 356          * read from disk.
 357          */
 358         for_each_file(td, f, i) {
 359                 if (!(f->flags & FIO_FILE_OPEN))
 360                         continue;
 361                 if (fio_io_sync(td, f))
 362                         break;
 363                 if (file_invalidate_cache(td, f))
 364                         break;
 365         }
 366
 367         if (td->error)
 368                 return;
 369
 370         td_set_runstate(td, TD_VERIFYING);
 371
 372         io_u = NULL;
 373         while (!td->terminate) {
 374                 int ret2, full;
 375
 376                 io_u = __get_io_u(td);
 377                 if (!io_u)
 378                         break;
 379
 380                 if (runtime_exceeded(td, &io_u->start_time)) {
 381                         put_io_u(td, io_u);
 382                         td->terminate = 1;
 383                         break;
 384                 }
 385
 386                 if (get_next_verify(td, io_u)) {
 387                         put_io_u(td, io_u);
 388                         break;
 389                 }
 390
 391                 if (td_io_prep(td, io_u)) {
 392                         put_io_u(td, io_u);
 393                         break;
 394                 }
 395
 396                 io_u->end_io = verify_io_u;
 397
 398                 ret = td_io_queue(td, io_u);
 399                 switch (ret) {
 400                 case FIO_Q_COMPLETED:
 401                         if (io_u->error)
 402                                 ret = -io_u->error;
 403                         else if (io_u->resid) {
 404                                 int bytes = io_u->xfer_buflen - io_u->resid;
 405                                 struct fio_file *f = io_u->file;
 406
 407                                 /*
 408                                  * zero read, fail
 409                                  */
 410                                 if (!bytes) {
 411                                         td_verror(td, EIO, "full resid");
 412                                         put_io_u(td, io_u);
 413                                         break;
 414                                 }
 415
 416                                 io_u->xfer_buflen = io_u->resid;
 417                                 io_u->xfer_buf += bytes;
 418                                 io_u->offset += bytes;
 419
 420                                 td->ts.short_io_u[io_u->ddir]++;
 421
 422                                 if (io_u->offset == f->real_file_size)
 423                                         goto sync_done;
 424
 425                                 requeue_io_u(td, &io_u);
 426                         } else {
 427 sync_done:
 428                                 ret = io_u_sync_complete(td, io_u);
 429                                 if (ret < 0)
 430                                         break;
 431                         }
 432                         continue;
 433                 case FIO_Q_QUEUED:
 434                         break;
 435                 case FIO_Q_BUSY:
 436                         requeue_io_u(td, &io_u);
 437                         ret2 = td_io_commit(td);
 438                         if (ret2 < 0)
 439                                 ret = ret2;
 440                         break;
 441                 default:
 442                         assert(ret < 0);
 443                         td_verror(td, -ret, "td_io_queue");
 444                         break;
 445                 }
 446
 447                 if (ret < 0 || td->error)
 448                         break;
 449
 450                 /*
 451                  * if we can queue more, do so. but check if there are
 452                  * completed io_u's first.
 453                  */
 454                 full = queue_full(td) || ret == FIO_Q_BUSY;
 455                 if (full || !td->o.iodepth_batch_complete) {
 456                         min_events = td->o.iodepth_batch_complete;
 457                         if (full && !min_events)
 458                                 min_events = 1;
 459
 460                         do {
 461                                 /*
 462                                  * Reap required number of io units, if any,
 463                                  * and do the verification on them through
 464                                  * the callback handler
 465                                  */
 466                                 if (io_u_queued_complete(td, min_events) < 0) {
 467                                         ret = -1;
 468                                         break;
 469                                 }
 470                         } while (full && (td->cur_depth > td->o.iodepth_low));
 471                 }
 472                 if (ret < 0)
 473                         break;
 474         }
 475
 476         if (!td->error) {
 477                 min_events = td->cur_depth;
 478
 479                 if (min_events)
 480                         ret = io_u_queued_complete(td, min_events);
 481         } else
 482                 cleanup_pending_aio(td);
 483
 484         td_set_runstate(td, TD_RUNNING);
 485 }
 486
 487 /*
 488  * Main IO worker function. It retrieves io_u's to process and queues
 489  * and reaps them, checking for rate and errors along the way.
 490  */
 491 static void do_io(struct thread_data *td)
 492 {
 493         struct timeval s;
 494         unsigned long usec;
 495         unsigned int i;
 496         int ret = 0;
 497
 498         if (in_ramp_time(td))
 499                 td_set_runstate(td, TD_RAMP);
 500         else
 501                 td_set_runstate(td, TD_RUNNING);
 502
 503         while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) {
 504                 struct timeval comp_time;
 505                 long bytes_done = 0;
 506                 int min_evts = 0;
 507                 struct io_u *io_u;
 508                 int ret2, full;
 509
 510                 if (td->terminate)
 511                         break;
 512
 513                 io_u = get_io_u(td);
 514                 if (!io_u)
 515                         break;
 516
 517                 memcpy(&s, &io_u->start_time, sizeof(s));
 518
 519                 if (runtime_exceeded(td, &s)) {
 520                         put_io_u(td, io_u);
 521                         td->terminate = 1;
 522                         break;
 523                 }
 524
 525                 /*
 526                  * Add verification end_io handler, if asked to verify
 527                  * a previously written file.
 528                  */
 529                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ) {
 530                         io_u->end_io = verify_io_u;
 531                         td_set_runstate(td, TD_VERIFYING);
 532                 } else if (in_ramp_time(td))
 533                         td_set_runstate(td, TD_RAMP);
 534                 else
 535                         td_set_runstate(td, TD_RUNNING);
 536
 537                 ret = td_io_queue(td, io_u);
 538                 switch (ret) {
 539                 case FIO_Q_COMPLETED:
 540                         if (io_u->error)
 541                                 ret = -io_u->error;
 542                         else if (io_u->resid) {
 543                                 int bytes = io_u->xfer_buflen - io_u->resid;
 544                                 struct fio_file *f = io_u->file;
 545
 546                                 /*
 547                                  * zero read, fail
 548                                  */
 549                                 if (!bytes) {
 550                                         td_verror(td, EIO, "full resid");
 551                                         put_io_u(td, io_u);
 552                                         break;
 553                                 }
 554
 555                                 io_u->xfer_buflen = io_u->resid;
 556                                 io_u->xfer_buf += bytes;
 557                                 io_u->offset += bytes;
 558
 559                                 td->ts.short_io_u[io_u->ddir]++;
 560
 561                                 if (io_u->offset == f->real_file_size)
 562                                         goto sync_done;
 563
 564                                 requeue_io_u(td, &io_u);
 565                         } else {
 566 sync_done:
 567                                 fio_gettime(&comp_time, NULL);
 568                                 bytes_done = io_u_sync_complete(td, io_u);
 569                                 if (bytes_done < 0)
 570                                         ret = bytes_done;
 571                         }
 572                         break;
 573                 case FIO_Q_QUEUED:
 574                         /*
 575                          * if the engine doesn't have a commit hook,
 576                          * the io_u is really queued. if it does have such
 577                          * a hook, it has to call io_u_queued() itself.
 578                          */
 579                         if (td->io_ops->commit == NULL)
 580                                 io_u_queued(td, io_u);
 581                         break;
 582                 case FIO_Q_BUSY:
 583                         requeue_io_u(td, &io_u);
 584                         ret2 = td_io_commit(td);
 585                         if (ret2 < 0)
 586                                 ret = ret2;
 587                         break;
 588                 default:
 589                         assert(ret < 0);
 590                         put_io_u(td, io_u);
 591                         break;
 592                 }
 593
 594                 if (ret < 0 || td->error)
 595                         break;
 596
 597                 /*
 598                  * See if we need to complete some commands
 599                  */
 600                 full = queue_full(td) || ret == FIO_Q_BUSY;
 601                 if (full || !td->o.iodepth_batch_complete) {
 602                         min_evts = td->o.iodepth_batch_complete;
 603                         if (full && !min_evts)
 604                                 min_evts = 1;
 605
 606                         fio_gettime(&comp_time, NULL);
 607
 608                         do {
 609                                 ret = io_u_queued_complete(td, min_evts);
 610                                 if (ret <= 0)
 611                                         break;
 612
 613                                 bytes_done += ret;
 614                         } while (full && (td->cur_depth > td->o.iodepth_low));
 615                 }
 616
 617                 if (ret < 0)
 618                         break;
 619                 if (!bytes_done)
 620                         continue;
 621
 622                 /*
 623                  * the rate is batched for now, it should work for batches
 624                  * of completions except the very first one which may look
 625                  * a little bursty
 626                  */
 627                 if (!in_ramp_time(td)) {
 628                         usec = utime_since(&s, &comp_time);
 629
 630                         rate_throttle(td, usec, bytes_done);
 631
 632                         if (check_min_rate(td, &comp_time)) {
 633                                 if (exitall_on_terminate)
 634                                         terminate_threads(td->groupid);
 635                                 td_verror(td, EIO, "check_min_rate");
 636                                 break;
 637                         }
 638                 }
 639
 640                 if (td->o.thinktime) {
 641                         unsigned long long b;
 642
 643                         b = td->io_blocks[0] + td->io_blocks[1];
 644                         if (!(b % td->o.thinktime_blocks)) {
 645                                 int left;
 646
 647                                 if (td->o.thinktime_spin)
 648                                         __usec_sleep(td->o.thinktime_spin);
 649
 650                                 left = td->o.thinktime - td->o.thinktime_spin;
 651                                 if (left)
 652                                         usec_sleep(td, left);
 653                         }
 654                 }
 655         }
 656
 657         if (td->o.fill_device && td->error == ENOSPC) {
 658                 td->error = 0;
 659                 td->terminate = 1;
 660         }
 661         if (!td->error) {
 662                 struct fio_file *f;
 663
 664                 i = td->cur_depth;
 665                 if (i)
 666                         ret = io_u_queued_complete(td, i);
 667
 668                 if (should_fsync(td) && td->o.end_fsync) {
 669                         td_set_runstate(td, TD_FSYNCING);
 670
 671                         for_each_file(td, f, i) {
 672                                 if (!(f->flags & FIO_FILE_OPEN))
 673                                         continue;
 674                                 fio_io_sync(td, f);
 675                         }
 676                 }
 677         } else
 678                 cleanup_pending_aio(td);
 679
 680         /*
 681          * stop job if we failed doing any IO
 682          */
 683         if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
 684                 td->done = 1;
 685 }
 686
 687 static void cleanup_io_u(struct thread_data *td)
 688 {
 689         struct flist_head *entry, *n;
 690         struct io_u *io_u;
 691
 692         flist_for_each_safe(entry, n, &td->io_u_freelist) {
 693                 io_u = flist_entry(entry, struct io_u, list);
 694
 695                 flist_del(&io_u->list);
 696                 free(io_u);
 697         }
 698
 699         free_io_mem(td);
 700 }
 701
 702 static int init_io_u(struct thread_data *td)
 703 {
 704         struct io_u *io_u;
 705         unsigned int max_bs;
 706         int i, max_units;
 707         char *p;
 708
 709         max_units = td->o.iodepth;
 710         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
 711         td->orig_buffer_size = (unsigned long long) max_bs
 712                                         * (unsigned long long) max_units;
 713
 714         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
 715                 unsigned long bs;
 716
 717                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
 718                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
 719         }
 720
 721         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
 722                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
 723                 return 1;
 724         }
 725
 726         if (allocate_io_mem(td))
 727                 return 1;
 728
 729         if (td->o.odirect)
 730                 p = ALIGN(td->orig_buffer);
 731         else
 732                 p = td->orig_buffer;
 733
 734         for (i = 0; i < max_units; i++) {
 735                 if (td->terminate)
 736                         return 1;
 737                 io_u = malloc(sizeof(*io_u));
 738                 memset(io_u, 0, sizeof(*io_u));
 739                 INIT_FLIST_HEAD(&io_u->list);
 740
 741                 if (!(td->io_ops->flags & FIO_NOIO)) {
 742                         io_u->buf = p + max_bs * i;
 743
 744                         if (td_write(td) && !td->o.refill_buffers)
 745                                 io_u_fill_buffer(td, io_u, max_bs);
 746                 }
 747
 748                 io_u->index = i;
 749                 io_u->flags = IO_U_F_FREE;
 750                 flist_add(&io_u->list, &td->io_u_freelist);
 751         }
 752
 753         return 0;
 754 }
 755
 756 static int switch_ioscheduler(struct thread_data *td)
 757 {
 758         char tmp[256], tmp2[128];
 759         FILE *f;
 760         int ret;
 761
 762         if (td->io_ops->flags & FIO_DISKLESSIO)
 763                 return 0;
 764
 765         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 766
 767         f = fopen(tmp, "r+");
 768         if (!f) {
 769                 if (errno == ENOENT) {
 770                         log_err("fio: os or kernel doesn't support IO scheduler"
 771                                 " switching\n");
 772                         return 0;
 773                 }
 774                 td_verror(td, errno, "fopen iosched");
 775                 return 1;
 776         }
 777
 778         /*
 779          * Set io scheduler.
 780          */
 781         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
 782         if (ferror(f) || ret != 1) {
 783                 td_verror(td, errno, "fwrite");
 784                 fclose(f);
 785                 return 1;
 786         }
 787
 788         rewind(f);
 789
 790         /*
 791          * Read back and check that the selected scheduler is now the default.
 792          */
 793         ret = fread(tmp, 1, sizeof(tmp), f);
 794         if (ferror(f) || ret < 0) {
 795                 td_verror(td, errno, "fread");
 796                 fclose(f);
 797                 return 1;
 798         }
 799
 800         sprintf(tmp2, "[%s]", td->o.ioscheduler);
 801         if (!strstr(tmp, tmp2)) {
 802                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
 803                 td_verror(td, EINVAL, "iosched_switch");
 804                 fclose(f);
 805                 return 1;
 806         }
 807
 808         fclose(f);
 809         return 0;
 810 }
 811
 812 static int keep_running(struct thread_data *td)
 813 {
 814         unsigned long long io_done;
 815
 816         if (td->done)
 817                 return 0;
 818         if (td->o.time_based)
 819                 return 1;
 820         if (td->o.loops) {
 821                 td->o.loops--;
 822                 return 1;
 823         }
 824
 825         io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
 826                         + td->io_skip_bytes;
 827         if (io_done < td->o.size)
 828                 return 1;
 829
 830         return 0;
 831 }
 832
 833 static void reset_io_counters(struct thread_data *td)
 834 {
 835         td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
 836         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
 837         td->zone_bytes = 0;
 838         td->rate_bytes = 0;
 839         td->rate_blocks = 0;
 840         td->rw_end_set[0] = td->rw_end_set[1] = 0;
 841
 842         td->last_was_sync = 0;
 843
 844         /*
 845          * reset file done count if we are to start over
 846          */
 847         if (td->o.time_based || td->o.loops)
 848                 td->nr_done_files = 0;
 849 }
 850
 851 void reset_all_stats(struct thread_data *td)
 852 {
 853         struct timeval tv;
 854         int i;
 855
 856         reset_io_counters(td);
 857
 858         for (i = 0; i < 2; i++) {
 859                 td->io_bytes[i] = 0;
 860                 td->io_blocks[i] = 0;
 861                 td->io_issues[i] = 0;
 862                 td->ts.total_io_u[i] = 0;
 863         }
 864
 865         fio_gettime(&tv, NULL);
 866         memcpy(&td->epoch, &tv, sizeof(tv));
 867         memcpy(&td->start, &tv, sizeof(tv));
 868 }
 869
 870 static int clear_io_state(struct thread_data *td)
 871 {
 872         struct fio_file *f;
 873         unsigned int i;
 874         int ret;
 875
 876         reset_io_counters(td);
 877
 878         close_files(td);
 879
 880         ret = 0;
 881         for_each_file(td, f, i) {
 882                 f->flags &= ~FIO_FILE_DONE;
 883                 ret = td_io_open_file(td, f);
 884                 if (ret)
 885                         break;
 886         }
 887
 888         return ret;
 889 }
 890
 891 /*
 892  * Entry point for the thread based jobs. The process based jobs end up
 893  * here as well, after a little setup.
 894  */
 895 static void *thread_main(void *data)
 896 {
 897         unsigned long long runtime[2], elapsed;
 898         struct thread_data *td = data;
 899         int clear_state;
 900
 901         if (!td->o.use_thread)
 902                 setsid();
 903
 904         td->pid = getpid();
 905
 906         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
 907
 908         INIT_FLIST_HEAD(&td->io_u_freelist);
 909         INIT_FLIST_HEAD(&td->io_u_busylist);
 910         INIT_FLIST_HEAD(&td->io_u_requeues);
 911         INIT_FLIST_HEAD(&td->io_log_list);
 912         INIT_FLIST_HEAD(&td->io_hist_list);
 913         td->io_hist_tree = RB_ROOT;
 914
 915         td_set_runstate(td, TD_INITIALIZED);
 916         fio_mutex_up(startup_mutex);
 917         fio_mutex_down(td->mutex);
 918
 919         /*
 920          * the ->mutex mutex is now no longer used, close it to avoid
 921          * eating a file descriptor
 922          */
 923         fio_mutex_remove(td->mutex);
 924
 925         /*
 926          * May alter parameters that init_io_u() will use, so we need to
 927          * do this first.
 928          */
 929         if (init_iolog(td))
 930                 goto err;
 931
 932         if (init_io_u(td))
 933                 goto err;
 934
 935         if (td->o.cpumask_set && fio_setaffinity(td) == -1) {
 936                 td_verror(td, errno, "cpu_set_affinity");
 937                 goto err;
 938         }
 939
 940         if (td->ioprio_set) {
 941                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
 942                         td_verror(td, errno, "ioprio_set");
 943                         goto err;
 944                 }
 945         }
 946
 947         if (nice(td->o.nice) == -1) {
 948                 td_verror(td, errno, "nice");
 949                 goto err;
 950         }
 951
 952         if (td->o.ioscheduler && switch_ioscheduler(td))
 953                 goto err;
 954
 955         if (!td->o.create_serialize && setup_files(td))
 956                 goto err;
 957
 958         if (td_io_init(td))
 959                 goto err;
 960
 961         if (open_files(td))
 962                 goto err;
 963
 964         if (init_random_map(td))
 965                 goto err;
 966
 967         if (td->o.exec_prerun) {
 968                 if (system(td->o.exec_prerun) < 0)
 969                         goto err;
 970         }
 971
 972         fio_gettime(&td->epoch, NULL);
 973         getrusage(RUSAGE_SELF, &td->ts.ru_start);
 974
 975         runtime[0] = runtime[1] = 0;
 976         clear_state = 0;
 977         while (keep_running(td)) {
 978                 fio_gettime(&td->start, NULL);
 979                 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
 980
 981                 if (td->o.ratemin)
 982                         memcpy(&td->lastrate, &td->ts.stat_sample_time,
 983                                                         sizeof(td->lastrate));
 984
 985                 if (clear_state && clear_io_state(td))
 986                         break;
 987
 988                 prune_io_piece_log(td);
 989
 990                 do_io(td);
 991
 992                 clear_state = 1;
 993
 994                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
 995                         if (td->rw_end_set[DDIR_READ])
 996                                 elapsed = utime_since(&td->start,
 997                                                       &td->rw_end[DDIR_READ]);
 998                         else
 999                                 elapsed = utime_since_now(&td->start);
1000
1001                         runtime[DDIR_READ] += elapsed;
1002                 }
1003                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1004                         if (td->rw_end_set[DDIR_WRITE])
1005                                 elapsed = utime_since(&td->start,
1006                                                       &td->rw_end[DDIR_WRITE]);
1007                         else
1008                                 elapsed = utime_since_now(&td->start);
1009
1010                         runtime[DDIR_WRITE] += elapsed;
1011                 }
1012
1013                 if (td->error || td->terminate)
1014                         break;
1015
1016                 if (!td->o.do_verify ||
1017                     td->o.verify == VERIFY_NONE ||
1018                     (td->io_ops->flags & FIO_UNIDIR))
1019                         continue;
1020
1021                 if (clear_io_state(td))
1022                         break;
1023
1024                 fio_gettime(&td->start, NULL);
1025
1026                 do_verify(td);
1027
1028                 runtime[DDIR_READ] += utime_since_now(&td->start);
1029
1030                 if (td->error || td->terminate)
1031                         break;
1032         }
1033
1034         update_rusage_stat(td);
1035         td->ts.runtime[0] = (runtime[0] + 999) / 1000;
1036         td->ts.runtime[1] = (runtime[1] + 999) / 1000;
1037         td->ts.total_run_time = mtime_since_now(&td->epoch);
1038         td->ts.io_bytes[0] = td->io_bytes[0];
1039         td->ts.io_bytes[1] = td->io_bytes[1];
1040
1041         if (td->ts.bw_log)
1042                 finish_log(td, td->ts.bw_log, "bw");
1043         if (td->ts.slat_log)
1044                 finish_log(td, td->ts.slat_log, "slat");
1045         if (td->ts.clat_log)
1046                 finish_log(td, td->ts.clat_log, "clat");
1047         if (td->o.exec_postrun) {
1048                 if (system(td->o.exec_postrun) < 0)
1049                         log_err("fio: postrun %s failed\n", td->o.exec_postrun);
1050         }
1051
1052         if (exitall_on_terminate)
1053                 terminate_threads(td->groupid);
1054
1055 err:
1056         if (td->error)
1057                 printf("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1058                                                         td->verror);
1059         close_and_free_files(td);
1060         close_ioengine(td);
1061         cleanup_io_u(td);
1062
1063         /*
1064          * do this very late, it will log file closing as well
1065          */
1066         if (td->o.write_iolog_file)
1067                 write_iolog_close(td);
1068
1069         options_mem_free(td);
1070         td_set_runstate(td, TD_EXITED);
1071         return (void *) (unsigned long) td->error;
1072 }
1073
1074 /*
1075  * We cannot pass the td data into a forked process, so attach the td and
1076  * pass it to the thread worker.
1077  */
1078 static int fork_main(int shmid, int offset)
1079 {
1080         struct thread_data *td;
1081         void *data, *ret;
1082
1083         data = shmat(shmid, NULL, 0);
1084         if (data == (void *) -1) {
1085                 int __err = errno;
1086
1087                 perror("shmat");
1088                 return __err;
1089         }
1090
1091         td = data + offset * sizeof(struct thread_data);
1092         ret = thread_main(td);
1093         shmdt(data);
1094         return (int) (unsigned long) ret;
1095 }
1096
1097 /*
1098  * Run over the job map and reap the threads that have exited, if any.
1099  */
1100 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1101 {
1102         struct thread_data *td;
1103         int i, cputhreads, realthreads, pending, status, ret;
1104
1105         /*
1106          * reap exited threads (TD_EXITED -> TD_REAPED)
1107          */
1108         realthreads = pending = cputhreads = 0;
1109         for_each_td(td, i) {
1110                 int flags = 0;
1111
1112                 /*
1113                  * ->io_ops is NULL for a thread that has closed its
1114                  * io engine
1115                  */
1116                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1117                         cputhreads++;
1118                 else
1119                         realthreads++;
1120
1121                 if (!td->pid) {
1122                         pending++;
1123                         continue;
1124                 }
1125                 if (td->runstate == TD_REAPED)
1126                         continue;
1127                 if (td->o.use_thread) {
1128                         if (td->runstate == TD_EXITED) {
1129                                 td_set_runstate(td, TD_REAPED);
1130                                 goto reaped;
1131                         }
1132                         continue;
1133                 }
1134
1135                 flags = WNOHANG;
1136                 if (td->runstate == TD_EXITED)
1137                         flags = 0;
1138
1139                 /*
1140                  * check if someone quit or got killed in an unusual way
1141                  */
1142                 ret = waitpid(td->pid, &status, flags);
1143                 if (ret < 0) {
1144                         if (errno == ECHILD) {
1145                                 log_err("fio: pid=%d disappeared %d\n",
1146                                                 (int) td->pid, td->runstate);
1147                                 td_set_runstate(td, TD_REAPED);
1148                                 goto reaped;
1149                         }
1150                         perror("waitpid");
1151                 } else if (ret == td->pid) {
1152                         if (WIFSIGNALED(status)) {
1153                                 int sig = WTERMSIG(status);
1154
1155                                 if (sig != SIGQUIT)
1156                                         log_err("fio: pid=%d, got signal=%d\n",
1157                                                         (int) td->pid, sig);
1158                                 td_set_runstate(td, TD_REAPED);
1159                                 goto reaped;
1160                         }
1161                         if (WIFEXITED(status)) {
1162                                 if (WEXITSTATUS(status) && !td->error)
1163                                         td->error = WEXITSTATUS(status);
1164
1165                                 td_set_runstate(td, TD_REAPED);
1166                                 goto reaped;
1167                         }
1168                 }
1169
1170                 /*
1171                  * thread is not dead, continue
1172                  */
1173                 pending++;
1174                 continue;
1175 reaped:
1176                 (*nr_running)--;
1177                 (*m_rate) -= td->o.ratemin;
1178                 (*t_rate) -= td->o.rate;
1179                 if (!td->pid)
1180                         pending--;
1181
1182                 if (td->error)
1183                         exit_value++;
1184
1185                 done_secs += mtime_since_now(&td->epoch) / 1000;
1186         }
1187
1188         if (*nr_running == cputhreads && !pending && realthreads)
1189                 terminate_threads(TERMINATE_ALL);
1190 }
1191
1192 /*
1193  * Main function for kicking off and reaping jobs, as needed.
1194  */
1195 static void run_threads(void)
1196 {
1197         struct thread_data *td;
1198         unsigned long spent;
1199         int i, todo, nr_running, m_rate, t_rate, nr_started;
1200
1201         if (fio_pin_memory())
1202                 return;
1203
1204         if (!terse_output) {
1205                 printf("Starting ");
1206                 if (nr_thread)
1207                         printf("%d thread%s", nr_thread,
1208                                                 nr_thread > 1 ? "s" : "");
1209                 if (nr_process) {
1210                         if (nr_thread)
1211                                 printf(" and ");
1212                         printf("%d process%s", nr_process,
1213                                                 nr_process > 1 ? "es" : "");
1214                 }
1215                 printf("\n");
1216                 fflush(stdout);
1217         }
1218
1219         set_sig_handlers();
1220
1221         todo = thread_number;
1222         nr_running = 0;
1223         nr_started = 0;
1224         m_rate = t_rate = 0;
1225
1226         for_each_td(td, i) {
1227                 print_status_init(td->thread_number - 1);
1228
1229                 if (!td->o.create_serialize) {
1230                         init_disk_util(td);
1231                         continue;
1232                 }
1233
1234                 /*
1235                  * do file setup here so it happens sequentially,
1236                  * we don't want X number of threads getting their
1237                  * client data interspersed on disk
1238                  */
1239                 if (setup_files(td)) {
1240                         exit_value++;
1241                         if (td->error)
1242                                 log_err("fio: pid=%d, err=%d/%s\n",
1243                                         (int) td->pid, td->error, td->verror);
1244                         td_set_runstate(td, TD_REAPED);
1245                         todo--;
1246                 } else {
1247                         struct fio_file *f;
1248                         unsigned int i;
1249
1250                         /*
1251                          * for sharing to work, each job must always open
1252                          * its own files. so close them, if we opened them
1253                          * for creation
1254                          */
1255                         for_each_file(td, f, i)
1256                                 td_io_close_file(td, f);
1257                 }
1258
1259                 init_disk_util(td);
1260         }
1261
1262         set_genesis_time();
1263
1264         while (todo) {
1265                 struct thread_data *map[MAX_JOBS];
1266                 struct timeval this_start;
1267                 int this_jobs = 0, left;
1268
1269                 /*
1270                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1271                  */
1272                 for_each_td(td, i) {
1273                         if (td->runstate != TD_NOT_CREATED)
1274                                 continue;
1275
1276                         /*
1277                          * never got a chance to start, killed by other
1278                          * thread for some reason
1279                          */
1280                         if (td->terminate) {
1281                                 todo--;
1282                                 continue;
1283                         }
1284
1285                         if (td->o.start_delay) {
1286                                 spent = mtime_since_genesis();
1287
1288                                 if (td->o.start_delay * 1000 > spent)
1289                                         continue;
1290                         }
1291
1292                         if (td->o.stonewall && (nr_started || nr_running)) {
1293                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1294                                                         td->o.name);
1295                                 break;
1296                         }
1297
1298                         /*
1299                          * Set state to created. Thread will transition
1300                          * to TD_INITIALIZED when it's done setting up.
1301                          */
1302                         td_set_runstate(td, TD_CREATED);
1303                         map[this_jobs++] = td;
1304                         nr_started++;
1305
1306                         if (td->o.use_thread) {
1307                                 dprint(FD_PROCESS, "will pthread_create\n");
1308                                 if (pthread_create(&td->thread, NULL,
1309                                                    thread_main, td)) {
1310                                         perror("pthread_create");
1311                                         nr_started--;
1312                                         break;
1313                                 }
1314                                 if (pthread_detach(td->thread) < 0)
1315                                         perror("pthread_detach");
1316                         } else {
1317                                 pid_t pid;
1318                                 dprint(FD_PROCESS, "will fork\n");
1319                                 pid = fork();
1320                                 if (!pid) {
1321                                         int ret = fork_main(shm_id, i);
1322
1323                                         _exit(ret);
1324                                 } else if (i == fio_debug_jobno)
1325                                         *fio_debug_jobp = pid;
1326                         }
1327                         fio_mutex_down(startup_mutex);
1328                 }
1329
1330                 /*
1331                  * Wait for the started threads to transition to
1332                  * TD_INITIALIZED.
1333                  */
1334                 fio_gettime(&this_start, NULL);
1335                 left = this_jobs;
1336                 while (left && !fio_abort) {
1337                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1338                                 break;
1339
1340                         usleep(100000);
1341
1342                         for (i = 0; i < this_jobs; i++) {
1343                                 td = map[i];
1344                                 if (!td)
1345                                         continue;
1346                                 if (td->runstate == TD_INITIALIZED) {
1347                                         map[i] = NULL;
1348                                         left--;
1349                                 } else if (td->runstate >= TD_EXITED) {
1350                                         map[i] = NULL;
1351                                         left--;
1352                                         todo--;
1353                                         nr_running++; /* work-around... */
1354                                 }
1355                         }
1356                 }
1357
1358                 if (left) {
1359                         log_err("fio: %d jobs failed to start\n", left);
1360                         for (i = 0; i < this_jobs; i++) {
1361                                 td = map[i];
1362                                 if (!td)
1363                                         continue;
1364                                 kill(td->pid, SIGTERM);
1365                         }
1366                         break;
1367                 }
1368
1369                 /*
1370                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1371                  */
1372                 for_each_td(td, i) {
1373                         if (td->runstate != TD_INITIALIZED)
1374                                 continue;
1375
1376                         if (in_ramp_time(td))
1377                                 td_set_runstate(td, TD_RAMP);
1378                         else
1379                                 td_set_runstate(td, TD_RUNNING);
1380                         nr_running++;
1381                         nr_started--;
1382                         m_rate += td->o.ratemin;
1383                         t_rate += td->o.rate;
1384                         todo--;
1385                         fio_mutex_up(td->mutex);
1386                 }
1387
1388                 reap_threads(&nr_running, &t_rate, &m_rate);
1389
1390                 if (todo)
1391                         usleep(100000);
1392         }
1393
1394         while (nr_running) {
1395                 reap_threads(&nr_running, &t_rate, &m_rate);
1396                 usleep(10000);
1397         }
1398
1399         update_io_ticks();
1400         fio_unpin_memory();
1401 }
1402
1403 int main(int argc, char *argv[])
1404 {
1405         long ps;
1406
1407         sinit();
1408
1409         /*
1410          * We need locale for number printing, if it isn't set then just
1411          * go with the US format.
1412          */
1413         if (!getenv("LC_NUMERIC"))
1414                 setlocale(LC_NUMERIC, "en_US");
1415
1416         if (parse_options(argc, argv))
1417                 return 1;
1418
1419         if (!thread_number)
1420                 return 0;
1421
1422         ps = sysconf(_SC_PAGESIZE);
1423         if (ps < 0) {
1424                 log_err("Failed to get page size\n");
1425                 return 1;
1426         }
1427
1428         page_size = ps;
1429         page_mask = ps - 1;
1430
1431         if (write_bw_log) {
1432                 setup_log(&agg_io_log[DDIR_READ]);
1433                 setup_log(&agg_io_log[DDIR_WRITE]);
1434         }
1435
1436         startup_mutex = fio_mutex_init(0);
1437
1438         set_genesis_time();
1439
1440         status_timer_arm();
1441
1442         run_threads();
1443
1444         if (!fio_abort) {
1445                 show_run_stats();
1446                 if (write_bw_log) {
1447                         __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1448                         __finish_log(agg_io_log[DDIR_WRITE],
1449                                         "agg-write_bw.log");
1450                 }
1451         }
1452
1453         fio_mutex_remove(startup_mutex);
1454         return exit_value;
1455 }