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