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