fio.c

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