fio.c

   1 /*
   2  * fio - the flexible io tester
   3  *
   4  * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
   5  *
   6  *  This program is free software; you can redistribute it and/or modify
   7  *  it under the terms of the GNU General Public License as published by
   8  *  the Free Software Foundation; either version 2 of the License, or
   9  *  (at your option) any later version.
  10  *
  11  *  This program is distributed in the hope that it will be useful,
  12  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
  13  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  14  *  GNU General Public License for more details.
  15  *
  16  *  You should have received a copy of the GNU General Public License
  17  *  along with this program; if not, write to the Free Software
  18  *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
  19  *
  20  */
  21 #include <stdio.h>
  22 #include <stdlib.h>
  23 #include <unistd.h>
  24 #include <fcntl.h>
  25 #include <string.h>
  26 #include <errno.h>
  27 #include <signal.h>
  28 #include <time.h>
  29 #include <math.h>
  30 #include <assert.h>
  31 #include <dirent.h>
  32 #include <libgen.h>
  33 #include <sys/types.h>
  34 #include <sys/stat.h>
  35 #include <sys/wait.h>
  36 #include <sys/ipc.h>
  37 #include <sys/shm.h>
  38 #include <sys/ioctl.h>
  39 #include <sys/mman.h>
  40
  41 #include "fio.h"
  42 #include "os.h"
  43
  44 #define MASK    (4095)
  45
  46 #define ALIGN(buf)      (char *) (((unsigned long) (buf) + MASK) & ~(MASK))
  47
  48 int groupid = 0;
  49 int thread_number = 0;
  50 static char run_str[MAX_JOBS + 1];
  51 int shm_id = 0;
  52 static LIST_HEAD(disk_list);
  53 static struct itimerval itimer;
  54 static struct timeval genesis;
  55
  56 static void update_io_ticks(void);
  57 static void disk_util_timer_arm(void);
  58 static void print_thread_status(void);
  59
  60 extern unsigned long long mlock_size;
  61
  62 /*
  63  * thread life cycle
  64  */
  65 enum {
  66         TD_NOT_CREATED = 0,
  67         TD_CREATED,
  68         TD_INITIALIZED,
  69         TD_RUNNING,
  70         TD_VERIFYING,
  71         TD_EXITED,
  72         TD_REAPED,
  73 };
  74
  75 #define should_fsync(td)        ((td_write(td) || td_rw(td)) && (!(td)->odirect || (td)->override_sync))
  76
  77 static sem_t startup_sem;
  78
  79 #define TERMINATE_ALL           (-1)
  80 #define JOB_START_TIMEOUT       (5 * 1000)
  81
  82 static void terminate_threads(int group_id)
  83 {
  84         int i;
  85
  86         for (i = 0; i < thread_number; i++) {
  87                 struct thread_data *td = &threads[i];
  88
  89                 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
  90                         td->terminate = 1;
  91                         td->start_delay = 0;
  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\n");
 106                         fflush(stdout);
 107                         terminate_threads(TERMINATE_ALL);
 108                         break;
 109         }
 110 }
 111
 112 static unsigned long utime_since(struct timeval *s, struct timeval *e)
 113 {
 114         double sec, usec;
 115
 116         sec = e->tv_sec - s->tv_sec;
 117         usec = e->tv_usec - s->tv_usec;
 118         if (sec > 0 && usec < 0) {
 119                 sec--;
 120                 usec += 1000000;
 121         }
 122
 123         sec *= (double) 1000000;
 124
 125         return sec + usec;
 126 }
 127
 128 static unsigned long utime_since_now(struct timeval *s)
 129 {
 130         struct timeval t;
 131
 132         gettimeofday(&t, NULL);
 133         return utime_since(s, &t);
 134 }
 135
 136 static unsigned long mtime_since(struct timeval *s, struct timeval *e)
 137 {
 138         double sec, usec;
 139
 140         sec = e->tv_sec - s->tv_sec;
 141         usec = e->tv_usec - s->tv_usec;
 142         if (sec > 0 && usec < 0) {
 143                 sec--;
 144                 usec += 1000000;
 145         }
 146
 147         sec *= (double) 1000;
 148         usec /= (double) 1000;
 149
 150         return sec + usec;
 151 }
 152
 153 static unsigned long mtime_since_now(struct timeval *s)
 154 {
 155         struct timeval t;
 156
 157         gettimeofday(&t, NULL);
 158         return mtime_since(s, &t);
 159 }
 160
 161 static inline unsigned long msec_now(struct timeval *s)
 162 {
 163         return s->tv_sec * 1000 + s->tv_usec / 1000;
 164 }
 165
 166 static unsigned long time_since_now(struct timeval *s)
 167 {
 168         return mtime_since_now(s) / 1000;
 169 }
 170
 171 static int random_map_free(struct thread_data *td, unsigned long long block)
 172 {
 173         unsigned int idx = RAND_MAP_IDX(td, block);
 174         unsigned int bit = RAND_MAP_BIT(td, block);
 175
 176         return (td->file_map[idx] & (1UL << bit)) == 0;
 177 }
 178
 179 static int get_next_free_block(struct thread_data *td, unsigned long long *b)
 180 {
 181         int i;
 182
 183         *b = 0;
 184         i = 0;
 185         while ((*b) * td->min_bs < td->io_size) {
 186                 if (td->file_map[i] != -1UL) {
 187                         *b += ffz(td->file_map[i]);
 188                         return 0;
 189                 }
 190
 191                 *b += BLOCKS_PER_MAP;
 192                 i++;
 193         }
 194
 195         return 1;
 196 }
 197
 198 static void mark_random_map(struct thread_data *td, struct io_u *io_u)
 199 {
 200         unsigned long long block = io_u->offset / (unsigned long long) td->min_bs;
 201         unsigned int blocks = 0;
 202
 203         while (blocks < (io_u->buflen / td->min_bs)) {
 204                 unsigned int idx, bit;
 205
 206                 if (!random_map_free(td, block))
 207                         break;
 208
 209                 idx = RAND_MAP_IDX(td, block);
 210                 bit = RAND_MAP_BIT(td, block);
 211
 212                 assert(idx < td->num_maps);
 213
 214                 td->file_map[idx] |= (1UL << bit);
 215                 block++;
 216                 blocks++;
 217         }
 218
 219         if ((blocks * td->min_bs) < io_u->buflen)
 220                 io_u->buflen = blocks * td->min_bs;
 221 }
 222
 223 static inline void add_stat_sample(struct io_stat *is, unsigned long val)
 224 {
 225         if (val > is->max_val)
 226                 is->max_val = val;
 227         if (val < is->min_val)
 228                 is->min_val = val;
 229
 230         is->val += val;
 231         is->val_sq += val * val;
 232         is->samples++;
 233 }
 234
 235 static void add_log_sample(struct thread_data *td, struct io_log *iolog,
 236                            unsigned long val, int ddir)
 237 {
 238         if (iolog->nr_samples == iolog->max_samples) {
 239                 int new_size = sizeof(struct io_sample) * iolog->max_samples*2;
 240
 241                 iolog->log = realloc(iolog->log, new_size);
 242                 iolog->max_samples <<= 1;
 243         }
 244
 245         iolog->log[iolog->nr_samples].val = val;
 246         iolog->log[iolog->nr_samples].time = mtime_since_now(&td->epoch);
 247         iolog->log[iolog->nr_samples].ddir = ddir;
 248         iolog->nr_samples++;
 249 }
 250
 251 static void add_clat_sample(struct thread_data *td, int ddir,unsigned long msec)
 252 {
 253         add_stat_sample(&td->clat_stat[ddir], msec);
 254
 255         if (td->clat_log)
 256                 add_log_sample(td, td->clat_log, msec, ddir);
 257 }
 258
 259 static void add_slat_sample(struct thread_data *td, int ddir,unsigned long msec)
 260 {
 261         add_stat_sample(&td->slat_stat[ddir], msec);
 262
 263         if (td->slat_log)
 264                 add_log_sample(td, td->slat_log, msec, ddir);
 265 }
 266
 267 static void add_bw_sample(struct thread_data *td, int ddir)
 268 {
 269         unsigned long spent = mtime_since_now(&td->stat_sample_time[ddir]);
 270         unsigned long rate;
 271
 272         if (spent < td->bw_avg_time)
 273                 return;
 274
 275         rate = (td->this_io_bytes[ddir] - td->stat_io_bytes[ddir]) / spent;
 276         add_stat_sample(&td->bw_stat[ddir], rate);
 277
 278         if (td->bw_log)
 279                 add_log_sample(td, td->bw_log, rate, ddir);
 280
 281         gettimeofday(&td->stat_sample_time[ddir], NULL);
 282         td->stat_io_bytes[ddir] = td->this_io_bytes[ddir];
 283 }
 284
 285 static int get_next_offset(struct thread_data *td, unsigned long long *offset)
 286 {
 287         unsigned long long b, rb;
 288         long r;
 289
 290         if (!td->sequential) {
 291                 unsigned long long max_blocks = td->io_size / td->min_bs;
 292                 int loops = 50;
 293
 294                 do {
 295                         lrand48_r(&td->random_state, &r);
 296                         b = ((max_blocks - 1) * r / (unsigned long long) (RAND_MAX+1.0));
 297                         rb = b + (td->file_offset / td->min_bs);
 298                         loops--;
 299                 } while (!random_map_free(td, rb) && loops);
 300
 301                 if (!loops) {
 302                         if (get_next_free_block(td, &b))
 303                                 return 1;
 304                 }
 305         } else
 306                 b = td->last_pos / td->min_bs;
 307
 308         *offset = (b * td->min_bs) + td->file_offset;
 309         if (*offset > td->real_file_size)
 310                 return 1;
 311
 312         return 0;
 313 }
 314
 315 static unsigned int get_next_buflen(struct thread_data *td)
 316 {
 317         unsigned int buflen;
 318         long r;
 319
 320         if (td->min_bs == td->max_bs)
 321                 buflen = td->min_bs;
 322         else {
 323                 lrand48_r(&td->bsrange_state, &r);
 324                 buflen = (1 + (double) (td->max_bs - 1) * r / (RAND_MAX + 1.0));
 325                 buflen = (buflen + td->min_bs - 1) & ~(td->min_bs - 1);
 326         }
 327
 328         if (buflen > td->io_size - td->this_io_bytes[td->ddir])
 329                 buflen = td->io_size - td->this_io_bytes[td->ddir];
 330
 331         return buflen;
 332 }
 333
 334 /*
 335  * busy looping version for the last few usec
 336  */
 337 static void __usec_sleep(unsigned int usec)
 338 {
 339         struct timeval start;
 340
 341         gettimeofday(&start, NULL);
 342         while (utime_since_now(&start) < usec)
 343                 nop;
 344 }
 345
 346 static void usec_sleep(struct thread_data *td, unsigned long usec)
 347 {
 348         struct timespec req, rem;
 349
 350         req.tv_sec = usec / 1000000;
 351         req.tv_nsec = usec * 1000 - req.tv_sec * 1000000;
 352
 353         do {
 354                 if (usec < 5000) {
 355                         __usec_sleep(usec);
 356                         break;
 357                 }
 358
 359                 rem.tv_sec = rem.tv_nsec = 0;
 360                 if (nanosleep(&req, &rem) < 0)
 361                         break;
 362
 363                 if ((rem.tv_sec + rem.tv_nsec) == 0)
 364                         break;
 365
 366                 req.tv_nsec = rem.tv_nsec;
 367                 req.tv_sec = rem.tv_sec;
 368
 369                 usec = rem.tv_sec * 1000000 + rem.tv_nsec / 1000;
 370         } while (!td->terminate);
 371 }
 372
 373 static void rate_throttle(struct thread_data *td, unsigned long time_spent,
 374                           unsigned int bytes)
 375 {
 376         unsigned long usec_cycle;
 377
 378         if (!td->rate)
 379                 return;
 380
 381         usec_cycle = td->rate_usec_cycle * (bytes / td->min_bs);
 382
 383         if (time_spent < usec_cycle) {
 384                 unsigned long s = usec_cycle - time_spent;
 385
 386                 td->rate_pending_usleep += s;
 387                 if (td->rate_pending_usleep >= 100000) {
 388                         usec_sleep(td, td->rate_pending_usleep);
 389                         td->rate_pending_usleep = 0;
 390                 }
 391         } else {
 392                 long overtime = time_spent - usec_cycle;
 393
 394                 td->rate_pending_usleep -= overtime;
 395         }
 396 }
 397
 398 static int check_min_rate(struct thread_data *td, struct timeval *now)
 399 {
 400         unsigned long spent;
 401         unsigned long rate;
 402         int ddir = td->ddir;
 403
 404         /*
 405          * allow a 2 second settle period in the beginning
 406          */
 407         if (mtime_since(&td->start, now) < 2000)
 408                 return 0;
 409
 410         /*
 411          * if rate blocks is set, sample is running
 412          */
 413         if (td->rate_bytes) {
 414                 spent = mtime_since(&td->lastrate, now);
 415                 if (spent < td->ratecycle)
 416                         return 0;
 417
 418                 rate = (td->this_io_bytes[ddir] - td->rate_bytes) / spent;
 419                 if (rate < td->ratemin) {
 420                         printf("Client%d: min rate %d not met, got %ldKiB/sec\n", td->thread_number, td->ratemin, rate);
 421                         if (rate_quit)
 422                                 terminate_threads(td->groupid);
 423                         return 1;
 424                 }
 425         }
 426
 427         td->rate_bytes = td->this_io_bytes[ddir];
 428         memcpy(&td->lastrate, now, sizeof(*now));
 429         return 0;
 430 }
 431
 432 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 433 {
 434         if (!td->timeout)
 435                 return 0;
 436         if (mtime_since(&td->epoch, t) >= td->timeout * 1000)
 437                 return 1;
 438
 439         return 0;
 440 }
 441
 442 static void fill_random_bytes(struct thread_data *td,
 443                               unsigned char *p, unsigned int len)
 444 {
 445         unsigned int todo;
 446         double r;
 447
 448         while (len) {
 449                 drand48_r(&td->verify_state, &r);
 450
 451                 /*
 452                  * lrand48_r seems to be broken and only fill the bottom
 453                  * 32-bits, even on 64-bit archs with 64-bit longs
 454                  */
 455                 todo = sizeof(r);
 456                 if (todo > len)
 457                         todo = len;
 458
 459                 memcpy(p, &r, todo);
 460
 461                 len -= todo;
 462                 p += todo;
 463         }
 464 }
 465
 466 static void hexdump(void *buffer, int len)
 467 {
 468         unsigned char *p = buffer;
 469         int i;
 470
 471         for (i = 0; i < len; i++)
 472                 printf("%02x", p[i]);
 473         printf("\n");
 474 }
 475
 476 static int verify_io_u_crc32(struct verify_header *hdr, struct io_u *io_u)
 477 {
 478         unsigned char *p = (unsigned char *) io_u->buf;
 479         unsigned long c;
 480         int ret;
 481
 482         p += sizeof(*hdr);
 483         c = crc32(p, hdr->len - sizeof(*hdr));
 484         ret = c != hdr->crc32;
 485
 486         if (ret) {
 487                 fprintf(stderr, "crc32: verify failed at %llu/%u\n", io_u->offset, io_u->buflen);
 488                 fprintf(stderr, "crc32: wanted %lx, got %lx\n", hdr->crc32, c);
 489         }
 490
 491         return ret;
 492 }
 493
 494 static int verify_io_u_md5(struct verify_header *hdr, struct io_u *io_u)
 495 {
 496         unsigned char *p = (unsigned char *) io_u->buf;
 497         struct md5_ctx md5_ctx;
 498         int ret;
 499
 500         memset(&md5_ctx, 0, sizeof(md5_ctx));
 501         p += sizeof(*hdr);
 502         md5_update(&md5_ctx, p, hdr->len - sizeof(*hdr));
 503
 504         ret = memcmp(hdr->md5_digest, md5_ctx.hash, sizeof(md5_ctx.hash));
 505         if (ret) {
 506                 fprintf(stderr, "md5: verify failed at %llu/%u\n", io_u->offset, io_u->buflen);
 507                 hexdump(hdr->md5_digest, sizeof(hdr->md5_digest));
 508                 hexdump(md5_ctx.hash, sizeof(md5_ctx.hash));
 509         }
 510
 511         return ret;
 512 }
 513
 514 static int verify_io_u(struct io_u *io_u)
 515 {
 516         struct verify_header *hdr = (struct verify_header *) io_u->buf;
 517         int ret;
 518
 519         if (hdr->fio_magic != FIO_HDR_MAGIC)
 520                 return 1;
 521
 522         if (hdr->verify_type == VERIFY_MD5)
 523                 ret = verify_io_u_md5(hdr, io_u);
 524         else if (hdr->verify_type == VERIFY_CRC32)
 525                 ret = verify_io_u_crc32(hdr, io_u);
 526         else {
 527                 fprintf(stderr, "Bad verify type %d\n", hdr->verify_type);
 528                 ret = 1;
 529         }
 530
 531         return ret;
 532 }
 533
 534 static void fill_crc32(struct verify_header *hdr, void *p, unsigned int len)
 535 {
 536         hdr->crc32 = crc32(p, len);
 537 }
 538
 539 static void fill_md5(struct verify_header *hdr, void *p, unsigned int len)
 540 {
 541         struct md5_ctx md5_ctx;
 542
 543         memset(&md5_ctx, 0, sizeof(md5_ctx));
 544         md5_update(&md5_ctx, p, len);
 545         memcpy(hdr->md5_digest, md5_ctx.hash, sizeof(md5_ctx.hash));
 546 }
 547
 548 static int get_rw_ddir(struct thread_data *td)
 549 {
 550         if (td_rw(td)) {
 551                 struct timeval now;
 552                 unsigned long elapsed;
 553
 554                 gettimeofday(&now, NULL);
 555                 elapsed = mtime_since_now(&td->rwmix_switch);
 556
 557                 /*
 558                  * Check if it's time to seed a new data direction.
 559                  */
 560                 if (elapsed >= td->rwmixcycle) {
 561                         unsigned long v;
 562                         long r;
 563
 564                         lrand48_r(&td->random_state, &r);
 565                         v = 100UL * r / (unsigned long) (RAND_MAX + 1.0);
 566                         if (v < td->rwmixread)
 567                                 td->rwmix_ddir = DDIR_READ;
 568                         else
 569                                 td->rwmix_ddir = DDIR_WRITE;
 570                         memcpy(&td->rwmix_switch, &now, sizeof(now));
 571                 }
 572                 return td->rwmix_ddir;
 573         } else if (td_read(td))
 574                 return DDIR_READ;
 575         else
 576                 return DDIR_WRITE;
 577 }
 578
 579 /*
 580  * fill body of io_u->buf with random data and add a header with the
 581  * (eg) sha1sum of that data.
 582  */
 583 static void populate_io_u(struct thread_data *td, struct io_u *io_u)
 584 {
 585         unsigned char *p = (unsigned char *) io_u->buf;
 586         struct verify_header hdr;
 587
 588         hdr.fio_magic = FIO_HDR_MAGIC;
 589         hdr.len = io_u->buflen;
 590         p += sizeof(hdr);
 591         fill_random_bytes(td, p, io_u->buflen - sizeof(hdr));
 592
 593         if (td->verify == VERIFY_MD5) {
 594                 fill_md5(&hdr, p, io_u->buflen - sizeof(hdr));
 595                 hdr.verify_type = VERIFY_MD5;
 596         } else {
 597                 fill_crc32(&hdr, p, io_u->buflen - sizeof(hdr));
 598                 hdr.verify_type = VERIFY_CRC32;
 599         }
 600
 601         memcpy(io_u->buf, &hdr, sizeof(hdr));
 602 }
 603
 604 static int td_io_prep(struct thread_data *td, struct io_u *io_u)
 605 {
 606         if (td->io_prep && td->io_prep(td, io_u))
 607                 return 1;
 608
 609         return 0;
 610 }
 611
 612 void put_io_u(struct thread_data *td, struct io_u *io_u)
 613 {
 614         list_del(&io_u->list);
 615         list_add(&io_u->list, &td->io_u_freelist);
 616         td->cur_depth--;
 617 }
 618
 619 static void write_iolog_put(struct thread_data *td, struct io_u *io_u)
 620 {
 621         fprintf(td->iolog_f, "%d,%llu,%u\n", io_u->ddir, io_u->offset, io_u->buflen);
 622 }
 623
 624 static int read_iolog_get(struct thread_data *td, struct io_u *io_u)
 625 {
 626         struct io_piece *ipo;
 627
 628         if (!list_empty(&td->io_log_list)) {
 629                 ipo = list_entry(td->io_log_list.next, struct io_piece, list);
 630                 list_del(&ipo->list);
 631                 io_u->offset = ipo->offset;
 632                 io_u->buflen = ipo->len;
 633                 io_u->ddir = ipo->ddir;
 634                 free(ipo);
 635                 return 0;
 636         }
 637
 638         return 1;
 639 }
 640
 641 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
 642 {
 643         /*
 644          * If using an iolog, grab next piece if any available.
 645          */
 646         if (td->read_iolog)
 647                 return read_iolog_get(td, io_u);
 648
 649         /*
 650          * No log, let the seq/rand engine retrieve the next position.
 651          */
 652         if (!get_next_offset(td, &io_u->offset)) {
 653                 io_u->buflen = get_next_buflen(td);
 654
 655                 if (io_u->buflen) {
 656                         io_u->ddir = get_rw_ddir(td);
 657
 658                         /*
 659                          * If using a write iolog, store this entry.
 660                          */
 661                         if (td->write_iolog)
 662                                 write_iolog_put(td, io_u);
 663
 664                         return 0;
 665                 }
 666         }
 667
 668         return 1;
 669 }
 670
 671 #define queue_full(td)  (list_empty(&(td)->io_u_freelist))
 672
 673 struct io_u *__get_io_u(struct thread_data *td)
 674 {
 675         struct io_u *io_u;
 676
 677         if (queue_full(td))
 678                 return NULL;
 679
 680         io_u = list_entry(td->io_u_freelist.next, struct io_u, list);
 681         io_u->error = 0;
 682         io_u->resid = 0;
 683         list_del(&io_u->list);
 684         list_add(&io_u->list, &td->io_u_busylist);
 685         td->cur_depth++;
 686         return io_u;
 687 }
 688
 689 static struct io_u *get_io_u(struct thread_data *td)
 690 {
 691         struct io_u *io_u;
 692
 693         io_u = __get_io_u(td);
 694         if (!io_u)
 695                 return NULL;
 696
 697         if (td->zone_bytes >= td->zone_size) {
 698                 td->zone_bytes = 0;
 699                 td->last_pos += td->zone_skip;
 700         }
 701
 702         if (fill_io_u(td, io_u)) {
 703                 put_io_u(td, io_u);
 704                 return NULL;
 705         }
 706
 707         if (io_u->buflen + io_u->offset > td->real_file_size)
 708                 io_u->buflen = td->real_file_size - io_u->offset;
 709
 710         if (!io_u->buflen) {
 711                 put_io_u(td, io_u);
 712                 return NULL;
 713         }
 714
 715         if (!td->read_iolog && !td->sequential)
 716                 mark_random_map(td, io_u);
 717
 718         td->last_pos += io_u->buflen;
 719
 720         if (td->verify != VERIFY_NONE)
 721                 populate_io_u(td, io_u);
 722
 723         if (td_io_prep(td, io_u)) {
 724                 put_io_u(td, io_u);
 725                 return NULL;
 726         }
 727
 728         gettimeofday(&io_u->start_time, NULL);
 729         return io_u;
 730 }
 731
 732 static inline void td_set_runstate(struct thread_data *td, int runstate)
 733 {
 734         td->old_runstate = td->runstate;
 735         td->runstate = runstate;
 736 }
 737
 738 static int get_next_verify(struct thread_data *td, struct io_u *io_u)
 739 {
 740         struct io_piece *ipo;
 741
 742         if (list_empty(&td->io_hist_list))
 743                 return 1;
 744
 745         ipo = list_entry(td->io_hist_list.next, struct io_piece, list);
 746         list_del(&ipo->list);
 747
 748         io_u->offset = ipo->offset;
 749         io_u->buflen = ipo->len;
 750         io_u->ddir = DDIR_READ;
 751         free(ipo);
 752         return 0;
 753 }
 754
 755 static void prune_io_piece_log(struct thread_data *td)
 756 {
 757         struct io_piece *ipo;
 758
 759         while (!list_empty(&td->io_hist_list)) {
 760                 ipo = list_entry(td->io_hist_list.next, struct io_piece, list);
 761
 762                 list_del(&ipo->list);
 763                 free(ipo);
 764         }
 765 }
 766
 767 /*
 768  * log a succesful write, so we can unwind the log for verify
 769  */
 770 static void log_io_piece(struct thread_data *td, struct io_u *io_u)
 771 {
 772         struct io_piece *ipo = malloc(sizeof(struct io_piece));
 773         struct list_head *entry;
 774
 775         INIT_LIST_HEAD(&ipo->list);
 776         ipo->offset = io_u->offset;
 777         ipo->len = io_u->buflen;
 778
 779         /*
 780          * for random io where the writes extend the file, it will typically
 781          * be laid out with the block scattered as written. it's faster to
 782          * read them in in that order again, so don't sort
 783          */
 784         if (td->sequential || !td->overwrite) {
 785                 list_add_tail(&ipo->list, &td->io_hist_list);
 786                 return;
 787         }
 788
 789         /*
 790          * for random io, sort the list so verify will run faster
 791          */
 792         entry = &td->io_hist_list;
 793         while ((entry = entry->prev) != &td->io_hist_list) {
 794                 struct io_piece *__ipo = list_entry(entry, struct io_piece, list);
 795
 796                 if (__ipo->offset < ipo->offset)
 797                         break;
 798         }
 799
 800         list_add(&ipo->list, entry);
 801 }
 802
 803 static void write_iolog_close(struct thread_data *td)
 804 {
 805         fflush(td->iolog_f);
 806         fclose(td->iolog_f);
 807         free(td->iolog_buf);
 808 }
 809
 810 static int init_iolog(struct thread_data *td)
 811 {
 812         unsigned long long offset;
 813         unsigned int bytes;
 814         char *str, *p;
 815         FILE *f;
 816         int rw, i, reads, writes;
 817
 818         if (!td->read_iolog && !td->write_iolog)
 819                 return 0;
 820
 821         if (td->read_iolog)
 822                 f = fopen(td->iolog_file, "r");
 823         else
 824                 f = fopen(td->iolog_file, "w");
 825
 826         if (!f) {
 827                 perror("fopen iolog");
 828                 printf("file %s, %d/%d\n", td->iolog_file, td->read_iolog, td->write_iolog);
 829                 return 1;
 830         }
 831
 832         /*
 833          * That's it for writing, setup a log buffer and we're done.
 834           */
 835         if (td->write_iolog) {
 836                 td->iolog_f = f;
 837                 td->iolog_buf = malloc(8192);
 838                 setvbuf(f, td->iolog_buf, _IOFBF, 8192);
 839                 return 0;
 840         }
 841
 842         /*
 843          * Read in the read iolog and store it, reuse the infrastructure
 844          * for doing verifications.
 845          */
 846         str = malloc(4096);
 847         reads = writes = i = 0;
 848         while ((p = fgets(str, 4096, f)) != NULL) {
 849                 struct io_piece *ipo;
 850
 851                 if (sscanf(p, "%d,%llu,%u", &rw, &offset, &bytes) != 3) {
 852                         fprintf(stderr, "bad iolog: %s\n", p);
 853                         continue;
 854                 }
 855                 if (rw == DDIR_READ)
 856                         reads++;
 857                 else if (rw == DDIR_WRITE)
 858                         writes++;
 859                 else {
 860                         fprintf(stderr, "bad ddir: %d\n", rw);
 861                         continue;
 862                 }
 863
 864                 ipo = malloc(sizeof(*ipo));
 865                 INIT_LIST_HEAD(&ipo->list);
 866                 ipo->offset = offset;
 867                 ipo->len = bytes;
 868                 if (bytes > td->max_bs)
 869                         td->max_bs = bytes;
 870                 ipo->ddir = rw;
 871                 list_add_tail(&ipo->list, &td->io_log_list);
 872                 i++;
 873         }
 874
 875         free(str);
 876         fclose(f);
 877
 878         if (!i)
 879                 return 1;
 880
 881         if (reads && !writes)
 882                 td->ddir = DDIR_READ;
 883         else if (!reads && writes)
 884                 td->ddir = DDIR_READ;
 885         else
 886                 td->iomix = 1;
 887
 888         return 0;
 889 }
 890
 891 static int sync_td(struct thread_data *td)
 892 {
 893         if (td->io_sync)
 894                 return td->io_sync(td);
 895
 896         return 0;
 897 }
 898
 899 static int io_u_getevents(struct thread_data *td, int min, int max,
 900                           struct timespec *t)
 901 {
 902         return td->io_getevents(td, min, max, t);
 903 }
 904
 905 static int io_u_queue(struct thread_data *td, struct io_u *io_u)
 906 {
 907         gettimeofday(&io_u->issue_time, NULL);
 908
 909         return td->io_queue(td, io_u);
 910 }
 911
 912 #define iocb_time(iocb) ((unsigned long) (iocb)->data)
 913
 914 static void io_completed(struct thread_data *td, struct io_u *io_u,
 915                          struct io_completion_data *icd)
 916 {
 917         struct timeval e;
 918         unsigned long msec;
 919
 920         gettimeofday(&e, NULL);
 921
 922         if (!io_u->error) {
 923                 unsigned int bytes = io_u->buflen - io_u->resid;
 924                 const int idx = io_u->ddir;
 925
 926                 td->io_blocks[idx]++;
 927                 td->io_bytes[idx] += bytes;
 928                 td->zone_bytes += bytes;
 929                 td->this_io_bytes[idx] += bytes;
 930
 931                 msec = mtime_since(&io_u->issue_time, &e);
 932
 933                 add_clat_sample(td, idx, msec);
 934                 add_bw_sample(td, idx);
 935
 936                 if ((td_rw(td) || td_write(td)) && idx == DDIR_WRITE)
 937                         log_io_piece(td, io_u);
 938
 939                 icd->bytes_done[idx] += bytes;
 940         } else
 941                 icd->error = io_u->error;
 942 }
 943
 944 static void ios_completed(struct thread_data *td,struct io_completion_data *icd)
 945 {
 946         struct io_u *io_u;
 947         int i;
 948
 949         icd->error = 0;
 950         icd->bytes_done[0] = icd->bytes_done[1] = 0;
 951
 952         for (i = 0; i < icd->nr; i++) {
 953                 io_u = td->io_event(td, i);
 954
 955                 io_completed(td, io_u, icd);
 956                 put_io_u(td, io_u);
 957         }
 958 }
 959
 960 static void cleanup_pending_aio(struct thread_data *td)
 961 {
 962         struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
 963         struct list_head *entry, *n;
 964         struct io_completion_data icd;
 965         struct io_u *io_u;
 966         int r;
 967
 968         /*
 969          * get immediately available events, if any
 970          */
 971         r = io_u_getevents(td, 0, td->cur_depth, &ts);
 972         if (r > 0) {
 973                 icd.nr = r;
 974                 ios_completed(td, &icd);
 975         }
 976
 977         /*
 978          * now cancel remaining active events
 979          */
 980         if (td->io_cancel) {
 981                 list_for_each_safe(entry, n, &td->io_u_busylist) {
 982                         io_u = list_entry(entry, struct io_u, list);
 983
 984                         r = td->io_cancel(td, io_u);
 985                         if (!r)
 986                                 put_io_u(td, io_u);
 987                 }
 988         }
 989
 990         if (td->cur_depth) {
 991                 r = io_u_getevents(td, td->cur_depth, td->cur_depth, NULL);
 992                 if (r > 0) {
 993                         icd.nr = r;
 994                         ios_completed(td, &icd);
 995                 }
 996         }
 997 }
 998
 999 static int do_io_u_verify(struct thread_data *td, struct io_u **io_u)
1000 {
1001         struct io_u *v_io_u = *io_u;
1002         int ret = 0;
1003
1004         if (v_io_u) {
1005                 ret = verify_io_u(v_io_u);
1006                 put_io_u(td, v_io_u);
1007                 *io_u = NULL;
1008         }
1009
1010         return ret;
1011 }
1012
1013 static void do_verify(struct thread_data *td)
1014 {
1015         struct timeval t;
1016         struct io_u *io_u, *v_io_u = NULL;
1017         struct io_completion_data icd;
1018         int ret;
1019
1020         td_set_runstate(td, TD_VERIFYING);
1021
1022         do {
1023                 if (td->terminate)
1024                         break;
1025
1026                 gettimeofday(&t, NULL);
1027                 if (runtime_exceeded(td, &t))
1028                         break;
1029
1030                 io_u = __get_io_u(td);
1031                 if (!io_u)
1032                         break;
1033
1034                 if (get_next_verify(td, io_u)) {
1035                         put_io_u(td, io_u);
1036                         break;
1037                 }
1038
1039                 if (td_io_prep(td, io_u)) {
1040                         put_io_u(td, io_u);
1041                         break;
1042                 }
1043
1044                 ret = io_u_queue(td, io_u);
1045                 if (ret) {
1046                         put_io_u(td, io_u);
1047                         td_verror(td, ret);
1048                         break;
1049                 }
1050
1051                 /*
1052                  * we have one pending to verify, do that while
1053                  * we are doing io on the next one
1054                  */
1055                 if (do_io_u_verify(td, &v_io_u))
1056                         break;
1057
1058                 ret = io_u_getevents(td, 1, 1, NULL);
1059                 if (ret != 1) {
1060                         if (ret < 0)
1061                                 td_verror(td, ret);
1062                         break;
1063                 }
1064
1065                 v_io_u = td->io_event(td, 0);
1066                 icd.nr = 1;
1067                 icd.error = 0;
1068                 io_completed(td, v_io_u, &icd);
1069
1070                 if (icd.error) {
1071                         td_verror(td, icd.error);
1072                         put_io_u(td, v_io_u);
1073                         v_io_u = NULL;
1074                         break;
1075                 }
1076
1077                 /*
1078                  * if we can't submit more io, we need to verify now
1079                  */
1080                 if (queue_full(td) && do_io_u_verify(td, &v_io_u))
1081                         break;
1082
1083         } while (1);
1084
1085         do_io_u_verify(td, &v_io_u);
1086
1087         if (td->cur_depth)
1088                 cleanup_pending_aio(td);
1089
1090         td_set_runstate(td, TD_RUNNING);
1091 }
1092
1093 static void do_io(struct thread_data *td)
1094 {
1095         struct io_completion_data icd;
1096         struct timeval s, e;
1097         unsigned long usec;
1098
1099         while (td->this_io_bytes[td->ddir] < td->io_size) {
1100                 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
1101                 struct timespec *timeout;
1102                 int ret, min_evts = 0;
1103                 struct io_u *io_u;
1104
1105                 if (td->terminate)
1106                         break;
1107
1108                 io_u = get_io_u(td);
1109                 if (!io_u)
1110                         break;
1111
1112                 memcpy(&s, &io_u->start_time, sizeof(s));
1113
1114                 ret = io_u_queue(td, io_u);
1115                 if (ret) {
1116                         put_io_u(td, io_u);
1117                         td_verror(td, ret);
1118                         break;
1119                 }
1120
1121                 add_slat_sample(td, io_u->ddir, mtime_since(&io_u->start_time, &io_u->issue_time));
1122
1123                 if (td->cur_depth < td->iodepth) {
1124                         timeout = &ts;
1125                         min_evts = 0;
1126                 } else {
1127                         timeout = NULL;
1128                         min_evts = 1;
1129                 }
1130
1131                 ret = io_u_getevents(td, min_evts, td->cur_depth, timeout);
1132                 if (ret < 0) {
1133                         td_verror(td, ret);
1134                         break;
1135                 } else if (!ret)
1136                         continue;
1137
1138                 icd.nr = ret;
1139                 ios_completed(td, &icd);
1140                 if (icd.error) {
1141                         td_verror(td, icd.error);
1142                         break;
1143                 }
1144
1145                 /*
1146                  * the rate is batched for now, it should work for batches
1147                  * of completions except the very first one which may look
1148                  * a little bursty
1149                  */
1150                 gettimeofday(&e, NULL);
1151                 usec = utime_since(&s, &e);
1152
1153                 rate_throttle(td, usec, icd.bytes_done[td->ddir]);
1154
1155                 if (check_min_rate(td, &e)) {
1156                         td_verror(td, ENOMEM);
1157                         break;
1158                 }
1159
1160                 if (runtime_exceeded(td, &e))
1161                         break;
1162
1163                 if (td->thinktime)
1164                         usec_sleep(td, td->thinktime);
1165
1166                 if (should_fsync(td) && td->fsync_blocks &&
1167                     (td->io_blocks[DDIR_WRITE] % td->fsync_blocks) == 0)
1168                         sync_td(td);
1169         }
1170
1171         if (td->cur_depth)
1172                 cleanup_pending_aio(td);
1173
1174         if (should_fsync(td) && td->end_fsync)
1175                 sync_td(td);
1176 }
1177
1178 static void cleanup_io(struct thread_data *td)
1179 {
1180         if (td->io_cleanup)
1181                 td->io_cleanup(td);
1182 }
1183
1184 static int init_io(struct thread_data *td)
1185 {
1186         if (td->io_engine == FIO_SYNCIO)
1187                 return fio_syncio_init(td);
1188         else if (td->io_engine == FIO_MMAPIO)
1189                 return fio_mmapio_init(td);
1190         else if (td->io_engine == FIO_LIBAIO)
1191                 return fio_libaio_init(td);
1192         else if (td->io_engine == FIO_POSIXAIO)
1193                 return fio_posixaio_init(td);
1194         else if (td->io_engine == FIO_SGIO)
1195                 return fio_sgio_init(td);
1196         else if (td->io_engine == FIO_SPLICEIO)
1197                 return fio_spliceio_init(td);
1198         else {
1199                 fprintf(stderr, "bad io_engine %d\n", td->io_engine);
1200                 return 1;
1201         }
1202 }
1203
1204 static void cleanup_io_u(struct thread_data *td)
1205 {
1206         struct list_head *entry, *n;
1207         struct io_u *io_u;
1208
1209         list_for_each_safe(entry, n, &td->io_u_freelist) {
1210                 io_u = list_entry(entry, struct io_u, list);
1211
1212                 list_del(&io_u->list);
1213                 free(io_u);
1214         }
1215
1216         if (td->mem_type == MEM_MALLOC)
1217                 free(td->orig_buffer);
1218         else if (td->mem_type == MEM_SHM) {
1219                 struct shmid_ds sbuf;
1220
1221                 shmdt(td->orig_buffer);
1222                 shmctl(td->shm_id, IPC_RMID, &sbuf);
1223         } else if (td->mem_type == MEM_MMAP)
1224                 munmap(td->orig_buffer, td->orig_buffer_size);
1225         else
1226                 fprintf(stderr, "Bad memory type %d\n", td->mem_type);
1227
1228         td->orig_buffer = NULL;
1229 }
1230
1231 static int init_io_u(struct thread_data *td)
1232 {
1233         struct io_u *io_u;
1234         int i, max_units;
1235         char *p;
1236
1237         if (td->io_engine & FIO_SYNCIO)
1238                 max_units = 1;
1239         else
1240                 max_units = td->iodepth;
1241
1242         td->orig_buffer_size = td->max_bs * max_units + MASK;
1243
1244         if (td->mem_type == MEM_MALLOC)
1245                 td->orig_buffer = malloc(td->orig_buffer_size);
1246         else if (td->mem_type == MEM_SHM) {
1247                 td->shm_id = shmget(IPC_PRIVATE, td->orig_buffer_size, IPC_CREAT | 0600);
1248                 if (td->shm_id < 0) {
1249                         td_verror(td, errno);
1250                         perror("shmget");
1251                         return 1;
1252                 }
1253
1254                 td->orig_buffer = shmat(td->shm_id, NULL, 0);
1255                 if (td->orig_buffer == (void *) -1) {
1256                         td_verror(td, errno);
1257                         perror("shmat");
1258                         td->orig_buffer = NULL;
1259                         return 1;
1260                 }
1261         } else if (td->mem_type == MEM_MMAP) {
1262                 td->orig_buffer = mmap(NULL, td->orig_buffer_size, PROT_READ | PROT_WRITE, MAP_PRIVATE | OS_MAP_ANON, 0, 0);
1263                 if (td->orig_buffer == MAP_FAILED) {
1264                         td_verror(td, errno);
1265                         perror("mmap");
1266                         td->orig_buffer = NULL;
1267                         return 1;
1268                 }
1269         }
1270
1271         p = ALIGN(td->orig_buffer);
1272         for (i = 0; i < max_units; i++) {
1273                 io_u = malloc(sizeof(*io_u));
1274                 memset(io_u, 0, sizeof(*io_u));
1275                 INIT_LIST_HEAD(&io_u->list);
1276
1277                 io_u->buf = p + td->max_bs * i;
1278                 io_u->index = i;
1279                 list_add(&io_u->list, &td->io_u_freelist);
1280         }
1281
1282         return 0;
1283 }
1284
1285 static void cleanup_allocs(struct thread_data *td)
1286 {
1287         if (td->directory)
1288                 free(td->directory);
1289         if (td->iolog_file)
1290                 free(td->iolog_file);
1291         if (td->exec_prerun)
1292                 free(td->exec_prerun);
1293         if (td->exec_postrun)
1294                 free(td->exec_postrun);
1295         if (td->ioscheduler)
1296                 free(td->ioscheduler);
1297         if (td->sysfs_root)
1298                 free(td->sysfs_root);
1299 }
1300
1301 static int create_file(struct thread_data *td, unsigned long long size,
1302                        int extend)
1303 {
1304         unsigned long long left;
1305         unsigned int bs;
1306         int r, oflags;
1307         char *b;
1308
1309         /*
1310          * unless specifically asked for overwrite, let normal io extend it
1311          */
1312         if (td_write(td) && !td->overwrite)
1313                 return 0;
1314
1315         if (!size) {
1316                 fprintf(stderr, "Need size for create\n");
1317                 td_verror(td, EINVAL);
1318                 return 1;
1319         }
1320
1321         if (!extend) {
1322                 oflags = O_CREAT | O_TRUNC;
1323                 printf("Client%d: Laying out IO file (%LuMiB)\n", td->thread_number, size >> 20);
1324         } else {
1325                 oflags = O_APPEND;
1326                 printf("Client%d: Extending IO file (%Lu -> %LuMiB)\n", td->thread_number, (td->file_size - size) >> 20, td->file_size >> 20);
1327         }
1328
1329         td->fd = open(td->file_name, O_WRONLY | oflags, 0644);
1330         if (td->fd < 0) {
1331                 td_verror(td, errno);
1332                 return 1;
1333         }
1334
1335         if (!extend && ftruncate(td->fd, td->file_size) == -1) {
1336                 td_verror(td, errno);
1337                 return 1;
1338         }
1339
1340         td->io_size = td->file_size;
1341         b = malloc(td->max_bs);
1342         memset(b, 0, td->max_bs);
1343
1344         left = size;
1345         while (left && !td->terminate) {
1346                 bs = td->max_bs;
1347                 if (bs > left)
1348                         bs = left;
1349
1350                 r = write(td->fd, b, bs);
1351
1352                 if (r == (int) bs) {
1353                         left -= bs;
1354                         continue;
1355                 } else {
1356                         if (r < 0)
1357                                 td_verror(td, errno);
1358                         else
1359                                 td_verror(td, EIO);
1360
1361                         break;
1362                 }
1363         }
1364
1365         if (td->terminate)
1366                 unlink(td->file_name);
1367         else if (td->create_fsync)
1368                 fsync(td->fd);
1369
1370         close(td->fd);
1371         td->fd = -1;
1372         free(b);
1373         return 0;
1374 }
1375
1376 static int file_size(struct thread_data *td)
1377 {
1378         struct stat st;
1379
1380         if (fstat(td->fd, &st) == -1) {
1381                 td_verror(td, errno);
1382                 return 1;
1383         }
1384
1385         td->real_file_size = st.st_size;
1386
1387         if (!td->file_size || td->file_size > td->real_file_size)
1388                 td->file_size = td->real_file_size;
1389
1390         td->file_size -= td->file_offset;
1391         return 0;
1392 }
1393
1394 static int bdev_size(struct thread_data *td)
1395 {
1396         unsigned long long bytes;
1397         int r;
1398
1399         r = blockdev_size(td->fd, &bytes);
1400         if (r) {
1401                 td_verror(td, r);
1402                 return 1;
1403         }
1404
1405         td->real_file_size = bytes;
1406
1407         /*
1408          * no extend possibilities, so limit size to device size if too large
1409          */
1410         if (!td->file_size || td->file_size > td->real_file_size)
1411                 td->file_size = td->real_file_size;
1412
1413         td->file_size -= td->file_offset;
1414         return 0;
1415 }
1416
1417 static int get_file_size(struct thread_data *td)
1418 {
1419         int ret = 0;
1420
1421         if (td->filetype == FIO_TYPE_FILE)
1422                 ret = file_size(td);
1423         else if (td->filetype == FIO_TYPE_BD)
1424                 ret = bdev_size(td);
1425         else
1426                 td->real_file_size = -1;
1427
1428         if (ret)
1429                 return ret;
1430
1431         if (td->file_offset > td->real_file_size) {
1432                 fprintf(stderr, "Client%d: offset extends end (%Lu > %Lu)\n", td->thread_number, td->file_offset, td->real_file_size);
1433                 return 1;
1434         }
1435
1436         td->io_size = td->file_size;
1437         if (td->io_size == 0) {
1438                 fprintf(stderr, "Client%d: no io blocks\n", td->thread_number);
1439                 td_verror(td, EINVAL);
1440                 return 1;
1441         }
1442
1443         if (!td->zone_size)
1444                 td->zone_size = td->io_size;
1445
1446         td->total_io_size = td->io_size * td->loops;
1447         return 0;
1448 }
1449
1450 static int setup_file_mmap(struct thread_data *td)
1451 {
1452         int flags;
1453
1454         if (td_rw(td))
1455                 flags = PROT_READ | PROT_WRITE;
1456         else if (td_write(td)) {
1457                 flags = PROT_WRITE;
1458
1459                 if (td->verify != VERIFY_NONE)
1460                         flags |= PROT_READ;
1461         } else
1462                 flags = PROT_READ;
1463
1464         td->mmap = mmap(NULL, td->file_size, flags, MAP_SHARED, td->fd, td->file_offset);
1465         if (td->mmap == MAP_FAILED) {
1466                 td->mmap = NULL;
1467                 td_verror(td, errno);
1468                 return 1;
1469         }
1470
1471         if (td->invalidate_cache) {
1472                 if (madvise(td->mmap, td->file_size, MADV_DONTNEED) < 0) {
1473                         td_verror(td, errno);
1474                         return 1;
1475                 }
1476         }
1477
1478         if (td->sequential) {
1479                 if (madvise(td->mmap, td->file_size, MADV_SEQUENTIAL) < 0) {
1480                         td_verror(td, errno);
1481                         return 1;
1482                 }
1483         } else {
1484                 if (madvise(td->mmap, td->file_size, MADV_RANDOM) < 0) {
1485                         td_verror(td, errno);
1486                         return 1;
1487                 }
1488         }
1489
1490         return 0;
1491 }
1492
1493 static int setup_file_plain(struct thread_data *td)
1494 {
1495         if (td->invalidate_cache) {
1496                 if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_DONTNEED) < 0) {
1497                         td_verror(td, errno);
1498                         return 1;
1499                 }
1500         }
1501
1502         if (td->sequential) {
1503                 if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_SEQUENTIAL) < 0) {
1504                         td_verror(td, errno);
1505                         return 1;
1506                 }
1507         } else {
1508                 if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_RANDOM) < 0) {
1509                         td_verror(td, errno);
1510                         return 1;
1511                 }
1512         }
1513
1514         return 0;
1515 }
1516
1517 static int setup_file(struct thread_data *td)
1518 {
1519         struct stat st;
1520         int flags = 0;
1521
1522         if (stat(td->file_name, &st) == -1) {
1523                 if (errno != ENOENT) {
1524                         td_verror(td, errno);
1525                         return 1;
1526                 }
1527                 if (!td->create_file) {
1528                         td_verror(td, ENOENT);
1529                         return 1;
1530                 }
1531                 if (create_file(td, td->file_size, 0))
1532                         return 1;
1533         } else if (td->filetype == FIO_TYPE_FILE) {
1534                 if (st.st_size < (off_t) td->file_size) {
1535                         if (create_file(td, td->file_size - st.st_size, 1))
1536                                 return 1;
1537                 }
1538         }
1539
1540         if (td->odirect)
1541                 flags |= O_DIRECT;
1542
1543         if (td_write(td) || td_rw(td)) {
1544                 if (td->filetype == FIO_TYPE_FILE) {
1545                         if (!td->overwrite)
1546                                 flags |= O_TRUNC;
1547
1548                         flags |= O_CREAT;
1549                 }
1550                 if (td->sync_io)
1551                         flags |= O_SYNC;
1552
1553                 flags |= O_RDWR;
1554
1555                 td->fd = open(td->file_name, flags, 0600);
1556         } else {
1557                 if (td->filetype == FIO_TYPE_CHAR)
1558                         flags |= O_RDWR;
1559                 else
1560                         flags |= O_RDONLY;
1561
1562                 td->fd = open(td->file_name, flags);
1563         }
1564
1565         if (td->fd == -1) {
1566                 td_verror(td, errno);
1567                 return 1;
1568         }
1569
1570         if (get_file_size(td))
1571                 return 1;
1572
1573         if (td->io_engine != FIO_MMAPIO)
1574                 return setup_file_plain(td);
1575         else
1576                 return setup_file_mmap(td);
1577 }
1578
1579 static int check_dev_match(dev_t dev, char *path)
1580 {
1581         unsigned int major, minor;
1582         char line[256], *p;
1583         FILE *f;
1584
1585         f = fopen(path, "r");
1586         if (!f) {
1587                 perror("open path");
1588                 return 1;
1589         }
1590
1591         p = fgets(line, sizeof(line), f);
1592         if (!p) {
1593                 fclose(f);
1594                 return 1;
1595         }
1596
1597         if (sscanf(p, "%u:%u", &major, &minor) != 2) {
1598                 fclose(f);
1599                 return 1;
1600         }
1601
1602         if (((major << 8) | minor) == dev) {
1603                 fclose(f);
1604                 return 0;
1605         }
1606
1607         fclose(f);
1608         return 1;
1609 }
1610
1611 static int find_block_dir(dev_t dev, char *path)
1612 {
1613         struct dirent *dir;
1614         struct stat st;
1615         int found = 0;
1616         DIR *D;
1617
1618         D = opendir(path);
1619         if (!D)
1620                 return 0;
1621
1622         while ((dir = readdir(D)) != NULL) {
1623                 char full_path[256];
1624
1625                 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1626                         continue;
1627                 if (!strcmp(dir->d_name, "device"))
1628                         continue;
1629
1630                 sprintf(full_path, "%s/%s", path, dir->d_name);
1631
1632                 if (!strcmp(dir->d_name, "dev")) {
1633                         if (!check_dev_match(dev, full_path)) {
1634                                 found = 1;
1635                                 break;
1636                         }
1637                 }
1638
1639                 if (stat(full_path, &st) == -1) {
1640                         perror("stat");
1641                         break;
1642                 }
1643
1644                 if (!S_ISDIR(st.st_mode) || S_ISLNK(st.st_mode))
1645                         continue;
1646
1647                 found = find_block_dir(dev, full_path);
1648                 if (found) {
1649                         strcpy(path, full_path);
1650                         break;
1651                 }
1652         }
1653
1654         closedir(D);
1655         return found;
1656 }
1657
1658 static int get_io_ticks(struct disk_util *du, struct disk_util_stat *dus)
1659 {
1660         unsigned in_flight;
1661         char line[256];
1662         FILE *f;
1663         char *p;
1664
1665         f = fopen(du->path, "r");
1666         if (!f)
1667                 return 1;
1668
1669         p = fgets(line, sizeof(line), f);
1670         if (!p) {
1671                 fclose(f);
1672                 return 1;
1673         }
1674
1675         if (sscanf(p, "%u %u %llu %u %u %u %llu %u %u %u %u\n", &dus->ios[0], &dus->merges[0], &dus->sectors[0], &dus->ticks[0], &dus->ios[1], &dus->merges[1], &dus->sectors[1], &dus->ticks[1], &in_flight, &dus->io_ticks, &dus->time_in_queue) != 11) {
1676                 fclose(f);
1677                 return 1;
1678         }
1679
1680         fclose(f);
1681         return 0;
1682 }
1683
1684 static void update_io_tick_disk(struct disk_util *du)
1685 {
1686         struct disk_util_stat __dus, *dus, *ldus;
1687         struct timeval t;
1688
1689         if (get_io_ticks(du, &__dus))
1690                 return;
1691
1692         dus = &du->dus;
1693         ldus = &du->last_dus;
1694
1695         dus->sectors[0] += (__dus.sectors[0] - ldus->sectors[0]);
1696         dus->sectors[1] += (__dus.sectors[1] - ldus->sectors[1]);
1697         dus->ios[0] += (__dus.ios[0] - ldus->ios[0]);
1698         dus->ios[1] += (__dus.ios[1] - ldus->ios[1]);
1699         dus->merges[0] += (__dus.merges[0] - ldus->merges[0]);
1700         dus->merges[1] += (__dus.merges[1] - ldus->merges[1]);
1701         dus->ticks[0] += (__dus.ticks[0] - ldus->ticks[0]);
1702         dus->ticks[1] += (__dus.ticks[1] - ldus->ticks[1]);
1703         dus->io_ticks += (__dus.io_ticks - ldus->io_ticks);
1704         dus->time_in_queue += (__dus.time_in_queue - ldus->time_in_queue);
1705
1706         gettimeofday(&t, NULL);
1707         du->msec += mtime_since(&du->time, &t);
1708         memcpy(&du->time, &t, sizeof(t));
1709         memcpy(ldus, &__dus, sizeof(__dus));
1710 }
1711
1712 static void update_io_ticks(void)
1713 {
1714         struct list_head *entry;
1715         struct disk_util *du;
1716
1717         list_for_each(entry, &disk_list) {
1718                 du = list_entry(entry, struct disk_util, list);
1719                 update_io_tick_disk(du);
1720         }
1721 }
1722
1723 static int disk_util_exists(dev_t dev)
1724 {
1725         struct list_head *entry;
1726         struct disk_util *du;
1727
1728         list_for_each(entry, &disk_list) {
1729                 du = list_entry(entry, struct disk_util, list);
1730
1731                 if (du->dev == dev)
1732                         return 1;
1733         }
1734
1735         return 0;
1736 }
1737
1738 static void disk_util_add(dev_t dev, char *path)
1739 {
1740         struct disk_util *du = malloc(sizeof(*du));
1741
1742         memset(du, 0, sizeof(*du));
1743         INIT_LIST_HEAD(&du->list);
1744         sprintf(du->path, "%s/stat", path);
1745         du->name = strdup(basename(path));
1746         du->dev = dev;
1747
1748         gettimeofday(&du->time, NULL);
1749         get_io_ticks(du, &du->last_dus);
1750
1751         list_add_tail(&du->list, &disk_list);
1752 }
1753
1754 static void init_disk_util(struct thread_data *td)
1755 {
1756         struct stat st;
1757         char foo[256], tmp[256];
1758         dev_t dev;
1759         char *p;
1760
1761         if (!td->do_disk_util)
1762                 return;
1763
1764         if (!stat(td->file_name, &st)) {
1765                 if (S_ISBLK(st.st_mode))
1766                         dev = st.st_rdev;
1767                 else
1768                         dev = st.st_dev;
1769         } else {
1770                 /*
1771                  * must be a file, open "." in that path
1772                  */
1773                 strcpy(foo, td->file_name);
1774                 p = dirname(foo);
1775                 if (stat(p, &st)) {
1776                         perror("disk util stat");
1777                         return;
1778                 }
1779
1780                 dev = st.st_dev;
1781         }
1782
1783         if (disk_util_exists(dev))
1784                 return;
1785
1786         sprintf(foo, "/sys/block");
1787         if (!find_block_dir(dev, foo))
1788                 return;
1789
1790         /*
1791          * If there's a ../queue/ directory there, we are inside a partition.
1792          * Check if that is the case and jump back. For loop/md/dm etc we
1793          * are already in the right spot.
1794          */
1795         sprintf(tmp, "%s/../queue", foo);
1796         if (!stat(tmp, &st)) {
1797                 p = dirname(foo);
1798                 sprintf(tmp, "%s/queue", p);
1799                 if (stat(tmp, &st)) {
1800                         fprintf(stderr, "unknown sysfs layout\n");
1801                         return;
1802                 }
1803                 sprintf(foo, "%s", p);
1804         }
1805
1806         td->sysfs_root = strdup(foo);
1807         disk_util_add(dev, foo);
1808 }
1809
1810 static int switch_ioscheduler(struct thread_data *td)
1811 {
1812         char tmp[256], tmp2[128];
1813         FILE *f;
1814         int ret;
1815
1816         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
1817
1818         f = fopen(tmp, "r+");
1819         if (!f) {
1820                 td_verror(td, errno);
1821                 return 1;
1822         }
1823
1824         /*
1825          * Set io scheduler.
1826          */
1827         ret = fwrite(td->ioscheduler, strlen(td->ioscheduler), 1, f);
1828         if (ferror(f) || ret != 1) {
1829                 td_verror(td, errno);
1830                 fclose(f);
1831                 return 1;
1832         }
1833
1834         rewind(f);
1835
1836         /*
1837          * Read back and check that the selected scheduler is now the default.
1838          */
1839         ret = fread(tmp, 1, sizeof(tmp), f);
1840         if (ferror(f) || ret < 0) {
1841                 td_verror(td, errno);
1842                 fclose(f);
1843                 return 1;
1844         }
1845
1846         sprintf(tmp2, "[%s]", td->ioscheduler);
1847         if (!strstr(tmp, tmp2)) {
1848                 fprintf(stderr, "fio: io scheduler %s not found\n", td->ioscheduler);
1849                 td_verror(td, EINVAL);
1850                 fclose(f);
1851                 return 1;
1852         }
1853
1854         fclose(f);
1855         return 0;
1856 }
1857
1858 static void disk_util_timer_arm(void)
1859 {
1860         itimer.it_value.tv_sec = 0;
1861         itimer.it_value.tv_usec = DISK_UTIL_MSEC * 1000;
1862         setitimer(ITIMER_REAL, &itimer, NULL);
1863 }
1864
1865 static void clear_io_state(struct thread_data *td)
1866 {
1867         if (td->io_engine == FIO_SYNCIO)
1868                 lseek(td->fd, SEEK_SET, 0);
1869
1870         td->last_pos = 0;
1871         td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0;
1872         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
1873         td->zone_bytes = 0;
1874
1875         if (td->file_map)
1876                 memset(td->file_map, 0, td->num_maps * sizeof(long));
1877 }
1878
1879 static void update_rusage_stat(struct thread_data *td)
1880 {
1881         if (!(td->runtime[0] + td->runtime[1]))
1882                 return;
1883
1884         getrusage(RUSAGE_SELF, &td->ru_end);
1885
1886         td->usr_time += mtime_since(&td->ru_start.ru_utime, &td->ru_end.ru_utime);
1887         td->sys_time += mtime_since(&td->ru_start.ru_stime, &td->ru_end.ru_stime);
1888         td->ctx += td->ru_end.ru_nvcsw + td->ru_end.ru_nivcsw - (td->ru_start.ru_nvcsw + td->ru_start.ru_nivcsw);
1889
1890
1891         memcpy(&td->ru_start, &td->ru_end, sizeof(td->ru_end));
1892 }
1893
1894 static void *thread_main(void *data)
1895 {
1896         struct thread_data *td = data;
1897
1898         if (!td->use_thread)
1899                 setsid();
1900
1901         td->pid = getpid();
1902
1903         INIT_LIST_HEAD(&td->io_u_freelist);
1904         INIT_LIST_HEAD(&td->io_u_busylist);
1905         INIT_LIST_HEAD(&td->io_hist_list);
1906         INIT_LIST_HEAD(&td->io_log_list);
1907
1908         if (init_io_u(td))
1909                 goto err;
1910
1911         if (fio_setaffinity(td) == -1) {
1912                 td_verror(td, errno);
1913                 goto err;
1914         }
1915
1916         if (init_io(td))
1917                 goto err;
1918
1919         if (init_iolog(td))
1920                 goto err;
1921
1922         if (td->ioprio) {
1923                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1924                         td_verror(td, errno);
1925                         goto err;
1926                 }
1927         }
1928
1929         if (nice(td->nice) < 0) {
1930                 td_verror(td, errno);
1931                 goto err;
1932         }
1933
1934         if (init_random_state(td))
1935                 goto err;
1936
1937         if (td->ioscheduler && switch_ioscheduler(td))
1938                 goto err;
1939
1940         td_set_runstate(td, TD_INITIALIZED);
1941         sem_post(&startup_sem);
1942         sem_wait(&td->mutex);
1943
1944         if (!td->create_serialize && setup_file(td))
1945                 goto err;
1946
1947         gettimeofday(&td->epoch, NULL);
1948
1949         if (td->exec_prerun)
1950                 system(td->exec_prerun);
1951
1952         while (td->loops--) {
1953                 getrusage(RUSAGE_SELF, &td->ru_start);
1954                 gettimeofday(&td->start, NULL);
1955                 memcpy(&td->stat_sample_time, &td->start, sizeof(td->start));
1956
1957                 if (td->ratemin)
1958                         memcpy(&td->lastrate, &td->stat_sample_time, sizeof(td->lastrate));
1959
1960                 clear_io_state(td);
1961                 prune_io_piece_log(td);
1962
1963                 do_io(td);
1964
1965                 td->runtime[td->ddir] += mtime_since_now(&td->start);
1966                 if (td_rw(td) && td->io_bytes[td->ddir ^ 1])
1967                         td->runtime[td->ddir ^ 1] = td->runtime[td->ddir];
1968
1969                 update_rusage_stat(td);
1970
1971                 if (td->error || td->terminate)
1972                         break;
1973
1974                 if (td->verify == VERIFY_NONE)
1975                         continue;
1976
1977                 clear_io_state(td);
1978                 gettimeofday(&td->start, NULL);
1979
1980                 do_verify(td);
1981
1982                 td->runtime[DDIR_READ] += mtime_since_now(&td->start);
1983
1984                 if (td->error || td->terminate)
1985                         break;
1986         }
1987
1988         if (td->bw_log)
1989                 finish_log(td, td->bw_log, "bw");
1990         if (td->slat_log)
1991                 finish_log(td, td->slat_log, "slat");
1992         if (td->clat_log)
1993                 finish_log(td, td->clat_log, "clat");
1994         if (td->write_iolog)
1995                 write_iolog_close(td);
1996         if (td->exec_postrun)
1997                 system(td->exec_postrun);
1998
1999         if (exitall_on_terminate)
2000                 terminate_threads(td->groupid);
2001
2002 err:
2003         if (td->fd != -1) {
2004                 close(td->fd);
2005                 td->fd = -1;
2006         }
2007         if (td->mmap)
2008                 munmap(td->mmap, td->file_size);
2009         cleanup_allocs(td);
2010         cleanup_io(td);
2011         cleanup_io_u(td);
2012         td_set_runstate(td, TD_EXITED);
2013         return NULL;
2014
2015 }
2016
2017 static void *fork_main(int shmid, int offset)
2018 {
2019         struct thread_data *td;
2020         void *data;
2021
2022         data = shmat(shmid, NULL, 0);
2023         if (data == (void *) -1) {
2024                 perror("shmat");
2025                 return NULL;
2026         }
2027
2028         td = data + offset * sizeof(struct thread_data);
2029         thread_main(td);
2030         shmdt(data);
2031         return NULL;
2032 }
2033
2034 static int calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max,
2035                     double *mean, double *dev)
2036 {
2037         double n;
2038
2039         if (is->samples == 0)
2040                 return 0;
2041
2042         *min = is->min_val;
2043         *max = is->max_val;
2044
2045         n = (double) is->samples;
2046         *mean = (double) is->val / n;
2047         *dev = sqrt(((double) is->val_sq - (*mean * *mean) / n) / (n - 1));
2048         if (!(*min + *max) && !(*mean + *dev))
2049                 return 0;
2050
2051         return 1;
2052 }
2053
2054 static void show_ddir_status(struct thread_data *td, struct group_run_stats *rs,
2055                              int ddir)
2056 {
2057         char *ddir_str[] = { "read ", "write" };
2058         unsigned long min, max;
2059         unsigned long long bw;
2060         double mean, dev;
2061
2062         if (!td->runtime[ddir])
2063                 return;
2064
2065         bw = td->io_bytes[ddir] / td->runtime[ddir];
2066         printf("  %s: io=%6lluMiB, bw=%6lluKiB/s, runt=%6lumsec\n", ddir_str[ddir], td->io_bytes[ddir] >> 20, bw, td->runtime[ddir]);
2067
2068         if (calc_lat(&td->slat_stat[ddir], &min, &max, &mean, &dev))
2069                 printf("    slat (msec): min=%5lu, max=%5lu, avg=%5.02f, dev=%5.02f\n", min, max, mean, dev);
2070
2071         if (calc_lat(&td->clat_stat[ddir], &min, &max, &mean, &dev))
2072                 printf("    clat (msec): min=%5lu, max=%5lu, avg=%5.02f, dev=%5.02f\n", min, max, mean, dev);
2073
2074         if (calc_lat(&td->bw_stat[ddir], &min, &max, &mean, &dev)) {
2075                 double p_of_agg;
2076
2077                 p_of_agg = mean * 100 / (double) rs->agg[ddir];
2078                 printf("    bw (KiB/s) : min=%5lu, max=%5lu, per=%3.2f%%, avg=%5.02f, dev=%5.02f\n", min, max, p_of_agg, mean, dev);
2079         }
2080 }
2081
2082 static void show_thread_status(struct thread_data *td,
2083                                struct group_run_stats *rs)
2084 {
2085         double usr_cpu, sys_cpu;
2086
2087         if (!(td->io_bytes[0] + td->io_bytes[1]) && !td->error)
2088                 return;
2089
2090         printf("Client%d (groupid=%d): err=%2d:\n", td->thread_number, td->groupid, td->error);
2091
2092         show_ddir_status(td, rs, td->ddir);
2093         if (td->io_bytes[td->ddir ^ 1])
2094                 show_ddir_status(td, rs, td->ddir ^ 1);
2095
2096         if (td->runtime[0] + td->runtime[1]) {
2097                 double runt = td->runtime[0] + td->runtime[1];
2098
2099                 usr_cpu = (double) td->usr_time * 100 / runt;
2100                 sys_cpu = (double) td->sys_time * 100 / runt;
2101         } else {
2102                 usr_cpu = 0;
2103                 sys_cpu = 0;
2104         }
2105
2106         printf("  cpu          : usr=%3.2f%%, sys=%3.2f%%, ctx=%lu\n", usr_cpu, sys_cpu, td->ctx);
2107 }
2108
2109 static void check_str_update(struct thread_data *td)
2110 {
2111         char c = run_str[td->thread_number - 1];
2112
2113         if (td->runstate == td->old_runstate)
2114                 return;
2115
2116         switch (td->runstate) {
2117                 case TD_REAPED:
2118                         c = '_';
2119                         break;
2120                 case TD_EXITED:
2121                         c = 'E';
2122                         break;
2123                 case TD_RUNNING:
2124                         if (td_rw(td)) {
2125                                 if (td->sequential)
2126                                         c = 'M';
2127                                 else
2128                                         c = 'm';
2129                         } else if (td_read(td)) {
2130                                 if (td->sequential)
2131                                         c = 'R';
2132                                 else
2133                                         c = 'r';
2134                         } else {
2135                                 if (td->sequential)
2136                                         c = 'W';
2137                                 else
2138                                         c = 'w';
2139                         }
2140                         break;
2141                 case TD_VERIFYING:
2142                         c = 'V';
2143                         break;
2144                 case TD_CREATED:
2145                         c = 'C';
2146                         break;
2147                 case TD_INITIALIZED:
2148                         c = 'I';
2149                         break;
2150                 case TD_NOT_CREATED:
2151                         c = 'P';
2152                         break;
2153                 default:
2154                         printf("state %d\n", td->runstate);
2155         }
2156
2157         run_str[td->thread_number - 1] = c;
2158         td->old_runstate = td->runstate;
2159 }
2160
2161 static void eta_to_str(char *str, int eta_sec)
2162 {
2163         unsigned int d, h, m, s;
2164         static int always_d, always_h;
2165
2166         d = h = m = s = 0;
2167
2168         s = eta_sec % 60;
2169         eta_sec /= 60;
2170         m = eta_sec % 60;
2171         eta_sec /= 60;
2172         h = eta_sec % 24;
2173         eta_sec /= 24;
2174         d = eta_sec;
2175
2176         if (d || always_d) {
2177                 always_d = 1;
2178                 str += sprintf(str, "%02dd:", d);
2179         }
2180         if (h || always_h) {
2181                 always_h = 1;
2182                 str += sprintf(str, "%02dh:", h);
2183         }
2184
2185         str += sprintf(str, "%02dm:", m);
2186         str += sprintf(str, "%02ds", s);
2187 }
2188
2189 static int thread_eta(struct thread_data *td, unsigned long elapsed)
2190 {
2191         unsigned long long bytes_total, bytes_done;
2192         unsigned int eta_sec = 0;
2193
2194         bytes_total = td->total_io_size;
2195
2196         /*
2197          * if writing, bytes_total will be twice the size. If mixing,
2198          * assume a 50/50 split and thus bytes_total will be 50% larger.
2199          */
2200         if (td->verify) {
2201                 if (td_rw(td))
2202                         bytes_total = bytes_total * 3 / 2;
2203                 else
2204                         bytes_total <<= 1;
2205         }
2206         if (td->zone_size && td->zone_skip)
2207                 bytes_total /= (td->zone_skip / td->zone_size);
2208
2209         if (td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING) {
2210                 double perc;
2211
2212                 bytes_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE];
2213                 perc = (double) bytes_done / (double) bytes_total;
2214                 if (perc > 1.0)
2215                         perc = 1.0;
2216
2217                 eta_sec = (elapsed * (1.0 / perc)) - elapsed;
2218
2219                 if (td->timeout && eta_sec > (td->timeout - elapsed))
2220                         eta_sec = td->timeout - elapsed;
2221         } else if (td->runstate == TD_NOT_CREATED || td->runstate == TD_CREATED
2222                         || td->runstate == TD_INITIALIZED) {
2223                 int t_eta = 0, r_eta = 0;
2224
2225                 /*
2226                  * We can only guess - assume it'll run the full timeout
2227                  * if given, otherwise assume it'll run at the specified rate.
2228                  */
2229                 if (td->timeout)
2230                         t_eta = td->timeout + td->start_delay - elapsed;
2231                 if (td->rate) {
2232                         r_eta = (bytes_total / 1024) / td->rate;
2233                         r_eta += td->start_delay - elapsed;
2234                 }
2235
2236                 if (r_eta && t_eta)
2237                         eta_sec = min(r_eta, t_eta);
2238                 else if (r_eta)
2239                         eta_sec = r_eta;
2240                 else if (t_eta)
2241                         eta_sec = t_eta;
2242                 else
2243                         eta_sec = INT_MAX;
2244         } else {
2245                 /*
2246                  * thread is already done
2247                  */
2248                 eta_sec = 0;
2249         }
2250
2251         return eta_sec;
2252 }
2253
2254 static void print_thread_status(void)
2255 {
2256         unsigned long elapsed = time_since_now(&genesis);
2257         int i, nr_running, t_rate, m_rate, *eta_secs, eta_sec;
2258         char eta_str[32];
2259         double perc = 0.0;
2260
2261         eta_secs = malloc(thread_number * sizeof(int));
2262         memset(eta_secs, 0, thread_number * sizeof(int));
2263
2264         nr_running = t_rate = m_rate = 0;
2265         for (i = 0; i < thread_number; i++) {
2266                 struct thread_data *td = &threads[i];
2267
2268                 if (td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING){
2269                         nr_running++;
2270                         t_rate += td->rate;
2271                         m_rate += td->ratemin;
2272                 }
2273
2274                 if (elapsed >= 3)
2275                         eta_secs[i] = thread_eta(td, elapsed);
2276                 else
2277                         eta_secs[i] = INT_MAX;
2278
2279                 check_str_update(td);
2280         }
2281
2282         if (exitall_on_terminate)
2283                 eta_sec = INT_MAX;
2284         else
2285                 eta_sec = 0;
2286
2287         for (i = 0; i < thread_number; i++) {
2288                 if (exitall_on_terminate) {
2289                         if (eta_secs[i] < eta_sec)
2290                                 eta_sec = eta_secs[i];
2291                 } else {
2292                         if (eta_secs[i] > eta_sec)
2293                                 eta_sec = eta_secs[i];
2294                 }
2295         }
2296
2297         if (eta_sec != INT_MAX && elapsed) {
2298                 perc = (double) elapsed / (double) (elapsed + eta_sec);
2299                 eta_to_str(eta_str, eta_sec);
2300         }
2301
2302         printf("Threads now running (%d)", nr_running);
2303         if (m_rate || t_rate)
2304                 printf(", commitrate %d/%dKiB/sec", t_rate, m_rate);
2305         if (eta_sec != INT_MAX) {
2306                 perc *= 100.0;
2307                 printf(": [%s] [%3.2f%% done] [eta %s]", run_str, perc,eta_str);
2308         }
2309         printf("\r");
2310         fflush(stdout);
2311         free(eta_secs);
2312 }
2313
2314 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
2315 {
2316         int i;
2317
2318         /*
2319          * reap exited threads (TD_EXITED -> TD_REAPED)
2320          */
2321         for (i = 0; i < thread_number; i++) {
2322                 struct thread_data *td = &threads[i];
2323
2324                 if (td->runstate != TD_EXITED)
2325                         continue;
2326
2327                 td_set_runstate(td, TD_REAPED);
2328
2329                 if (td->use_thread) {
2330                         long ret;
2331
2332                         if (pthread_join(td->thread, (void *) &ret))
2333                                 perror("thread_join");
2334                 } else
2335                         waitpid(td->pid, NULL, 0);
2336
2337                 (*nr_running)--;
2338                 (*m_rate) -= td->ratemin;
2339                 (*t_rate) -= td->rate;
2340         }
2341 }
2342
2343 static void fio_unpin_memory(void *pinned)
2344 {
2345         if (pinned) {
2346                 if (munlock(pinned, mlock_size) < 0)
2347                         perror("munlock");
2348                 munmap(pinned, mlock_size);
2349         }
2350 }
2351
2352 static void *fio_pin_memory(void)
2353 {
2354         long pagesize, pages;
2355         void *ptr;
2356
2357         if (!mlock_size)
2358                 return NULL;
2359
2360         /*
2361          * Don't allow mlock of more than real_mem-128MB
2362          */
2363         pagesize = sysconf(_SC_PAGESIZE);
2364         pages = sysconf(_SC_PHYS_PAGES);
2365         if (pages != -1 && pagesize != -1) {
2366                 unsigned long long real_mem = pages * pagesize;
2367
2368                 if ((mlock_size + 128 * 1024 * 1024) > real_mem) {
2369                         mlock_size = real_mem - 128 * 1024 * 1024;
2370                         printf("fio: limiting mlocked memory to %lluMiB\n",
2371                                                         mlock_size >> 20);
2372                 }
2373         }
2374
2375         ptr = mmap(NULL, mlock_size, PROT_READ | PROT_WRITE, MAP_PRIVATE | OS_MAP_ANON, 0, 0);
2376         if (!ptr) {
2377                 perror("malloc locked mem");
2378                 return NULL;
2379         }
2380         if (mlock(ptr, mlock_size) < 0) {
2381                 munmap(ptr, mlock_size);
2382                 perror("mlock");
2383                 return NULL;
2384         }
2385
2386         return ptr;
2387 }
2388
2389 static void run_threads(void)
2390 {
2391         struct thread_data *td;
2392         unsigned long spent;
2393         int i, todo, nr_running, m_rate, t_rate, nr_started;
2394         void *mlocked_mem;
2395
2396         mlocked_mem = fio_pin_memory();
2397
2398         printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
2399         fflush(stdout);
2400
2401         signal(SIGINT, sig_handler);
2402         signal(SIGALRM, sig_handler);
2403
2404         todo = thread_number;
2405         nr_running = 0;
2406         nr_started = 0;
2407         m_rate = t_rate = 0;
2408
2409         for (i = 0; i < thread_number; i++) {
2410                 td = &threads[i];
2411
2412                 run_str[td->thread_number - 1] = 'P';
2413
2414                 init_disk_util(td);
2415
2416                 if (!td->create_serialize)
2417                         continue;
2418
2419                 /*
2420                  * do file setup here so it happens sequentially,
2421                  * we don't want X number of threads getting their
2422                  * client data interspersed on disk
2423                  */
2424                 if (setup_file(td)) {
2425                         td_set_runstate(td, TD_REAPED);
2426                         todo--;
2427                 }
2428         }
2429
2430         gettimeofday(&genesis, NULL);
2431
2432         while (todo) {
2433                 struct thread_data *map[MAX_JOBS];
2434                 struct timeval this_start;
2435                 int this_jobs = 0, left;
2436
2437                 /*
2438                  * create threads (TD_NOT_CREATED -> TD_CREATED)
2439                  */
2440                 for (i = 0; i < thread_number; i++) {
2441                         td = &threads[i];
2442
2443                         if (td->runstate != TD_NOT_CREATED)
2444                                 continue;
2445
2446                         /*
2447                          * never got a chance to start, killed by other
2448                          * thread for some reason
2449                          */
2450                         if (td->terminate) {
2451                                 todo--;
2452                                 continue;
2453                         }
2454
2455                         if (td->start_delay) {
2456                                 spent = mtime_since_now(&genesis);
2457
2458                                 if (td->start_delay * 1000 > spent)
2459                                         continue;
2460                         }
2461
2462                         if (td->stonewall && (nr_started || nr_running))
2463                                 break;
2464
2465                         /*
2466                          * Set state to created. Thread will transition
2467                          * to TD_INITIALIZED when it's done setting up.
2468                          */
2469                         td_set_runstate(td, TD_CREATED);
2470                         map[this_jobs++] = td;
2471                         sem_init(&startup_sem, 0, 1);
2472                         nr_started++;
2473
2474                         if (td->use_thread) {
2475                                 if (pthread_create(&td->thread, NULL, thread_main, td)) {
2476                                         perror("thread_create");
2477                                         nr_started--;
2478                                 }
2479                         } else {
2480                                 if (fork())
2481                                         sem_wait(&startup_sem);
2482                                 else {
2483                                         fork_main(shm_id, i);
2484                                         exit(0);
2485                                 }
2486                         }
2487                 }
2488
2489                 /*
2490                  * Wait for the started threads to transition to
2491                  * TD_INITIALIZED.
2492                  */
2493                 printf("fio: Waiting for threads to initialize...\n");
2494                 gettimeofday(&this_start, NULL);
2495                 left = this_jobs;
2496                 while (left) {
2497                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2498                                 break;
2499
2500                         usleep(100000);
2501
2502                         for (i = 0; i < this_jobs; i++) {
2503                                 td = map[i];
2504                                 if (!td)
2505                                         continue;
2506                                 if (td->runstate == TD_INITIALIZED) {
2507                                         map[i] = NULL;
2508                                         left--;
2509                                 } else if (td->runstate >= TD_EXITED) {
2510                                         map[i] = NULL;
2511                                         left--;
2512                                         todo--;
2513                                         nr_running++; /* work-around... */
2514                                 }
2515                         }
2516                 }
2517
2518                 if (left) {
2519                         fprintf(stderr, "fio: %d jobs failed to start\n", left);
2520                         for (i = 0; i < this_jobs; i++) {
2521                                 td = map[i];
2522                                 if (!td)
2523                                         continue;
2524                                 kill(td->pid, SIGTERM);
2525                         }
2526                         break;
2527                 }
2528
2529                 /*
2530                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
2531                  */
2532                 printf("fio: Go for launch\n");
2533                 for (i = 0; i < thread_number; i++) {
2534                         td = &threads[i];
2535
2536                         if (td->runstate != TD_INITIALIZED)
2537                                 continue;
2538
2539                         td_set_runstate(td, TD_RUNNING);
2540                         nr_running++;
2541                         nr_started--;
2542                         m_rate += td->ratemin;
2543                         t_rate += td->rate;
2544                         todo--;
2545                         sem_post(&td->mutex);
2546                 }
2547
2548                 reap_threads(&nr_running, &t_rate, &m_rate);
2549
2550                 if (todo)
2551                         usleep(100000);
2552         }
2553
2554         while (nr_running) {
2555                 reap_threads(&nr_running, &t_rate, &m_rate);
2556                 usleep(10000);
2557         }
2558
2559         update_io_ticks();
2560         fio_unpin_memory(mlocked_mem);
2561 }
2562
2563 static void show_group_stats(struct group_run_stats *rs, int id)
2564 {
2565         printf("\nRun status group %d (all jobs):\n", id);
2566
2567         if (rs->max_run[DDIR_READ])
2568                 printf("   READ: io=%lluMiB, aggrb=%llu, minb=%llu, maxb=%llu, mint=%llumsec, maxt=%llumsec\n", rs->io_mb[0], rs->agg[0], rs->min_bw[0], rs->max_bw[0], rs->min_run[0], rs->max_run[0]);
2569         if (rs->max_run[DDIR_WRITE])
2570                 printf("  WRITE: io=%lluMiB, aggrb=%llu, minb=%llu, maxb=%llu, mint=%llumsec, maxt=%llumsec\n", rs->io_mb[1], rs->agg[1], rs->min_bw[1], rs->max_bw[1], rs->min_run[1], rs->max_run[1]);
2571 }
2572
2573 static void show_disk_util(void)
2574 {
2575         struct disk_util_stat *dus;
2576         struct list_head *entry;
2577         struct disk_util *du;
2578         double util;
2579
2580         printf("\nDisk stats (read/write):\n");
2581
2582         list_for_each(entry, &disk_list) {
2583                 du = list_entry(entry, struct disk_util, list);
2584                 dus = &du->dus;
2585
2586                 util = (double) 100 * du->dus.io_ticks / (double) du->msec;
2587                 if (util > 100.0)
2588                         util = 100.0;
2589
2590                 printf("  %s: ios=%u/%u, merge=%u/%u, ticks=%u/%u, in_queue=%u, util=%3.2f%%\n", du->name, dus->ios[0], dus->ios[1], dus->merges[0], dus->merges[1], dus->ticks[0], dus->ticks[1], dus->time_in_queue, util);
2591         }
2592 }
2593
2594 static void show_run_stats(void)
2595 {
2596         struct group_run_stats *runstats, *rs;
2597         struct thread_data *td;
2598         int i;
2599
2600         runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
2601
2602         for (i = 0; i < groupid + 1; i++) {
2603                 rs = &runstats[i];
2604
2605                 memset(rs, 0, sizeof(*rs));
2606                 rs->min_bw[0] = rs->min_run[0] = ~0UL;
2607                 rs->min_bw[1] = rs->min_run[1] = ~0UL;
2608         }
2609
2610         for (i = 0; i < thread_number; i++) {
2611                 unsigned long long rbw, wbw;
2612
2613                 td = &threads[i];
2614
2615                 if (td->error) {
2616                         printf("Client%d: %s\n", td->thread_number, td->verror);
2617                         continue;
2618                 }
2619
2620                 rs = &runstats[td->groupid];
2621
2622                 if (td->runtime[0] < rs->min_run[0] || !rs->min_run[0])
2623                         rs->min_run[0] = td->runtime[0];
2624                 if (td->runtime[0] > rs->max_run[0])
2625                         rs->max_run[0] = td->runtime[0];
2626                 if (td->runtime[1] < rs->min_run[1] || !rs->min_run[1])
2627                         rs->min_run[1] = td->runtime[1];
2628                 if (td->runtime[1] > rs->max_run[1])
2629                         rs->max_run[1] = td->runtime[1];
2630
2631                 rbw = wbw = 0;
2632                 if (td->runtime[0])
2633                         rbw = td->io_bytes[0] / (unsigned long long) td->runtime[0];
2634                 if (td->runtime[1])
2635                         wbw = td->io_bytes[1] / (unsigned long long) td->runtime[1];
2636
2637                 if (rbw < rs->min_bw[0])
2638                         rs->min_bw[0] = rbw;
2639                 if (wbw < rs->min_bw[1])
2640                         rs->min_bw[1] = wbw;
2641                 if (rbw > rs->max_bw[0])
2642                         rs->max_bw[0] = rbw;
2643                 if (wbw > rs->max_bw[1])
2644                         rs->max_bw[1] = wbw;
2645
2646                 rs->io_mb[0] += td->io_bytes[0] >> 20;
2647                 rs->io_mb[1] += td->io_bytes[1] >> 20;
2648         }
2649
2650         for (i = 0; i < groupid + 1; i++) {
2651                 rs = &runstats[i];
2652
2653                 if (rs->max_run[0])
2654                         rs->agg[0] = (rs->io_mb[0]*1024*1000) / rs->max_run[0];
2655                 if (rs->max_run[1])
2656                         rs->agg[1] = (rs->io_mb[1]*1024*1000) / rs->max_run[1];
2657         }
2658
2659         /*
2660          * don't overwrite last signal output
2661          */
2662         printf("\n");
2663
2664         for (i = 0; i < thread_number; i++) {
2665                 td = &threads[i];
2666                 rs = &runstats[td->groupid];
2667
2668                 show_thread_status(td, rs);
2669         }
2670
2671         for (i = 0; i < groupid + 1; i++)
2672                 show_group_stats(&runstats[i], i);
2673
2674         show_disk_util();
2675 }
2676
2677 int main(int argc, char *argv[])
2678 {
2679         if (parse_options(argc, argv))
2680                 return 1;
2681
2682         if (!thread_number) {
2683                 printf("Nothing to do\n");
2684                 return 1;
2685         }
2686
2687         disk_util_timer_arm();
2688
2689         run_threads();
2690         show_run_stats();
2691
2692         return 0;
2693 }