fio.c

   1 /*
   2  * fio - the flexible io tester
   3  *
   4  * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
   5  * Copyright (C) 2006 Jens Axboe <axboe@kernel.dk>
   6  *
   7  * The license below covers all files distributed with fio unless otherwise
   8  * noted in the file itself.
   9  *
  10  *  This program is free software; you can redistribute it and/or modify
  11  *  it under the terms of the GNU General Public License version 2 as
  12  *  published by the Free Software Foundation.
  13  *
  14  *  This program is distributed in the hope that it will be useful,
  15  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  *  GNU General Public License for more details.
  18  *
  19  *  You should have received a copy of the GNU General Public License
  20  *  along with this program; if not, write to the Free Software
  21  *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
  22  *
  23  */
  24 #include <unistd.h>
  25 #include <fcntl.h>
  26 #include <string.h>
  27 #include <signal.h>
  28 #include <time.h>
  29 #include <locale.h>
  30 #include <assert.h>
  31 #include <sys/stat.h>
  32 #include <sys/wait.h>
  33 #include <sys/ipc.h>
  34 #include <sys/shm.h>
  35 #include <sys/mman.h>
  36
  37 #include "fio.h"
  38 #include "hash.h"
  39 #include "smalloc.h"
  40
  41 unsigned long page_mask;
  42 unsigned long page_size;
  43 #define ALIGN(buf)      \
  44         (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
  45
  46 int groupid = 0;
  47 int thread_number = 0;
  48 int nr_process = 0;
  49 int nr_thread = 0;
  50 int shm_id = 0;
  51 int temp_stall_ts;
  52
  53 static struct fio_mutex *startup_mutex;
  54 static volatile int fio_abort;
  55 static int exit_value;
  56
  57 struct io_log *agg_io_log[2];
  58
  59 #define TERMINATE_ALL           (-1)
  60 #define JOB_START_TIMEOUT       (5 * 1000)
  61
  62 static inline void td_set_runstate(struct thread_data *td, int runstate)
  63 {
  64         if (td->runstate == runstate)
  65                 return;
  66
  67         dprint(FD_PROCESS, "pid=%d: runstate %d -> %d\n", td->pid, td->runstate,
  68                                                                 runstate);
  69         td->runstate = runstate;
  70 }
  71
  72 static void terminate_threads(int group_id)
  73 {
  74         struct thread_data *td;
  75         int i;
  76
  77         dprint(FD_PROCESS, "terminate group_id=%d\n", group_id);
  78
  79         for_each_td(td, i) {
  80                 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
  81                         dprint(FD_PROCESS, "setting terminate on %s/%d\n",
  82                                                         td->o.name, td->pid);
  83                         td->terminate = 1;
  84                         td->o.start_delay = 0;
  85
  86                         /*
  87                          * if the thread is running, just let it exit
  88                          */
  89                         if (td->runstate < TD_RUNNING)
  90                                 kill(td->pid, SIGQUIT);
  91                         else {
  92                                 struct ioengine_ops *ops = td->io_ops;
  93
  94                                 if (ops && (ops->flags & FIO_SIGQUIT))
  95                                         kill(td->pid, SIGQUIT);
  96                         }
  97                 }
  98         }
  99 }
 100
 101 static void sig_handler(int sig)
 102 {
 103         switch (sig) {
 104         case SIGALRM:
 105                 update_io_ticks();
 106                 disk_util_timer_arm();
 107                 print_thread_status();
 108                 break;
 109         default:
 110                 printf("\nfio: terminating on signal %d\n", sig);
 111                 fflush(stdout);
 112                 terminate_threads(TERMINATE_ALL);
 113                 break;
 114         }
 115 }
 116
 117 /*
 118  * Check if we are above the minimum rate given.
 119  */
 120 static int check_min_rate(struct thread_data *td, struct timeval *now)
 121 {
 122         unsigned long long bytes = 0;
 123         unsigned long iops = 0;
 124         unsigned long spent;
 125         unsigned long rate;
 126
 127         /*
 128          * No minimum rate set, always ok
 129          */
 130         if (!td->o.ratemin && !td->o.rate_iops_min)
 131                 return 0;
 132
 133         /*
 134          * allow a 2 second settle period in the beginning
 135          */
 136         if (mtime_since(&td->start, now) < 2000)
 137                 return 0;
 138
 139         if (td_read(td)) {
 140                 iops += td->io_blocks[DDIR_READ];
 141                 bytes += td->this_io_bytes[DDIR_READ];
 142         }
 143         if (td_write(td)) {
 144                 iops += td->io_blocks[DDIR_WRITE];
 145                 bytes += td->this_io_bytes[DDIR_WRITE];
 146         }
 147
 148         /*
 149          * if rate blocks is set, sample is running
 150          */
 151         if (td->rate_bytes || td->rate_blocks) {
 152                 spent = mtime_since(&td->lastrate, now);
 153                 if (spent < td->o.ratecycle)
 154                         return 0;
 155
 156                 if (td->o.rate) {
 157                         /*
 158                          * check bandwidth specified rate
 159                          */
 160                         if (bytes < td->rate_bytes) {
 161                                 log_err("%s: min rate %u not met\n", td->o.name,
 162                                                                 td->o.ratemin);
 163                                 return 1;
 164                         } else {
 165                                 rate = (bytes - td->rate_bytes) / spent;
 166                                 if (rate < td->o.ratemin ||
 167                                     bytes < td->rate_bytes) {
 168                                         log_err("%s: min rate %u not met, got"
 169                                                 " %luKiB/sec\n", td->o.name,
 170                                                         td->o.ratemin, rate);
 171                                         return 1;
 172                                 }
 173                         }
 174                 } else {
 175                         /*
 176                          * checks iops specified rate
 177                          */
 178                         if (iops < td->o.rate_iops) {
 179                                 log_err("%s: min iops rate %u not met\n",
 180                                                 td->o.name, td->o.rate_iops);
 181                                 return 1;
 182                         } else {
 183                                 rate = (iops - td->rate_blocks) / spent;
 184                                 if (rate < td->o.rate_iops_min ||
 185                                     iops < td->rate_blocks) {
 186                                         log_err("%s: min iops rate %u not met,"
 187                                                 " got %lu\n", td->o.name,
 188                                                         td->o.rate_iops_min,
 189                                                         rate);
 190                                 }
 191                         }
 192                 }
 193         }
 194
 195         td->rate_bytes = bytes;
 196         td->rate_blocks = iops;
 197         memcpy(&td->lastrate, now, sizeof(*now));
 198         return 0;
 199 }
 200
 201 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 202 {
 203         if (!td->o.timeout)
 204                 return 0;
 205         if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
 206                 return 1;
 207
 208         return 0;
 209 }
 210
 211 /*
 212  * When job exits, we can cancel the in-flight IO if we are using async
 213  * io. Attempt to do so.
 214  */
 215 static void cleanup_pending_aio(struct thread_data *td)
 216 {
 217         struct list_head *entry, *n;
 218         struct io_u *io_u;
 219         int r;
 220
 221         /*
 222          * get immediately available events, if any
 223          */
 224         r = io_u_queued_complete(td, 0);
 225         if (r < 0)
 226                 return;
 227
 228         /*
 229          * now cancel remaining active events
 230          */
 231         if (td->io_ops->cancel) {
 232                 list_for_each_safe(entry, n, &td->io_u_busylist) {
 233                         io_u = list_entry(entry, struct io_u, list);
 234
 235                         /*
 236                          * if the io_u isn't in flight, then that generally
 237                          * means someone leaked an io_u. complain but fix
 238                          * it up, so we don't stall here.
 239                          */
 240                         if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
 241                                 log_err("fio: non-busy IO on busy list\n");
 242                                 put_io_u(td, io_u);
 243                         } else {
 244                                 r = td->io_ops->cancel(td, io_u);
 245                                 if (!r)
 246                                         put_io_u(td, io_u);
 247                         }
 248                 }
 249         }
 250
 251         if (td->cur_depth)
 252                 r = io_u_queued_complete(td, td->cur_depth);
 253 }
 254
 255 /*
 256  * Helper to handle the final sync of a file. Works just like the normal
 257  * io path, just does everything sync.
 258  */
 259 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 260 {
 261         struct io_u *io_u = __get_io_u(td);
 262         int ret;
 263
 264         if (!io_u)
 265                 return 1;
 266
 267         io_u->ddir = DDIR_SYNC;
 268         io_u->file = f;
 269
 270         if (td_io_prep(td, io_u)) {
 271                 put_io_u(td, io_u);
 272                 return 1;
 273         }
 274
 275 requeue:
 276         ret = td_io_queue(td, io_u);
 277         if (ret < 0) {
 278                 td_verror(td, io_u->error, "td_io_queue");
 279                 put_io_u(td, io_u);
 280                 return 1;
 281         } else if (ret == FIO_Q_QUEUED) {
 282                 if (io_u_queued_complete(td, 1) < 0)
 283                         return 1;
 284         } else if (ret == FIO_Q_COMPLETED) {
 285                 if (io_u->error) {
 286                         td_verror(td, io_u->error, "td_io_queue");
 287                         return 1;
 288                 }
 289
 290                 if (io_u_sync_complete(td, io_u) < 0)
 291                         return 1;
 292         } else if (ret == FIO_Q_BUSY) {
 293                 if (td_io_commit(td))
 294                         return 1;
 295                 goto requeue;
 296         }
 297
 298         return 0;
 299 }
 300
 301 /*
 302  * The main verify engine. Runs over the writes we previously submitted,
 303  * reads the blocks back in, and checks the crc/md5 of the data.
 304  */
 305 static void do_verify(struct thread_data *td)
 306 {
 307         struct fio_file *f;
 308         struct io_u *io_u;
 309         int ret, min_events;
 310         unsigned int i;
 311
 312         /*
 313          * sync io first and invalidate cache, to make sure we really
 314          * read from disk.
 315          */
 316         for_each_file(td, f, i) {
 317                 if (!(f->flags & FIO_FILE_OPEN))
 318                         continue;
 319                 if (fio_io_sync(td, f))
 320                         break;
 321                 if (file_invalidate_cache(td, f))
 322                         break;
 323         }
 324
 325         if (td->error)
 326                 return;
 327
 328         td_set_runstate(td, TD_VERIFYING);
 329
 330         io_u = NULL;
 331         while (!td->terminate) {
 332                 int ret2;
 333
 334                 io_u = __get_io_u(td);
 335                 if (!io_u)
 336                         break;
 337
 338                 if (runtime_exceeded(td, &io_u->start_time)) {
 339                         put_io_u(td, io_u);
 340                         td->terminate = 1;
 341                         break;
 342                 }
 343
 344                 if (get_next_verify(td, io_u)) {
 345                         put_io_u(td, io_u);
 346                         break;
 347                 }
 348
 349                 if (td_io_prep(td, io_u)) {
 350                         put_io_u(td, io_u);
 351                         break;
 352                 }
 353
 354                 io_u->end_io = verify_io_u;
 355
 356                 ret = td_io_queue(td, io_u);
 357                 switch (ret) {
 358                 case FIO_Q_COMPLETED:
 359                         if (io_u->error)
 360                                 ret = -io_u->error;
 361                         else if (io_u->resid) {
 362                                 int bytes = io_u->xfer_buflen - io_u->resid;
 363                                 struct fio_file *f = io_u->file;
 364
 365                                 /*
 366                                  * zero read, fail
 367                                  */
 368                                 if (!bytes) {
 369                                         td_verror(td, ENODATA, "full resid");
 370                                         put_io_u(td, io_u);
 371                                         break;
 372                                 }
 373
 374                                 io_u->xfer_buflen = io_u->resid;
 375                                 io_u->xfer_buf += bytes;
 376                                 io_u->offset += bytes;
 377
 378                                 td->ts.short_io_u[io_u->ddir]++;
 379
 380                                 if (io_u->offset == f->real_file_size)
 381                                         goto sync_done;
 382
 383                                 requeue_io_u(td, &io_u);
 384                         } else {
 385 sync_done:
 386                                 ret = io_u_sync_complete(td, io_u);
 387                                 if (ret < 0)
 388                                         break;
 389                         }
 390                         continue;
 391                 case FIO_Q_QUEUED:
 392                         break;
 393                 case FIO_Q_BUSY:
 394                         requeue_io_u(td, &io_u);
 395                         ret2 = td_io_commit(td);
 396                         if (ret2 < 0)
 397                                 ret = ret2;
 398                         break;
 399                 default:
 400                         assert(ret < 0);
 401                         td_verror(td, -ret, "td_io_queue");
 402                         break;
 403                 }
 404
 405                 if (ret < 0 || td->error)
 406                         break;
 407
 408                 /*
 409                  * if we can queue more, do so. but check if there are
 410                  * completed io_u's first.
 411                  */
 412                 min_events = 0;
 413                 if (queue_full(td) || ret == FIO_Q_BUSY) {
 414                         min_events = 1;
 415
 416                         if (td->cur_depth > td->o.iodepth_low)
 417                                 min_events = td->cur_depth - td->o.iodepth_low;
 418                 }
 419
 420                 /*
 421                  * Reap required number of io units, if any, and do the
 422                  * verification on them through the callback handler
 423                  */
 424                 if (io_u_queued_complete(td, min_events) < 0)
 425                         break;
 426         }
 427
 428         if (!td->error) {
 429                 min_events = td->cur_depth;
 430
 431                 if (min_events)
 432                         ret = io_u_queued_complete(td, min_events);
 433         } else
 434                 cleanup_pending_aio(td);
 435
 436         td_set_runstate(td, TD_RUNNING);
 437 }
 438
 439 /*
 440  * Main IO worker function. It retrieves io_u's to process and queues
 441  * and reaps them, checking for rate and errors along the way.
 442  */
 443 static void do_io(struct thread_data *td)
 444 {
 445         struct timeval s;
 446         unsigned long usec;
 447         unsigned int i;
 448         int ret = 0;
 449
 450         td_set_runstate(td, TD_RUNNING);
 451
 452         while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) {
 453                 struct timeval comp_time;
 454                 long bytes_done = 0;
 455                 int min_evts = 0;
 456                 struct io_u *io_u;
 457                 int ret2;
 458
 459                 if (td->terminate)
 460                         break;
 461
 462                 io_u = get_io_u(td);
 463                 if (!io_u)
 464                         break;
 465
 466                 memcpy(&s, &io_u->start_time, sizeof(s));
 467
 468                 if (runtime_exceeded(td, &s)) {
 469                         put_io_u(td, io_u);
 470                         td->terminate = 1;
 471                         break;
 472                 }
 473
 474                 /*
 475                  * Add verification end_io handler, if asked to verify
 476                  * a previously written file.
 477                  */
 478                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ) {
 479                         io_u->end_io = verify_io_u;
 480                         td_set_runstate(td, TD_VERIFYING);
 481                 } else
 482                         td_set_runstate(td, TD_RUNNING);
 483
 484                 ret = td_io_queue(td, io_u);
 485                 switch (ret) {
 486                 case FIO_Q_COMPLETED:
 487                         if (io_u->error)
 488                                 ret = -io_u->error;
 489                         else if (io_u->resid) {
 490                                 int bytes = io_u->xfer_buflen - io_u->resid;
 491                                 struct fio_file *f = io_u->file;
 492
 493                                 /*
 494                                  * zero read, fail
 495                                  */
 496                                 if (!bytes) {
 497                                         td_verror(td, ENODATA, "full resid");
 498                                         put_io_u(td, io_u);
 499                                         break;
 500                                 }
 501
 502                                 io_u->xfer_buflen = io_u->resid;
 503                                 io_u->xfer_buf += bytes;
 504                                 io_u->offset += bytes;
 505
 506                                 td->ts.short_io_u[io_u->ddir]++;
 507
 508                                 if (io_u->offset == f->real_file_size)
 509                                         goto sync_done;
 510
 511                                 requeue_io_u(td, &io_u);
 512                         } else {
 513 sync_done:
 514                                 fio_gettime(&comp_time, NULL);
 515                                 bytes_done = io_u_sync_complete(td, io_u);
 516                                 if (bytes_done < 0)
 517                                         ret = bytes_done;
 518                         }
 519                         break;
 520                 case FIO_Q_QUEUED:
 521                         /*
 522                          * if the engine doesn't have a commit hook,
 523                          * the io_u is really queued. if it does have such
 524                          * a hook, it has to call io_u_queued() itself.
 525                          */
 526                         if (td->io_ops->commit == NULL)
 527                                 io_u_queued(td, io_u);
 528                         break;
 529                 case FIO_Q_BUSY:
 530                         requeue_io_u(td, &io_u);
 531                         ret2 = td_io_commit(td);
 532                         if (ret2 < 0)
 533                                 ret = ret2;
 534                         break;
 535                 default:
 536                         assert(ret < 0);
 537                         put_io_u(td, io_u);
 538                         break;
 539                 }
 540
 541                 if (ret < 0 || td->error)
 542                         break;
 543
 544                 /*
 545                  * See if we need to complete some commands
 546                  */
 547                 if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
 548                         min_evts = 0;
 549                         if (queue_full(td) || ret == FIO_Q_BUSY) {
 550                                 min_evts = 1;
 551
 552                                 if (td->cur_depth > td->o.iodepth_low)
 553                                         min_evts = td->cur_depth
 554                                                         - td->o.iodepth_low;
 555                         }
 556
 557                         fio_gettime(&comp_time, NULL);
 558                         bytes_done = io_u_queued_complete(td, min_evts);
 559                         if (bytes_done < 0)
 560                                 break;
 561                 }
 562
 563                 if (!bytes_done)
 564                         continue;
 565
 566                 /*
 567                  * the rate is batched for now, it should work for batches
 568                  * of completions except the very first one which may look
 569                  * a little bursty
 570                  */
 571                 usec = utime_since(&s, &comp_time);
 572
 573                 rate_throttle(td, usec, bytes_done);
 574
 575                 if (check_min_rate(td, &comp_time)) {
 576                         if (exitall_on_terminate)
 577                                 terminate_threads(td->groupid);
 578                         td_verror(td, ENODATA, "check_min_rate");
 579                         break;
 580                 }
 581
 582                 if (td->o.thinktime) {
 583                         unsigned long long b;
 584
 585                         b = td->io_blocks[0] + td->io_blocks[1];
 586                         if (!(b % td->o.thinktime_blocks)) {
 587                                 int left;
 588
 589                                 if (td->o.thinktime_spin)
 590                                         __usec_sleep(td->o.thinktime_spin);
 591
 592                                 left = td->o.thinktime - td->o.thinktime_spin;
 593                                 if (left)
 594                                         usec_sleep(td, left);
 595                         }
 596                 }
 597         }
 598
 599         if (td->o.fill_device && td->error == ENOSPC) {
 600                 td->error = 0;
 601                 td->terminate = 1;
 602         }
 603         if (!td->error) {
 604                 struct fio_file *f;
 605
 606                 i = td->cur_depth;
 607                 if (i)
 608                         ret = io_u_queued_complete(td, i);
 609
 610                 if (should_fsync(td) && td->o.end_fsync) {
 611                         td_set_runstate(td, TD_FSYNCING);
 612
 613                         for_each_file(td, f, i) {
 614                                 if (!(f->flags & FIO_FILE_OPEN))
 615                                         continue;
 616                                 fio_io_sync(td, f);
 617                         }
 618                 }
 619         } else
 620                 cleanup_pending_aio(td);
 621
 622         /*
 623          * stop job if we failed doing any IO
 624          */
 625         if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
 626                 td->done = 1;
 627 }
 628
 629 static void cleanup_io_u(struct thread_data *td)
 630 {
 631         struct list_head *entry, *n;
 632         struct io_u *io_u;
 633
 634         list_for_each_safe(entry, n, &td->io_u_freelist) {
 635                 io_u = list_entry(entry, struct io_u, list);
 636
 637                 list_del(&io_u->list);
 638                 free(io_u);
 639         }
 640
 641         free_io_mem(td);
 642 }
 643
 644 /*
 645  * "randomly" fill the buffer contents
 646  */
 647 static void fill_io_buf(struct thread_data *td, struct io_u *io_u, int max_bs)
 648 {
 649         long *ptr = io_u->buf;
 650
 651         if (!td->o.zero_buffers) {
 652                 while ((void *) ptr - io_u->buf < max_bs) {
 653                         *ptr = rand() * GOLDEN_RATIO_PRIME;
 654                         ptr++;
 655                 }
 656         } else
 657                 memset(ptr, 0, max_bs);
 658 }
 659
 660 static int init_io_u(struct thread_data *td)
 661 {
 662         struct io_u *io_u;
 663         unsigned int max_bs;
 664         int i, max_units;
 665         char *p;
 666
 667         max_units = td->o.iodepth;
 668         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
 669         td->orig_buffer_size = (unsigned long long) max_bs
 670                                         * (unsigned long long) max_units;
 671
 672         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
 673                 unsigned long bs;
 674
 675                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
 676                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
 677         }
 678
 679         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
 680                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
 681                 return 1;
 682         }
 683
 684         if (allocate_io_mem(td))
 685                 return 1;
 686
 687         if (td->o.odirect)
 688                 p = ALIGN(td->orig_buffer);
 689         else
 690                 p = td->orig_buffer;
 691
 692         for (i = 0; i < max_units; i++) {
 693                 if (td->terminate)
 694                         return 1;
 695                 io_u = malloc(sizeof(*io_u));
 696                 memset(io_u, 0, sizeof(*io_u));
 697                 INIT_LIST_HEAD(&io_u->list);
 698
 699                 if (!(td->io_ops->flags & FIO_NOIO)) {
 700                         io_u->buf = p + max_bs * i;
 701
 702                         if (td_write(td))
 703                                 fill_io_buf(td, io_u, max_bs);
 704                 }
 705
 706                 io_u->index = i;
 707                 io_u->flags = IO_U_F_FREE;
 708                 list_add(&io_u->list, &td->io_u_freelist);
 709         }
 710
 711         io_u_init_timeout();
 712
 713         return 0;
 714 }
 715
 716 static int switch_ioscheduler(struct thread_data *td)
 717 {
 718         char tmp[256], tmp2[128];
 719         FILE *f;
 720         int ret;
 721
 722         if (td->io_ops->flags & FIO_DISKLESSIO)
 723                 return 0;
 724
 725         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 726
 727         f = fopen(tmp, "r+");
 728         if (!f) {
 729                 if (errno == ENOENT) {
 730                         log_err("fio: os or kernel doesn't support IO scheduler"
 731                                 " switching\n");
 732                         return 0;
 733                 }
 734                 td_verror(td, errno, "fopen iosched");
 735                 return 1;
 736         }
 737
 738         /*
 739          * Set io scheduler.
 740          */
 741         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
 742         if (ferror(f) || ret != 1) {
 743                 td_verror(td, errno, "fwrite");
 744                 fclose(f);
 745                 return 1;
 746         }
 747
 748         rewind(f);
 749
 750         /*
 751          * Read back and check that the selected scheduler is now the default.
 752          */
 753         ret = fread(tmp, 1, sizeof(tmp), f);
 754         if (ferror(f) || ret < 0) {
 755                 td_verror(td, errno, "fread");
 756                 fclose(f);
 757                 return 1;
 758         }
 759
 760         sprintf(tmp2, "[%s]", td->o.ioscheduler);
 761         if (!strstr(tmp, tmp2)) {
 762                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
 763                 td_verror(td, EINVAL, "iosched_switch");
 764                 fclose(f);
 765                 return 1;
 766         }
 767
 768         fclose(f);
 769         return 0;
 770 }
 771
 772 static int keep_running(struct thread_data *td)
 773 {
 774         unsigned long long io_done;
 775
 776         if (td->done)
 777                 return 0;
 778         if (td->o.time_based)
 779                 return 1;
 780         if (td->o.loops) {
 781                 td->o.loops--;
 782                 return 1;
 783         }
 784
 785         io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
 786                         + td->io_skip_bytes;
 787         if (io_done < td->o.size)
 788                 return 1;
 789
 790         return 0;
 791 }
 792
 793 static int clear_io_state(struct thread_data *td)
 794 {
 795         struct fio_file *f;
 796         unsigned int i;
 797         int ret;
 798
 799         td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
 800         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
 801         td->zone_bytes = 0;
 802         td->rate_bytes = 0;
 803         td->rate_blocks = 0;
 804         td->rw_end_set[0] = td->rw_end_set[1] = 0;
 805
 806         td->last_was_sync = 0;
 807
 808         /*
 809          * reset file done count if we are to start over
 810          */
 811         if (td->o.time_based || td->o.loops)
 812                 td->nr_done_files = 0;
 813
 814         close_files(td);
 815
 816         ret = 0;
 817         for_each_file(td, f, i) {
 818                 f->flags &= ~FIO_FILE_DONE;
 819                 ret = td_io_open_file(td, f);
 820                 if (ret)
 821                         break;
 822         }
 823
 824         return ret;
 825 }
 826
 827 /*
 828  * Entry point for the thread based jobs. The process based jobs end up
 829  * here as well, after a little setup.
 830  */
 831 static void *thread_main(void *data)
 832 {
 833         unsigned long long runtime[2], elapsed;
 834         struct thread_data *td = data;
 835         int clear_state;
 836
 837         if (!td->o.use_thread)
 838                 setsid();
 839
 840         td->pid = getpid();
 841
 842         dprint(FD_PROCESS, "jobs pid=%d started\n", td->pid);
 843
 844         INIT_LIST_HEAD(&td->io_u_freelist);
 845         INIT_LIST_HEAD(&td->io_u_busylist);
 846         INIT_LIST_HEAD(&td->io_u_requeues);
 847         INIT_LIST_HEAD(&td->io_log_list);
 848         INIT_LIST_HEAD(&td->io_hist_list);
 849         td->io_hist_tree = RB_ROOT;
 850
 851         td_set_runstate(td, TD_INITIALIZED);
 852         fio_mutex_up(startup_mutex);
 853         fio_mutex_down(td->mutex);
 854
 855         /*
 856          * the ->mutex mutex is now no longer used, close it to avoid
 857          * eating a file descriptor
 858          */
 859         fio_mutex_remove(td->mutex);
 860
 861         /*
 862          * May alter parameters that init_io_u() will use, so we need to
 863          * do this first.
 864          */
 865         if (init_iolog(td))
 866                 goto err;
 867
 868         if (init_io_u(td))
 869                 goto err;
 870
 871         if (td->o.cpumask_set && fio_setaffinity(td) == -1) {
 872                 td_verror(td, errno, "cpu_set_affinity");
 873                 goto err;
 874         }
 875
 876         if (td->ioprio_set) {
 877                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
 878                         td_verror(td, errno, "ioprio_set");
 879                         goto err;
 880                 }
 881         }
 882
 883         if (nice(td->o.nice) == -1) {
 884                 td_verror(td, errno, "nice");
 885                 goto err;
 886         }
 887
 888         if (td->o.ioscheduler && switch_ioscheduler(td))
 889                 goto err;
 890
 891         if (!td->o.create_serialize && setup_files(td))
 892                 goto err;
 893
 894         if (td_io_init(td))
 895                 goto err;
 896
 897         if (open_files(td))
 898                 goto err;
 899
 900         if (init_random_map(td))
 901                 goto err;
 902
 903         if (td->o.exec_prerun) {
 904                 if (system(td->o.exec_prerun) < 0)
 905                         goto err;
 906         }
 907
 908         fio_gettime(&td->epoch, NULL);
 909         memcpy(&td->timeout_end, &td->epoch, sizeof(td->epoch));
 910         getrusage(RUSAGE_SELF, &td->ts.ru_start);
 911
 912         runtime[0] = runtime[1] = 0;
 913         clear_state = 0;
 914         while (keep_running(td)) {
 915                 fio_gettime(&td->start, NULL);
 916                 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
 917
 918                 if (td->o.ratemin)
 919                         memcpy(&td->lastrate, &td->ts.stat_sample_time,
 920                                                         sizeof(td->lastrate));
 921
 922                 if (clear_state && clear_io_state(td))
 923                         break;
 924
 925                 prune_io_piece_log(td);
 926
 927                 do_io(td);
 928
 929                 clear_state = 1;
 930
 931                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
 932                         if (td->rw_end_set[DDIR_READ])
 933                                 elapsed = utime_since(&td->start,
 934                                                       &td->rw_end[DDIR_READ]);
 935                         else
 936                                 elapsed = utime_since_now(&td->start);
 937
 938                         runtime[DDIR_READ] += elapsed;
 939                 }
 940                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
 941                         if (td->rw_end_set[DDIR_WRITE])
 942                                 elapsed = utime_since(&td->start,
 943                                                       &td->rw_end[DDIR_WRITE]);
 944                         else
 945                                 elapsed = utime_since_now(&td->start);
 946
 947                         runtime[DDIR_WRITE] += elapsed;
 948                 }
 949
 950                 if (td->error || td->terminate)
 951                         break;
 952
 953                 if (!td->o.do_verify ||
 954                     td->o.verify == VERIFY_NONE ||
 955                     (td->io_ops->flags & FIO_UNIDIR))
 956                         continue;
 957
 958                 if (clear_io_state(td))
 959                         break;
 960
 961                 fio_gettime(&td->start, NULL);
 962
 963                 do_verify(td);
 964
 965                 runtime[DDIR_READ] += utime_since_now(&td->start);
 966
 967                 if (td->error || td->terminate)
 968                         break;
 969         }
 970
 971         update_rusage_stat(td);
 972         td->ts.runtime[0] = (runtime[0] + 999) / 1000;
 973         td->ts.runtime[1] = (runtime[1] + 999) / 1000;
 974         td->ts.total_run_time = mtime_since_now(&td->epoch);
 975         td->ts.io_bytes[0] = td->io_bytes[0];
 976         td->ts.io_bytes[1] = td->io_bytes[1];
 977
 978         if (td->ts.bw_log)
 979                 finish_log(td, td->ts.bw_log, "bw");
 980         if (td->ts.slat_log)
 981                 finish_log(td, td->ts.slat_log, "slat");
 982         if (td->ts.clat_log)
 983                 finish_log(td, td->ts.clat_log, "clat");
 984         if (td->o.exec_postrun) {
 985                 if (system(td->o.exec_postrun) < 0)
 986                         log_err("fio: postrun %s failed\n", td->o.exec_postrun);
 987         }
 988
 989         if (exitall_on_terminate)
 990                 terminate_threads(td->groupid);
 991
 992 err:
 993         if (td->error)
 994                 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error,
 995                                                         td->verror);
 996         close_and_free_files(td);
 997         close_ioengine(td);
 998         cleanup_io_u(td);
 999
1000         /*
1001          * do this very late, it will log file closing as well
1002          */
1003         if (td->o.write_iolog_file)
1004                 write_iolog_close(td);
1005
1006         options_mem_free(td);
1007         td_set_runstate(td, TD_EXITED);
1008         return (void *) (unsigned long) td->error;
1009 }
1010
1011 /*
1012  * We cannot pass the td data into a forked process, so attach the td and
1013  * pass it to the thread worker.
1014  */
1015 static int fork_main(int shmid, int offset)
1016 {
1017         struct thread_data *td;
1018         void *data, *ret;
1019
1020         data = shmat(shmid, NULL, 0);
1021         if (data == (void *) -1) {
1022                 int __err = errno;
1023
1024                 perror("shmat");
1025                 return __err;
1026         }
1027
1028         td = data + offset * sizeof(struct thread_data);
1029         ret = thread_main(td);
1030         shmdt(data);
1031         return (int) (unsigned long) ret;
1032 }
1033
1034 /*
1035  * Run over the job map and reap the threads that have exited, if any.
1036  */
1037 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1038 {
1039         struct thread_data *td;
1040         int i, cputhreads, realthreads, pending, status, ret;
1041
1042         /*
1043          * reap exited threads (TD_EXITED -> TD_REAPED)
1044          */
1045         realthreads = pending = cputhreads = 0;
1046         for_each_td(td, i) {
1047                 int flags = 0;
1048
1049                 /*
1050                  * ->io_ops is NULL for a thread that has closed its
1051                  * io engine
1052                  */
1053                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1054                         cputhreads++;
1055                 else
1056                         realthreads++;
1057
1058                 if (!td->pid) {
1059                         pending++;
1060                         continue;
1061                 }
1062                 if (td->runstate == TD_REAPED)
1063                         continue;
1064                 if (td->o.use_thread) {
1065                         if (td->runstate == TD_EXITED) {
1066                                 td_set_runstate(td, TD_REAPED);
1067                                 goto reaped;
1068                         }
1069                         continue;
1070                 }
1071
1072                 flags = WNOHANG;
1073                 if (td->runstate == TD_EXITED)
1074                         flags = 0;
1075
1076                 /*
1077                  * check if someone quit or got killed in an unusual way
1078                  */
1079                 ret = waitpid(td->pid, &status, flags);
1080                 if (ret < 0) {
1081                         if (errno == ECHILD) {
1082                                 log_err("fio: pid=%d disappeared %d\n", td->pid,
1083                                                                 td->runstate);
1084                                 td_set_runstate(td, TD_REAPED);
1085                                 goto reaped;
1086                         }
1087                         perror("waitpid");
1088                 } else if (ret == td->pid) {
1089                         if (WIFSIGNALED(status)) {
1090                                 int sig = WTERMSIG(status);
1091
1092                                 if (sig != SIGQUIT)
1093                                         log_err("fio: pid=%d, got signal=%d\n",
1094                                                                 td->pid, sig);
1095                                 td_set_runstate(td, TD_REAPED);
1096                                 goto reaped;
1097                         }
1098                         if (WIFEXITED(status)) {
1099                                 if (WEXITSTATUS(status) && !td->error)
1100                                         td->error = WEXITSTATUS(status);
1101
1102                                 td_set_runstate(td, TD_REAPED);
1103                                 goto reaped;
1104                         }
1105                 }
1106
1107                 /*
1108                  * thread is not dead, continue
1109                  */
1110                 pending++;
1111                 continue;
1112 reaped:
1113                 (*nr_running)--;
1114                 (*m_rate) -= td->o.ratemin;
1115                 (*t_rate) -= td->o.rate;
1116                 if (!td->pid)
1117                         pending--;
1118
1119                 if (td->error)
1120                         exit_value++;
1121         }
1122
1123         if (*nr_running == cputhreads && !pending && realthreads)
1124                 terminate_threads(TERMINATE_ALL);
1125 }
1126
1127 /*
1128  * Main function for kicking off and reaping jobs, as needed.
1129  */
1130 static void run_threads(void)
1131 {
1132         struct thread_data *td;
1133         unsigned long spent;
1134         int i, todo, nr_running, m_rate, t_rate, nr_started;
1135
1136         if (fio_pin_memory())
1137                 return;
1138
1139         if (!terse_output) {
1140                 printf("Starting ");
1141                 if (nr_thread)
1142                         printf("%d thread%s", nr_thread,
1143                                                 nr_thread > 1 ? "s" : "");
1144                 if (nr_process) {
1145                         if (nr_thread)
1146                                 printf(" and ");
1147                         printf("%d process%s", nr_process,
1148                                                 nr_process > 1 ? "es" : "");
1149                 }
1150                 printf("\n");
1151                 fflush(stdout);
1152         }
1153
1154         signal(SIGINT, sig_handler);
1155         signal(SIGALRM, sig_handler);
1156
1157         todo = thread_number;
1158         nr_running = 0;
1159         nr_started = 0;
1160         m_rate = t_rate = 0;
1161
1162         for_each_td(td, i) {
1163                 print_status_init(td->thread_number - 1);
1164
1165                 if (!td->o.create_serialize) {
1166                         init_disk_util(td);
1167                         continue;
1168                 }
1169
1170                 /*
1171                  * do file setup here so it happens sequentially,
1172                  * we don't want X number of threads getting their
1173                  * client data interspersed on disk
1174                  */
1175                 if (setup_files(td)) {
1176                         exit_value++;
1177                         if (td->error)
1178                                 log_err("fio: pid=%d, err=%d/%s\n", td->pid,
1179                                                         td->error, td->verror);
1180                         td_set_runstate(td, TD_REAPED);
1181                         todo--;
1182                 } else {
1183                         struct fio_file *f;
1184                         unsigned int i;
1185
1186                         /*
1187                          * for sharing to work, each job must always open
1188                          * its own files. so close them, if we opened them
1189                          * for creation
1190                          */
1191                         for_each_file(td, f, i)
1192                                 td_io_close_file(td, f);
1193                 }
1194
1195                 init_disk_util(td);
1196         }
1197
1198         set_genesis_time();
1199
1200         while (todo) {
1201                 struct thread_data *map[MAX_JOBS];
1202                 struct timeval this_start;
1203                 int this_jobs = 0, left;
1204
1205                 /*
1206                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1207                  */
1208                 for_each_td(td, i) {
1209                         if (td->runstate != TD_NOT_CREATED)
1210                                 continue;
1211
1212                         /*
1213                          * never got a chance to start, killed by other
1214                          * thread for some reason
1215                          */
1216                         if (td->terminate) {
1217                                 todo--;
1218                                 continue;
1219                         }
1220
1221                         if (td->o.start_delay) {
1222                                 spent = mtime_since_genesis();
1223
1224                                 if (td->o.start_delay * 1000 > spent)
1225                                         continue;
1226                         }
1227
1228                         if (td->o.stonewall && (nr_started || nr_running)) {
1229                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1230                                                         td->o.name);
1231                                 break;
1232                         }
1233
1234                         /*
1235                          * Set state to created. Thread will transition
1236                          * to TD_INITIALIZED when it's done setting up.
1237                          */
1238                         td_set_runstate(td, TD_CREATED);
1239                         map[this_jobs++] = td;
1240                         nr_started++;
1241
1242                         if (td->o.use_thread) {
1243                                 dprint(FD_PROCESS, "will pthread_create\n");
1244                                 if (pthread_create(&td->thread, NULL,
1245                                                    thread_main, td)) {
1246                                         perror("pthread_create");
1247                                         nr_started--;
1248                                         break;
1249                                 }
1250                                 if (pthread_detach(td->thread) < 0)
1251                                         perror("pthread_detach");
1252                         } else {
1253                                 dprint(FD_PROCESS, "will fork\n");
1254                                 if (!fork()) {
1255                                         int ret = fork_main(shm_id, i);
1256
1257                                         exit(ret);
1258                                 }
1259                         }
1260                         fio_mutex_down(startup_mutex);
1261                 }
1262
1263                 /*
1264                  * Wait for the started threads to transition to
1265                  * TD_INITIALIZED.
1266                  */
1267                 fio_gettime(&this_start, NULL);
1268                 left = this_jobs;
1269                 while (left && !fio_abort) {
1270                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1271                                 break;
1272
1273                         usleep(100000);
1274
1275                         for (i = 0; i < this_jobs; i++) {
1276                                 td = map[i];
1277                                 if (!td)
1278                                         continue;
1279                                 if (td->runstate == TD_INITIALIZED) {
1280                                         map[i] = NULL;
1281                                         left--;
1282                                 } else if (td->runstate >= TD_EXITED) {
1283                                         map[i] = NULL;
1284                                         left--;
1285                                         todo--;
1286                                         nr_running++; /* work-around... */
1287                                 }
1288                         }
1289                 }
1290
1291                 if (left) {
1292                         log_err("fio: %d jobs failed to start\n", left);
1293                         for (i = 0; i < this_jobs; i++) {
1294                                 td = map[i];
1295                                 if (!td)
1296                                         continue;
1297                                 kill(td->pid, SIGTERM);
1298                         }
1299                         break;
1300                 }
1301
1302                 /*
1303                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1304                  */
1305                 for_each_td(td, i) {
1306                         if (td->runstate != TD_INITIALIZED)
1307                                 continue;
1308
1309                         td_set_runstate(td, TD_RUNNING);
1310                         nr_running++;
1311                         nr_started--;
1312                         m_rate += td->o.ratemin;
1313                         t_rate += td->o.rate;
1314                         todo--;
1315                         fio_mutex_up(td->mutex);
1316                 }
1317
1318                 reap_threads(&nr_running, &t_rate, &m_rate);
1319
1320                 if (todo)
1321                         usleep(100000);
1322         }
1323
1324         while (nr_running) {
1325                 reap_threads(&nr_running, &t_rate, &m_rate);
1326                 usleep(10000);
1327         }
1328
1329         update_io_ticks();
1330         fio_unpin_memory();
1331 }
1332
1333 int main(int argc, char *argv[])
1334 {
1335         long ps;
1336
1337         sinit();
1338
1339         /*
1340          * We need locale for number printing, if it isn't set then just
1341          * go with the US format.
1342          */
1343         if (!getenv("LC_NUMERIC"))
1344                 setlocale(LC_NUMERIC, "en_US");
1345
1346         if (parse_options(argc, argv))
1347                 return 1;
1348
1349         if (!thread_number)
1350                 return 0;
1351
1352         ps = sysconf(_SC_PAGESIZE);
1353         if (ps < 0) {
1354                 log_err("Failed to get page size\n");
1355                 return 1;
1356         }
1357
1358         page_size = ps;
1359         page_mask = ps - 1;
1360
1361         if (write_bw_log) {
1362                 setup_log(&agg_io_log[DDIR_READ]);
1363                 setup_log(&agg_io_log[DDIR_WRITE]);
1364         }
1365
1366         startup_mutex = fio_mutex_init(0);
1367
1368         set_genesis_time();
1369
1370         disk_util_timer_arm();
1371
1372         run_threads();
1373
1374         if (!fio_abort) {
1375                 show_run_stats();
1376                 if (write_bw_log) {
1377                         __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1378                         __finish_log(agg_io_log[DDIR_WRITE],
1379                                         "agg-write_bw.log");
1380                 }
1381         }
1382
1383         fio_mutex_remove(startup_mutex);
1384         return exit_value;
1385 }