fio.c

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