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 <ctype.h>
  30 #include <sched.h>
  31 #include <libaio.h>
  32 #include <math.h>
  33 #include <limits.h>
  34 #include <sys/time.h>
  35 #include <sys/types.h>
  36 #include <sys/stat.h>
  37 #include <sys/wait.h>
  38 #include <semaphore.h>
  39 #include <sys/ipc.h>
  40 #include <sys/shm.h>
  41 #include <asm/unistd.h>
  42
  43 #include "list.h"
  44
  45 #define MAX_JOBS        (1024)
  46
  47 /*
  48  * assume we don't have _get either, if _set isn't defined
  49  */
  50 #ifndef __NR_ioprio_set
  51 #if defined(__i386__)
  52 #define __NR_ioprio_set         289
  53 #define __NR_ioprio_get         290
  54 #elif defined(__powerpc__) || defined(__powerpc64__)
  55 #define __NR_ioprio_set         273
  56 #define __NR_ioprio_get         274
  57 #elif defined(__x86_64__)
  58 #define __NR_ioprio_set         251
  59 #define __NR_ioprio_get         252
  60 #elif defined(__ia64__)
  61 #define __NR_ioprio_set         1274
  62 #define __NR_ioprio_get         1275
  63 #elif defined(__alpha__)
  64 #define __NR_ioprio_set         442
  65 #define __NR_ioprio_get         443
  66 #elif defined(__s390x__) || defined(__s390__)
  67 #define __NR_ioprio_set         282
  68 #define __NR_ioprio_get         283
  69 #else
  70 #error "Unsupported arch"
  71 #endif
  72 #endif
  73
  74 #ifndef __NR_fadvise64
  75 #if defined(__i386__)
  76 #define __NR_fadvise64          250
  77 #elif defined(__powerpc__) || defined(__powerpc64__)
  78 #define __NR_fadvise64          233
  79 #elif defined(__x86_64__)
  80 #define __NR_fadvise64          221
  81 #elif defined(__ia64__)
  82 #define __NR_fadvise64          1234
  83 #elif defined(__alpha__)
  84 #define __NR_fadvise64          413
  85 #elif defined(__s390x__) || defined(__s390__)
  86 #define __NR_fadvise64          253
  87 #else
  88 #error "Unsupported arch"
  89 #endif
  90 #endif
  91
  92 static int ioprio_set(int which, int who, int ioprio)
  93 {
  94         return syscall(__NR_ioprio_set, which, who, ioprio);
  95 }
  96
  97 /*
  98  * we want fadvise64 really, but it's so tangled... later
  99  */
 100 static int fadvise(int fd, loff_t offset, size_t len, int advice)
 101 {
 102 #if 0
 103         return syscall(__NR_fadvise64, fd, offset, offset >> 32, len, advice);
 104 #else
 105         return posix_fadvise(fd, (off_t) offset, len, advice);
 106 #endif
 107 }
 108
 109 enum {
 110         IOPRIO_WHO_PROCESS = 1,
 111         IOPRIO_WHO_PGRP,
 112         IOPRIO_WHO_USER,
 113 };
 114
 115 #define IOPRIO_CLASS_SHIFT      13
 116
 117 #define MASK    (4095)
 118
 119 #define DEF_BS          (4096)
 120 #define DEF_TIMEOUT     (30)
 121 #define DEF_RATE_CYCLE  (1000)
 122 #define DEF_ODIRECT     (1)
 123 #define DEF_SEQUENTIAL  (1)
 124 #define DEF_RAND_REPEAT (1)
 125 #define DEF_OVERWRITE   (0)
 126 #define DEF_CREATE      (1)
 127 #define DEF_INVALIDATE  (1)
 128
 129 #define ALIGN(buf)      (char *) (((unsigned long) (buf) + MASK) & ~(MASK))
 130
 131 static int repeatable = DEF_RAND_REPEAT;
 132 static int rate_quit = 1;
 133
 134 static int thread_number;
 135 static char *ini_file;
 136
 137 static int max_jobs = MAX_JOBS;
 138
 139 static char run_str[MAX_JOBS + 1];
 140
 141 static int shm_id;
 142
 143 enum {
 144         DDIR_READ = 0,
 145         DDIR_WRITE,
 146 };
 147
 148 /*
 149  * thread life cycle
 150  */
 151 enum {
 152         TD_NOT_CREATED = 0,
 153         TD_CREATED,
 154         TD_STARTED,
 155         TD_EXITED,
 156         TD_REAPED,
 157 };
 158
 159 /*
 160  * The io unit
 161  */
 162 struct io_u {
 163         struct iocb iocb;
 164         struct timeval start_time;
 165         struct timeval issue_time;
 166
 167         void *mem;
 168         char *buf;
 169         unsigned int buflen;
 170         off_t offset;
 171
 172         struct list_head list;
 173 };
 174
 175 struct io_stat {
 176         unsigned long val;
 177         unsigned long val_sq;
 178         unsigned long max_val;
 179         unsigned long min_val;
 180         unsigned long samples;
 181 };
 182
 183 #define td_read(td)             ((td)->ddir == DDIR_READ)
 184 #define should_fsync(td)        (!td_read(td) && !(td)->odirect)
 185
 186 struct thread_data {
 187         char file_name[256];
 188         int thread_number;
 189         int error;
 190         int fd;
 191         pid_t pid;
 192         volatile int terminate;
 193         volatile int runstate;
 194         unsigned int ddir;
 195         unsigned int ioprio;
 196         unsigned int sequential;
 197         unsigned int bs;
 198         unsigned int odirect;
 199         unsigned int delay_sleep;
 200         unsigned int fsync_blocks;
 201         unsigned int start_delay;
 202         unsigned int timeout;
 203         unsigned int use_aio;
 204         unsigned int create_file;
 205         unsigned int overwrite;
 206         unsigned int invalidate_cache;
 207         unsigned long long file_size;
 208         unsigned long long file_offset;
 209         cpu_set_t cpumask;
 210
 211         off_t cur_off;
 212
 213         io_context_t aio_ctx;
 214         unsigned int aio_depth;
 215         struct io_event *aio_events;
 216
 217         unsigned int cur_depth;
 218         struct list_head io_u_freelist;
 219         struct list_head io_u_busylist;
 220
 221         unsigned int rate;
 222         unsigned int ratemin;
 223         unsigned int ratecycle;
 224         unsigned long rate_usec_cycle;
 225         long rate_pending_usleep;
 226         unsigned long rate_blocks;
 227         struct timeval lastrate;
 228
 229         unsigned long runtime;          /* sec */
 230         unsigned long blocks;
 231         unsigned long io_blocks;
 232         unsigned long last_block;
 233         sem_t mutex;
 234         struct drand48_data random_state;
 235
 236         /*
 237          * bandwidth and latency stats
 238          */
 239         struct io_stat clat_stat;               /* completion latency */
 240         struct io_stat slat_stat;               /* submission latency */
 241
 242         struct io_stat bw_stat;                 /* bandwidth stats */
 243         unsigned long stat_io_blocks;
 244         struct timeval stat_sample_time;
 245
 246         struct timeval start;
 247 };
 248
 249 static struct thread_data *threads;
 250 static struct thread_data def_thread;
 251
 252 static sem_t startup_sem;
 253
 254 static void sig_handler(int sig)
 255 {
 256         int i;
 257
 258         for (i = 0; i < thread_number; i++) {
 259                 struct thread_data *td = &threads[i];
 260
 261                 td->terminate = 1;
 262                 td->start_delay = 0;
 263         }
 264 }
 265
 266 static int init_random_state(struct thread_data *td)
 267 {
 268         unsigned long seed = 123;
 269
 270         if (td->sequential)
 271                 return 0;
 272
 273         if (!repeatable) {
 274                 int fd = open("/dev/random", O_RDONLY);
 275
 276                 if (fd == -1) {
 277                         td->error = errno;
 278                         return 1;
 279                 }
 280
 281                 if (read(fd, &seed, sizeof(seed)) < (int) sizeof(seed)) {
 282                         td->error = EIO;
 283                         close(fd);
 284                         return 1;
 285                 }
 286
 287                 close(fd);
 288         }
 289
 290         srand48_r(seed, &td->random_state);
 291         return 0;
 292 }
 293
 294 static unsigned long utime_since(struct timeval *s, struct timeval *e)
 295 {
 296         double sec, usec;
 297
 298         sec = e->tv_sec - s->tv_sec;
 299         usec = e->tv_usec - s->tv_usec;
 300         if (sec > 0 && usec < 0) {
 301                 sec--;
 302                 usec += 1000000;
 303         }
 304
 305         sec *= (double) 1000000;
 306
 307         return sec + usec;
 308 }
 309
 310 static unsigned long mtime_since(struct timeval *s, struct timeval *e)
 311 {
 312         double sec, usec;
 313
 314         sec = e->tv_sec - s->tv_sec;
 315         usec = e->tv_usec - s->tv_usec;
 316         if (sec > 0 && usec < 0) {
 317                 sec--;
 318                 usec += 1000000;
 319         }
 320
 321         sec *= (double) 1000;
 322         usec /= (double) 1000;
 323
 324         return sec + usec;
 325 }
 326
 327 static unsigned long mtime_since_now(struct timeval *s)
 328 {
 329         struct timeval t;
 330
 331         gettimeofday(&t, NULL);
 332         return mtime_since(s, &t);
 333 }
 334
 335 static inline unsigned long msec_now(struct timeval *s)
 336 {
 337         return s->tv_sec * 1000 + s->tv_usec / 1000;
 338 }
 339
 340 static unsigned long get_next_offset(struct thread_data *td)
 341 {
 342         unsigned long b;
 343         long r;
 344
 345         if (!td->sequential) {
 346                 lrand48_r(&td->random_state, &r);
 347                 b = (1+(double) (td->blocks-1) * r / (RAND_MAX+1.0));
 348         } else {
 349                 b = td->last_block;
 350                 td->last_block++;
 351         }
 352
 353         return b * td->bs + td->file_offset;
 354 }
 355
 356 static inline void add_stat_sample(struct thread_data *td, struct io_stat *is,
 357                                    unsigned long val)
 358 {
 359         if (val > is->max_val)
 360                 is->max_val = val;
 361         if (val < is->min_val)
 362                 is->min_val = val;
 363
 364         is->val += val;
 365         is->val_sq += val * val;
 366         is->samples++;
 367 }
 368
 369 static void add_clat_sample(struct thread_data *td, unsigned long msec)
 370 {
 371         add_stat_sample(td, &td->clat_stat, msec);
 372 }
 373
 374 static void add_slat_sample(struct thread_data *td, unsigned long msec)
 375 {
 376         add_stat_sample(td, &td->slat_stat, msec);
 377 }
 378
 379 static void add_bw_sample(struct thread_data *td, unsigned long msec)
 380 {
 381         unsigned long spent = mtime_since_now(&td->stat_sample_time);
 382         unsigned long rate;
 383
 384         if (spent < 500)
 385                 return;
 386
 387         rate = ((td->io_blocks - td->stat_io_blocks) * td->bs) / spent;
 388         add_stat_sample(td, &td->bw_stat, rate);
 389
 390         gettimeofday(&td->stat_sample_time, NULL);
 391         td->stat_io_blocks = td->io_blocks;
 392 }
 393
 394 static void usec_sleep(int usec)
 395 {
 396         struct timespec req = { .tv_sec = 0, .tv_nsec = usec * 1000 };
 397         struct timespec rem;
 398
 399         do {
 400                 rem.tv_sec = rem.tv_nsec = 0;
 401                 nanosleep(&req, &rem);
 402                 if (!rem.tv_nsec)
 403                         break;
 404
 405                 req.tv_nsec = rem.tv_nsec;
 406         } while (1);
 407 }
 408
 409 static void rate_throttle(struct thread_data *td, unsigned long time_spent)
 410 {
 411         if (!td->rate)
 412                 return;
 413
 414         if (time_spent < td->rate_usec_cycle) {
 415                 unsigned long s = td->rate_usec_cycle - time_spent;
 416
 417                 td->rate_pending_usleep += s;
 418                 if (td->rate_pending_usleep >= 100000) {
 419                         usec_sleep(td->rate_pending_usleep);
 420                         td->rate_pending_usleep = 0;
 421                 }
 422         } else {
 423                 long overtime = time_spent - td->rate_usec_cycle;
 424
 425                 td->rate_pending_usleep -= overtime;
 426         }
 427 }
 428
 429 static int check_min_rate(struct thread_data *td, struct timeval *now)
 430 {
 431         unsigned long spent;
 432         unsigned long rate;
 433
 434         /*
 435          * allow a 2 second settle period in the beginning
 436          */
 437         if (mtime_since(&td->start, now) < 2000)
 438                 return 0;
 439
 440         /*
 441          * if rate blocks is set, sample is running
 442          */
 443         if (td->rate_blocks) {
 444                 spent = mtime_since(&td->lastrate, now);
 445                 if (spent < td->ratecycle)
 446                         return 0;
 447
 448                 rate = ((td->io_blocks - td->rate_blocks) * td->bs) / spent;
 449                 if (rate < td->ratemin) {
 450                         printf("Client%d: min rate %d not met, got %ldKiB/sec\n", td->thread_number, td->ratemin, rate);
 451                         if (rate_quit)
 452                                 sig_handler(0);
 453                         return 1;
 454                 }
 455         }
 456
 457         td->rate_blocks = td->io_blocks;
 458         memcpy(&td->lastrate, now, sizeof(*now));
 459         return 0;
 460 }
 461
 462 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 463 {
 464         if (mtime_since(&td->start, t) >= td->timeout * 1000)
 465                 return 1;
 466
 467         return 0;
 468 }
 469
 470 static void put_io_u(struct thread_data *td, struct io_u *io_u)
 471 {
 472         list_del(&io_u->list);
 473         list_add(&io_u->list, &td->io_u_freelist);
 474         td->cur_depth--;
 475 }
 476
 477 static struct io_u *get_io_u(struct thread_data *td)
 478 {
 479         struct io_u *io_u;
 480
 481         if (list_empty(&td->io_u_freelist))
 482                 return NULL;
 483
 484         io_u = list_entry(td->io_u_freelist.next, struct io_u, list);
 485         list_del(&io_u->list);
 486         list_add(&io_u->list, &td->io_u_busylist);
 487
 488         io_u->offset = get_next_offset(td);
 489
 490         if (td->use_aio) {
 491                 if (td_read(td))
 492                         io_prep_pread(&io_u->iocb, td->fd, io_u->buf, io_u->buflen, io_u->offset);
 493                 else
 494                         io_prep_pwrite(&io_u->iocb, td->fd, io_u->buf, io_u->buflen, io_u->offset);
 495         }
 496
 497         gettimeofday(&io_u->start_time, NULL);
 498         td->cur_depth++;
 499         return io_u;
 500 }
 501
 502 static void do_sync_io(struct thread_data *td)
 503 {
 504         unsigned long blocks, msec, usec;
 505         struct timeval e;
 506
 507         td->cur_off = 0;
 508
 509         for (blocks = 0; blocks < td->blocks; blocks++) {
 510                 struct io_u *io_u;
 511                 int ret;
 512
 513                 if (td->terminate)
 514                         break;
 515
 516                 io_u = get_io_u(td);
 517
 518                 if (td->cur_off != io_u->offset) {
 519                         if (lseek(td->fd, io_u->offset, SEEK_SET) == -1) {
 520                                 td->error = errno;
 521                                 break;
 522                         }
 523                 }
 524
 525                 if (td->delay_sleep)
 526                         usec_sleep(td->delay_sleep);
 527
 528                 if (td_read(td))
 529                         ret = read(td->fd, io_u->buf, io_u->buflen);
 530                 else
 531                         ret = write(td->fd, io_u->buf, io_u->buflen);
 532
 533                 if (ret < (int) io_u->buflen) {
 534                         if (ret == -1)
 535                                 td->error = errno;
 536                         break;
 537                 }
 538
 539                 td->io_blocks++;
 540                 td->cur_off = io_u->offset + io_u->buflen;
 541
 542                 if (should_fsync(td) && td->fsync_blocks &&
 543                     (td->io_blocks % td->fsync_blocks) == 0)
 544                         fsync(td->fd);
 545
 546                 gettimeofday(&e, NULL);
 547
 548                 usec = utime_since(&io_u->start_time, &e);
 549
 550                 rate_throttle(td, usec);
 551
 552                 if (check_min_rate(td, &e)) {
 553                         td->error = ENODATA;
 554                         break;
 555                 }
 556
 557                 msec = usec / 1000;
 558                 add_clat_sample(td, msec);
 559                 add_bw_sample(td, msec);
 560
 561                 if (runtime_exceeded(td, &e))
 562                         break;
 563
 564                 put_io_u(td, io_u);
 565         }
 566
 567         if (should_fsync(td))
 568                 fsync(td->fd);
 569 }
 570
 571 static int io_u_queue(struct thread_data *td, struct io_u *io_u)
 572 {
 573         struct iocb *iocb = &io_u->iocb;
 574         int ret;
 575
 576         do {
 577                 ret = io_submit(td->aio_ctx, 1, &iocb);
 578                 if (ret == 1)
 579                         return 0;
 580                 else if (ret == EAGAIN)
 581                         usleep(100);
 582                 else if (ret == EINTR)
 583                         continue;
 584                 else
 585                         break;
 586         } while (1);
 587
 588         return ret;
 589 }
 590
 591 #define iocb_time(iocb) ((unsigned long) (iocb)->data)
 592 #define ev_to_iou(ev)   (struct io_u *) ((unsigned long) (ev)->obj)
 593
 594 static void ios_completed(struct thread_data *td, int nr)
 595 {
 596         unsigned long msec;
 597         struct io_u *io_u;
 598         struct timeval e;
 599         int i;
 600
 601         gettimeofday(&e, NULL);
 602
 603         for (i = 0; i < nr; i++) {
 604                 td->io_blocks++;
 605
 606                 io_u = ev_to_iou(td->aio_events + i);
 607
 608                 msec = mtime_since(&io_u->issue_time, &e);
 609
 610                 add_clat_sample(td, msec);
 611                 add_bw_sample(td, msec);
 612
 613                 put_io_u(td, io_u);
 614         }
 615 }
 616
 617 static void cleanup_pending_aio(struct thread_data *td)
 618 {
 619         struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
 620         struct list_head *entry, *n;
 621         struct io_u *io_u;
 622         int r;
 623
 624         /*
 625          * get immediately available events, if any
 626          */
 627         r = io_getevents(td->aio_ctx, 0, td->cur_depth, td->aio_events, &ts);
 628         if (r > 0)
 629                 ios_completed(td, r);
 630
 631         /*
 632          * now cancel remaining active events
 633          */
 634         list_for_each_safe(entry, n, &td->io_u_busylist) {
 635                 io_u = list_entry(entry, struct io_u, list);
 636
 637                 r = io_cancel(td->aio_ctx, &io_u->iocb, td->aio_events);
 638                 if (!r)
 639                         put_io_u(td, io_u);
 640         }
 641
 642         if (td->cur_depth) {
 643                 r = io_getevents(td->aio_ctx, td->cur_depth, td->cur_depth, td->aio_events, NULL);
 644                 if (r > 0)
 645                         ios_completed(td, r);
 646         }
 647 }
 648
 649 static void do_async_io(struct thread_data *td)
 650 {
 651         struct timeval s, e;
 652         unsigned long blocks, usec;
 653
 654         for (blocks = 0; blocks < td->blocks; blocks++) {
 655                 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
 656                 struct timespec *timeout;
 657                 int ret, min_evts = 0;
 658                 struct io_u *io_u;
 659
 660                 if (td->terminate)
 661                         break;
 662
 663                 if (td->delay_sleep)
 664                         usec_sleep(td->delay_sleep);
 665
 666                 io_u = get_io_u(td);
 667
 668                 memcpy(&s, &io_u->start_time, sizeof(s));
 669
 670                 ret = io_u_queue(td, io_u);
 671                 if (ret) {
 672                         put_io_u(td, io_u);
 673                         td->error = ret;
 674                         break;
 675                 }
 676
 677                 gettimeofday(&io_u->issue_time, NULL);
 678                 add_slat_sample(td, mtime_since(&io_u->start_time, &io_u->issue_time));
 679
 680                 if (td->cur_depth < td->aio_depth) {
 681                         timeout = &ts;
 682                         min_evts = 0;
 683                 } else {
 684                         timeout = NULL;
 685                         min_evts = 1;
 686                 }
 687
 688                 ret = io_getevents(td->aio_ctx, min_evts, td->cur_depth, td->aio_events, timeout);
 689                 if (ret < 0) {
 690                         td->error = errno;
 691                         break;
 692                 } else if (!ret)
 693                         continue;
 694
 695                 ios_completed(td, ret);
 696
 697                 /*
 698                  * the rate is batched for now, it should work for batches
 699                  * of completions except the very first one which may look
 700                  * a little bursty
 701                  */
 702                 gettimeofday(&e, NULL);
 703                 usec = utime_since(&s, &e);
 704
 705                 rate_throttle(td, usec);
 706
 707                 if (check_min_rate(td, &e)) {
 708                         td->error = ENODATA;
 709                         break;
 710                 }
 711
 712                 if (runtime_exceeded(td, &e))
 713                         break;
 714         }
 715
 716         if (td->cur_depth)
 717                 cleanup_pending_aio(td);
 718 }
 719
 720 static void cleanup_aio(struct thread_data *td)
 721 {
 722         io_destroy(td->aio_ctx);
 723
 724         if (td->aio_events)
 725                 free(td->aio_events);
 726 }
 727
 728 static int init_aio(struct thread_data *td)
 729 {
 730         if (io_queue_init(td->aio_depth, &td->aio_ctx)) {
 731                 td->error = errno;
 732                 return 1;
 733         }
 734
 735         td->aio_events = malloc(td->aio_depth * sizeof(struct io_event));
 736         return 0;
 737 }
 738
 739 static void cleanup_io_u(struct thread_data *td)
 740 {
 741         struct list_head *entry, *n;
 742         struct io_u *io_u;
 743
 744         list_for_each_safe(entry, n, &td->io_u_freelist) {
 745                 io_u = list_entry(entry, struct io_u, list);
 746
 747                 list_del(&io_u->list);
 748                 free(io_u->mem);
 749                 free(io_u);
 750         }
 751 }
 752
 753 static void init_io_u(struct thread_data *td)
 754 {
 755         struct io_u *io_u;
 756         int i, max_units;
 757
 758         if (!td->use_aio)
 759                 max_units = 1;
 760         else
 761                 max_units = td->aio_depth;
 762
 763         INIT_LIST_HEAD(&td->io_u_freelist);
 764         INIT_LIST_HEAD(&td->io_u_busylist);
 765
 766         for (i = 0; i < max_units; i++) {
 767                 io_u = malloc(sizeof(*io_u));
 768                 memset(io_u, 0, sizeof(*io_u));
 769                 INIT_LIST_HEAD(&io_u->list);
 770
 771                 io_u->mem = malloc(td->bs + MASK);
 772                 io_u->buf = ALIGN(io_u->mem);
 773                 io_u->buflen = td->bs;
 774
 775                 list_add(&io_u->list, &td->io_u_freelist);
 776         }
 777 }
 778
 779 static int create_file(struct thread_data *td)
 780 {
 781         unsigned int i;
 782         char *b;
 783
 784         /*
 785          * unless specifically asked for overwrite, let normal io extend it
 786          */
 787         if (!td_read(td) && !td->overwrite)
 788                 return 0;
 789
 790         if (!td->file_size) {
 791                 fprintf(stderr, "Need size for create\n");
 792                 td->error = EINVAL;
 793                 return 1;
 794         }
 795
 796         td->fd = open(td->file_name, O_WRONLY | O_CREAT | O_TRUNC, 0644);
 797         if (td->fd < 0) {
 798                 td->error = errno;
 799                 return 1;
 800         }
 801
 802         td->blocks = td->file_size / td->bs;
 803         b = malloc(td->bs);
 804         memset(b, 0, td->bs);
 805
 806         for (i = 0; i < td->blocks; i++) {
 807                 int r = write(td->fd, b, td->bs);
 808
 809                 if (r == td->bs)
 810                         continue;
 811                 else {
 812                         if (r < 0)
 813                                 td->error = errno;
 814                         else
 815                                 td->error = EIO;
 816
 817                         break;
 818                 }
 819         }
 820
 821         fsync(td->fd);
 822         close(td->fd);
 823         td->fd = -1;
 824         free(b);
 825         return 0;
 826 }
 827
 828 static int file_exists(struct thread_data *td)
 829 {
 830         struct stat st;
 831
 832         if (stat(td->file_name, &st) != -1)
 833                 return 1;
 834
 835         return errno != ENOENT;
 836 }
 837
 838 static int setup_file(struct thread_data *td)
 839 {
 840         struct stat st;
 841         int flags = 0;
 842
 843         if (!file_exists(td)) {
 844                 if (!td->create_file) {
 845                         td->error = ENOENT;
 846                         return 1;
 847                 }
 848                 if (create_file(td))
 849                         return 1;
 850         }
 851
 852         if (td->odirect)
 853                 flags |= O_DIRECT;
 854
 855         if (td_read(td))
 856                 td->fd = open(td->file_name, flags | O_RDONLY);
 857         else {
 858                 if (!td->overwrite)
 859                         flags |= O_TRUNC;
 860
 861                 td->fd = open(td->file_name, flags | O_WRONLY | O_CREAT, 0600);
 862         }
 863
 864         if (td->fd == -1) {
 865                 td->error = errno;
 866                 return 1;
 867         }
 868
 869         if (td_read(td)) {
 870                 if (fstat(td->fd, &st) == -1) {
 871                         td->error = errno;
 872                         return 1;
 873                 }
 874
 875                 if (td->file_size > st.st_size)
 876                         st.st_size = td->file_size;
 877         } else {
 878                 if (!td->file_size)
 879                         td->file_size = 1024 * 1024 * 1024;
 880
 881                 st.st_size = td->file_size;
 882         }
 883
 884         td->blocks = (st.st_size - td->file_offset) / td->bs;
 885         if (!td->blocks) {
 886                 fprintf(stderr, "Client%d: no io blocks\n", td->thread_number);
 887                 td->error = EINVAL;
 888                 return 1;
 889         }
 890
 891         if (td->invalidate_cache) {
 892                 if (fadvise(td->fd, 0, st.st_size, POSIX_FADV_DONTNEED) < 0) {
 893                         td->error = errno;
 894                         return 1;
 895                 }
 896         }
 897
 898         return 0;
 899 }
 900
 901 static void *thread_main(int shm_id, int offset, char *argv[])
 902 {
 903         struct thread_data *td;
 904         int ret = 1;
 905         void *data;
 906
 907         setsid();
 908
 909         data = shmat(shm_id, NULL, 0);
 910         td = data + offset * sizeof(struct thread_data);
 911         td->pid = getpid();
 912
 913         init_io_u(td);
 914
 915         if (sched_setaffinity(td->pid, sizeof(td->cpumask), &td->cpumask) == -1) {
 916                 td->error = errno;
 917                 goto err;
 918         }
 919
 920         sprintf(argv[0], "fio%d", offset);
 921
 922         if (td->use_aio && init_aio(td))
 923                 goto err;
 924
 925         if (init_random_state(td))
 926                 goto err;
 927
 928         if (td->ioprio) {
 929                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
 930                         td->error = errno;
 931                         goto err;
 932                 }
 933         }
 934
 935         if (setup_file(td))
 936                 goto err;
 937
 938         sem_post(&startup_sem);
 939         sem_wait(&td->mutex);
 940
 941         gettimeofday(&td->start, NULL);
 942
 943         if (td->ratemin)
 944                 memcpy(&td->lastrate, &td->start, sizeof(td->start));
 945
 946         memcpy(&td->stat_sample_time, &td->start, sizeof(td->start));
 947
 948         if (!td->use_aio)
 949                 do_sync_io(td);
 950         else
 951                 do_async_io(td);
 952
 953         td->runtime = mtime_since_now(&td->start);
 954         ret = 0;
 955 err:
 956         if (td->use_aio)
 957                 cleanup_aio(td);
 958         if (td->fd != -1) {
 959                 close(td->fd);
 960                 td->fd = -1;
 961         }
 962         cleanup_io_u(td);
 963         if (ret) {
 964                 sem_post(&startup_sem);
 965                 sem_wait(&td->mutex);
 966         }
 967         td->runstate = TD_EXITED;
 968         shmdt(data);
 969         return NULL;
 970 }
 971
 972 static void free_shm(void)
 973 {
 974         shmdt(threads);
 975 }
 976
 977 static int calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max,
 978                     double *mean, double *dev)
 979 {
 980         double n;
 981
 982         if (is->samples == 0)
 983                 return 0;
 984
 985         *min = is->min_val;
 986         *max = is->max_val;
 987
 988         n = (double) is->samples;
 989         *mean = (double) is->val / n;
 990         *dev = sqrt(((double) is->val_sq - (*mean * *mean) / n) / (n - 1));
 991         return 1;
 992 }
 993
 994 static void show_thread_status(struct thread_data *td)
 995 {
 996         int prio, prio_class;
 997         unsigned long min, max, bw = 0;
 998         double mean, dev;
 999
1000         if (!td->io_blocks && !td->error)
1001                 return;
1002
1003         if (td->runtime)
1004                 bw = (td->io_blocks * td->bs) / td->runtime;
1005
1006         prio = td->ioprio & 0xff;
1007         prio_class = td->ioprio >> IOPRIO_CLASS_SHIFT;
1008
1009         printf("Client%d: err=%2d, io=%6luMiB, bw=%6luKiB/s\n", td->thread_number, td->error, td->io_blocks * td->bs >> 20, bw);
1010
1011         if (calc_lat(&td->slat_stat, &min, &max, &mean, &dev))
1012                 printf("  slat (msec): min=%5lu, max=%5lu, avg=%5.02f, dev=%5.02f\n", min, max, mean, dev);
1013         if (calc_lat(&td->clat_stat, &min, &max, &mean, &dev))
1014                 printf("  clat (msec): min=%5lu, max=%5lu, avg=%5.02f, dev=%5.02f\n", min, max, mean, dev);
1015         if (calc_lat(&td->bw_stat, &min, &max, &mean, &dev))
1016                 printf("  bw (KiB/s) : min=%5lu, max=%5lu, avg=%5.02f, dev=%5.02f\n", min, max, mean, dev);
1017 }
1018
1019 static int setup_rate(struct thread_data *td)
1020 {
1021         int nr_reads_per_sec;
1022
1023         if (!td->rate)
1024                 return 0;
1025
1026         if (td->rate < td->ratemin) {
1027                 fprintf(stderr, "min rate larger than nominal rate\n");
1028                 return -1;
1029         }
1030
1031         nr_reads_per_sec = td->rate * 1024 / td->bs;
1032         td->rate_usec_cycle = 1000000 / nr_reads_per_sec;
1033         td->rate_pending_usleep = 0;
1034         return 0;
1035 }
1036
1037 static struct thread_data *get_new_job(int global)
1038 {
1039         struct thread_data *td;
1040
1041         if (global)
1042                 return &def_thread;
1043         if (thread_number >= max_jobs)
1044                 return NULL;
1045
1046         td = &threads[thread_number++];
1047         memset(td, 0, sizeof(*td));
1048
1049         td->fd = -1;
1050         td->thread_number = thread_number;
1051
1052         td->ddir = def_thread.ddir;
1053         td->ioprio = def_thread.ioprio;
1054         td->sequential = def_thread.sequential;
1055         td->bs = def_thread.bs;
1056         td->odirect = def_thread.odirect;
1057         td->delay_sleep = def_thread.delay_sleep;
1058         td->fsync_blocks = def_thread.fsync_blocks;
1059         td->start_delay = def_thread.start_delay;
1060         td->timeout = def_thread.timeout;
1061         td->use_aio = def_thread.use_aio;
1062         td->create_file = def_thread.create_file;
1063         td->overwrite = def_thread.overwrite;
1064         td->invalidate_cache = def_thread.invalidate_cache;
1065         td->file_size = def_thread.file_size;
1066         td->file_size = def_thread.file_offset;
1067         td->rate = def_thread.rate;
1068         td->ratemin = def_thread.ratemin;
1069         td->ratecycle = def_thread.ratecycle;
1070         td->aio_depth = def_thread.aio_depth;
1071         memcpy(&td->cpumask, &def_thread.cpumask, sizeof(td->cpumask));
1072
1073         return td;
1074 }
1075
1076 static void put_job(struct thread_data *td)
1077 {
1078         memset(&threads[td->thread_number - 1], 0, sizeof(*td));
1079         thread_number--;
1080 }
1081
1082 static int add_job(struct thread_data *td, const char *filename, int prioclass,
1083                    int prio)
1084 {
1085         if (td == &def_thread)
1086                 return 0;
1087
1088         strcpy(td->file_name, filename);
1089         sem_init(&td->mutex, 1, 0);
1090         td->ioprio = (prioclass << IOPRIO_CLASS_SHIFT) | prio;
1091
1092         td->clat_stat.min_val = ULONG_MAX;
1093         td->slat_stat.min_val = ULONG_MAX;
1094         td->bw_stat.min_val = ULONG_MAX;
1095
1096         run_str[td->thread_number - 1] = 'P';
1097
1098         if (td->use_aio) {
1099                 if (!td->aio_depth)
1100                         td->aio_depth = 1;
1101
1102                 /*
1103                  * no buffered aio support for now
1104                  */
1105                 if (!td->odirect)
1106                         td->odirect = 1;
1107         }
1108
1109         if (setup_rate(td))
1110                 return -1;
1111
1112         printf("Client%d: file=%s, rw=%d, prio=%d/%d, seq=%d, odir=%d, bs=%d, rate=%d, aio=%d, aio_depth=%d\n", td->thread_number, filename, td->ddir, prioclass, prio, td->sequential, td->odirect, td->bs, td->rate, td->use_aio, td->aio_depth);
1113         return 0;
1114 }
1115
1116 static void fill_cpu_mask(cpu_set_t cpumask, int cpu)
1117 {
1118         unsigned int i;
1119
1120         CPU_ZERO(&cpumask);
1121
1122         for (i = 0; i < sizeof(int) * 8; i++) {
1123                 if ((1 << i) & cpu)
1124                         CPU_SET(i, &cpumask);
1125         }
1126 }
1127
1128 static void fill_option(const char *input, char *output)
1129 {
1130         int i;
1131
1132         i = 0;
1133         while (input[i] != ',' && input[i] != '}' && input[i] != '\0') {
1134                 output[i] = input[i];
1135                 i++;
1136         }
1137
1138         output[i] = '\0';
1139 }
1140
1141 /*
1142  * convert string after '=' into decimal value, noting any size suffix
1143  */
1144 static int str_cnv(char *p, unsigned long long *val)
1145 {
1146         unsigned long mult;
1147         char *str;
1148         int len;
1149
1150         str = strstr(p, "=");
1151         if (!str)
1152                 return 1;
1153
1154         str++;
1155         len = strlen(str);
1156         mult = 1;
1157
1158         switch (str[len - 2]) {
1159                 case 'k':
1160                 case 'K':
1161                         mult = 1024;
1162                         break;
1163                 case 'm':
1164                 case 'M':
1165                         mult = 1024 * 1024;
1166                         break;
1167                 case 'g':
1168                 case 'G':
1169                         mult = 1024 * 1024 * 1024;
1170                         break;
1171         }
1172
1173         *val = strtoul(str, NULL, 10);
1174         if (*val == ULONG_MAX && errno == ERANGE)
1175                 return 1;
1176
1177         *val *= mult;
1178         return 0;
1179
1180 }
1181
1182 /*
1183  * job key words:
1184  *
1185  * file=
1186  * bs=
1187  * rw=
1188  * direct=
1189  */
1190 static void parse_jobs_cmd(int argc, char *argv[], int index)
1191 {
1192         struct thread_data *td;
1193         unsigned int prio, prioclass, cpu;
1194         char *string, *filename, *p, *c;
1195         int i;
1196
1197         string = malloc(256);
1198         filename = malloc(256);
1199
1200         for (i = index; i < argc; i++) {
1201                 p = argv[i];
1202
1203                 c = strpbrk(p, "{");
1204                 if (!c)
1205                         break;
1206
1207                 filename[0] = 0;
1208
1209                 td = get_new_job(0);
1210                 if (!td)
1211                         break;
1212
1213                 prioclass = 2;
1214                 prio = 4;
1215
1216                 c = strstr(p, "rw=");
1217                 if (c) {
1218                         c += 3;
1219                         if (*c == '0')
1220                                 td->ddir = DDIR_READ;
1221                         else
1222                                 td->ddir = DDIR_WRITE;
1223                 }
1224
1225                 c = strstr(p, "prio=");
1226                 if (c) {
1227                         c += 5;
1228                         prio = *c - '0';
1229                 }
1230
1231                 c = strstr(p, "prioclass=");
1232                 if (c) {
1233                         c += 10;
1234                         prioclass = *c - '0';
1235                 }
1236
1237                 c = strstr(p, "file=");
1238                 if (c) {
1239                         c += 5;
1240                         fill_option(c, filename);
1241                 }
1242
1243                 c = strstr(p, "bs=");
1244                 if (c) {
1245                         c += 3;
1246                         fill_option(c, string);
1247                         td->bs = strtoul(string, NULL, 10);
1248                         td->bs <<= 10;
1249                 }
1250
1251                 c = strstr(p, "direct=");
1252                 if (c) {
1253                         c += 7;
1254                         if (*c != '0')
1255                                 td->odirect = 1;
1256                         else
1257                                 td->odirect = 0;
1258                 }
1259
1260                 c = strstr(p, "delay=");
1261                 if (c) {
1262                         c += 6;
1263                         fill_option(c, string);
1264                         td->delay_sleep = strtoul(string, NULL, 10);
1265                 }
1266
1267                 c = strstr(p, "rate=");
1268                 if (c) {
1269                         c += 5;
1270                         fill_option(c, string);
1271                         td->rate = strtoul(string, NULL, 10);
1272                 }
1273
1274                 c = strstr(p, "ratemin=");
1275                 if (c) {
1276                         c += 8;
1277                         fill_option(c, string);
1278                         td->ratemin = strtoul(string, NULL, 10);
1279                 }
1280
1281                 c = strstr(p, "ratecycle=");
1282                 if (c) {
1283                         c += 10;
1284                         fill_option(c, string);
1285                         td->ratecycle = strtoul(string, NULL, 10);
1286                 }
1287
1288                 c = strstr(p, "cpumask=");
1289                 if (c) {
1290                         c += 8;
1291                         fill_option(c, string);
1292                         cpu = strtoul(string, NULL, 10);
1293                         fill_cpu_mask(td->cpumask, cpu);
1294                 }
1295
1296                 c = strstr(p, "fsync=");
1297                 if (c) {
1298                         c += 6;
1299                         fill_option(c, string);
1300                         td->fsync_blocks = strtoul(string, NULL, 10);
1301                 }
1302
1303                 c = strstr(p, "startdelay=");
1304                 if (c) {
1305                         c += 11;
1306                         fill_option(c, string);
1307                         td->start_delay = strtoul(string, NULL, 10);
1308                 }
1309
1310                 c = strstr(p, "timeout=");
1311                 if (c) {
1312                         c += 8;
1313                         fill_option(c, string);
1314                         td->timeout = strtoul(string, NULL, 10);
1315                 }
1316
1317                 c = strstr(p, "invalidate=");
1318                 if (c) {
1319                         c += 11;
1320                         if (*c != '0')
1321                                 td->invalidate_cache = 1;
1322                         else
1323                                 td->invalidate_cache = 0;
1324                 }
1325
1326                 c = strstr(p, "size=");
1327                 if (c) {
1328                         c += 5;
1329                         str_cnv(c, &td->file_size);
1330                 }
1331
1332                 c = strstr(p, "offset=");
1333                 if (c) {
1334                         c += 7;
1335                         str_cnv(c, &td->file_offset);
1336                 }
1337
1338                 c = strstr(p, "aio_depth=");
1339                 if (c) {
1340                         c += 10;
1341                         fill_option(c, string);
1342                         td->aio_depth = strtoul(string, NULL, 10);
1343                 }
1344
1345                 c = strstr(p, "aio");
1346                 if (c)
1347                         td->use_aio = 1;
1348
1349                 c = strstr(p, "create");
1350                 if (c)
1351                         td->create_file = 1;
1352
1353                 c = strstr(p, "overwrite");
1354                 if (c)
1355                         td->overwrite = 1;
1356
1357                 c = strstr(p, "random");
1358                 if (c)
1359                         td->sequential = 0;
1360                 c = strstr(p, "sequential");
1361                 if (c)
1362                         td->sequential = 1;
1363
1364                 if (add_job(td, filename, prioclass, prio))
1365                         put_job(td);
1366         }
1367
1368         free(string);
1369         free(filename);
1370 }
1371
1372 static int check_strcnv(char *p, char *name, unsigned long long *val)
1373 {
1374         if (!strstr(p, name))
1375                 return 1;
1376
1377         return str_cnv(p, val);
1378 }
1379
1380 static int check_int(char *p, char *name, unsigned int *val)
1381 {
1382         char str[128];
1383
1384         sprintf(str, "%s=%%d", name);
1385         if (sscanf(p, str, val) == 1)
1386                 return 0;
1387
1388         sprintf(str, "%s = %%d", name);
1389         if (sscanf(p, str, val) == 1)
1390                 return 0;
1391
1392         return 1;
1393 }
1394
1395 static int is_empty_or_comment(char *line)
1396 {
1397         unsigned int i;
1398
1399         for (i = 0; i < strlen(line); i++) {
1400                 if (line[i] == ';')
1401                         return 1;
1402                 if (!isspace(line[i]) && !iscntrl(line[i]))
1403                         return 0;
1404         }
1405
1406         return 1;
1407 }
1408
1409 static int parse_jobs_ini(char *file)
1410 {
1411         unsigned int prioclass, prio, cpu, global;
1412         struct thread_data *td;
1413         char *string, *name;
1414         fpos_t off;
1415         FILE *f;
1416         char *p;
1417
1418         f = fopen(file, "r");
1419         if (!f) {
1420                 perror("fopen");
1421                 return 1;
1422         }
1423
1424         string = malloc(4096);
1425         name = malloc(256);
1426
1427         while ((p = fgets(string, 4096, f)) != NULL) {
1428                 if (is_empty_or_comment(p))
1429                         continue;
1430                 if (sscanf(p, "[%s]", name) != 1)
1431                         continue;
1432
1433                 global = !strncmp(name, "global", 6);
1434
1435                 name[strlen(name) - 1] = '\0';
1436
1437                 td = get_new_job(global);
1438                 if (!td)
1439                         break;
1440
1441                 prioclass = 2;
1442                 prio = 4;
1443
1444                 fgetpos(f, &off);
1445                 while ((p = fgets(string, 4096, f)) != NULL) {
1446                         if (is_empty_or_comment(p))
1447                                 continue;
1448                         if (strstr(p, "["))
1449                                 break;
1450                         if (!check_int(p, "bs", &td->bs)) {
1451                                 td->bs <<= 10;
1452                                 fgetpos(f, &off);
1453                                 continue;
1454                         }
1455                         if (!check_int(p, "rw", &td->ddir)) {
1456                                 fgetpos(f, &off);
1457                                 continue;
1458                         }
1459                         if (!check_int(p, "prio", &prio)) {
1460                                 fgetpos(f, &off);
1461                                 continue;
1462                         }
1463                         if (!check_int(p, "prioclass", &prioclass)) {
1464                                 fgetpos(f, &off);
1465                                 continue;
1466                         }
1467                         if (!check_int(p, "direct", &td->odirect)) {
1468                                 fgetpos(f, &off);
1469                                 continue;
1470                         }
1471                         if (!check_int(p, "rate", &td->rate)) {
1472                                 fgetpos(f, &off);
1473                                 continue;
1474                         }
1475                         if (!check_int(p, "ratemin", &td->ratemin)) {
1476                                 fgetpos(f, &off);
1477                                 continue;
1478                         }
1479                         if (!check_int(p, "ratecycle", &td->ratecycle)) {
1480                                 fgetpos(f, &off);
1481                                 continue;
1482                         }
1483                         if (!check_int(p, "delay", &td->delay_sleep)) {
1484                                 fgetpos(f, &off);
1485                                 continue;
1486                         }
1487                         if (!check_int(p, "cpumask", &cpu)) {
1488                                 fill_cpu_mask(td->cpumask, cpu);
1489                                 fgetpos(f, &off);
1490                                 continue;
1491                         }
1492                         if (!check_int(p, "fsync", &td->fsync_blocks)) {
1493                                 fgetpos(f, &off);
1494                                 continue;
1495                         }
1496                         if (!check_int(p, "startdelay", &td->start_delay)) {
1497                                 fgetpos(f, &off);
1498                                 continue;
1499                         }
1500                         if (!check_int(p, "timeout", &td->timeout)) {
1501                                 fgetpos(f, &off);
1502                                 continue;
1503                         }
1504                         if (!check_int(p, "invalidate",&td->invalidate_cache)) {
1505                                 fgetpos(f, &off);
1506                                 continue;
1507                         }
1508                         if (!check_int(p, "aio_depth", &td->aio_depth)) {
1509                                 fgetpos(f, &off);
1510                                 continue;
1511                         }
1512                         if (!check_strcnv(p, "size", &td->file_size)) {
1513                                 fgetpos(f, &off);
1514                                 continue;
1515                         }
1516                         if (!check_strcnv(p, "offset", &td->file_offset)) {
1517                                 fgetpos(f, &off);
1518                                 continue;
1519                         }
1520                         if (!strncmp(p, "sequential", 10)) {
1521                                 td->sequential = 1;
1522                                 fgetpos(f, &off);
1523                                 continue;
1524                         }
1525                         if (!strncmp(p, "random", 6)) {
1526                                 td->sequential = 0;
1527                                 fgetpos(f, &off);
1528                                 continue;
1529                         }
1530                         if (!strncmp(p, "aio", 3)) {
1531                                 td->use_aio = 1;
1532                                 fgetpos(f, &off);
1533                                 continue;
1534                         }
1535                         if (!strncmp(p, "create", 6)) {
1536                                 td->create_file = 1;
1537                                 fgetpos(f, &off);
1538                                 continue;
1539                         }
1540                         if (!strncmp(p, "overwrite", 9)) {
1541                                 td->overwrite = 1;
1542                                 fgetpos(f, &off);
1543                                 continue;
1544                         }
1545                         printf("Client%d: bad option %s\n",td->thread_number,p);
1546                 }
1547                 fsetpos(f, &off);
1548
1549                 if (add_job(td, name, prioclass, prio))
1550                         put_job(td);
1551         }
1552
1553         free(string);
1554         free(name);
1555         fclose(f);
1556         return 0;
1557 }
1558
1559 static int parse_options(int argc, char *argv[])
1560 {
1561         int i;
1562
1563         for (i = 1; i < argc; i++) {
1564                 char *parm = argv[i];
1565
1566                 if (parm[0] != '-')
1567                         break;
1568
1569                 parm++;
1570                 switch (*parm) {
1571                         case 's':
1572                                 parm++;
1573                                 def_thread.sequential = !!atoi(parm);
1574                                 break;
1575                         case 'b':
1576                                 parm++;
1577                                 def_thread.bs = atoi(parm);
1578                                 def_thread.bs <<= 10;
1579                                 if (!def_thread.bs) {
1580                                         printf("bad block size\n");
1581                                         def_thread.bs = DEF_BS;
1582                                 }
1583                                 break;
1584                         case 't':
1585                                 parm++;
1586                                 def_thread.timeout = atoi(parm);
1587                                 break;
1588                         case 'r':
1589                                 parm++;
1590                                 repeatable = !!atoi(parm);
1591                                 break;
1592                         case 'R':
1593                                 parm++;
1594                                 rate_quit = !!atoi(parm);
1595                                 break;
1596                         case 'o':
1597                                 parm++;
1598                                 def_thread.odirect = !!atoi(parm);
1599                                 break;
1600                         case 'f':
1601                                 if (i + 1 >= argc) {
1602                                         printf("-f needs file as arg\n");
1603                                         break;
1604                                 }
1605                                 ini_file = strdup(argv[i+1]);
1606                                 i++;
1607                                 break;
1608                         default:
1609                                 printf("bad option %s\n", argv[i]);
1610                                 break;
1611                 }
1612         }
1613
1614         return i;
1615 }
1616
1617 static void print_thread_status(struct thread_data *td, int nr_running,
1618                                 int t_rate, int m_rate)
1619 {
1620         printf("Threads now running: %d", nr_running);
1621         if (m_rate || t_rate)
1622                 printf(", commitrate %d/%dKiB/sec", t_rate, m_rate);
1623         printf(" : [%s]\r", run_str);
1624         fflush(stdout);
1625 }
1626
1627 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1628 {
1629         int i;
1630
1631         /*
1632          * reap exited threads (TD_EXITED -> TD_REAPED)
1633          */
1634         for (i = 0; i < thread_number; i++) {
1635                 struct thread_data *td = &threads[i];
1636
1637                 if (td->runstate != TD_EXITED)
1638                         continue;
1639
1640                 td->runstate = TD_REAPED;
1641                 run_str[td->thread_number - 1] = '_';
1642                 waitpid(td->pid, NULL, 0);
1643                 (*nr_running)--;
1644                 (*m_rate) -= td->ratemin;
1645                 (*t_rate) -= td->rate;
1646
1647                 if (td->terminate)
1648                         continue;
1649
1650                 print_thread_status(td, *nr_running, *t_rate, *m_rate);
1651         }
1652 }
1653
1654 static void run_threads(char *argv[])
1655 {
1656         struct timeval genesis;
1657         struct thread_data *td;
1658         unsigned long spent;
1659         int i, todo, nr_running, m_rate, t_rate;
1660
1661         gettimeofday(&genesis, NULL);
1662
1663         printf("Starting %d threads\n", thread_number);
1664         fflush(stdout);
1665
1666         signal(SIGINT, sig_handler);
1667
1668         todo = thread_number;
1669         nr_running = 0;
1670         m_rate = t_rate = 0;
1671
1672         while (todo) {
1673                 /*
1674                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1675                  */
1676                 for (i = 0; i < thread_number; i++) {
1677                         td = &threads[i];
1678
1679                         if (td->runstate != TD_NOT_CREATED)
1680                                 continue;
1681
1682                         /*
1683                          * never got a chance to start, killed by other
1684                          * thread for some reason
1685                          */
1686                         if (td->terminate) {
1687                                 todo--;
1688                                 continue;
1689                         }
1690
1691                         if (td->start_delay) {
1692                                 spent = mtime_since_now(&genesis);
1693
1694                                 if (td->start_delay * 1000 > spent)
1695                                         continue;
1696                         }
1697
1698                         td->runstate = TD_CREATED;
1699                         run_str[td->thread_number - 1] = 'C';
1700                         sem_init(&startup_sem, 1, 1);
1701                         todo--;
1702
1703                         if (fork())
1704                                 sem_wait(&startup_sem);
1705                         else {
1706                                 thread_main(shm_id, i, argv);
1707                                 exit(0);
1708                         }
1709                 }
1710
1711                 /*
1712                  * start created threads (TD_CREATED -> TD_STARTED)
1713                  */
1714                 for (i = 0; i < thread_number; i++) {
1715                         struct thread_data *td = &threads[i];
1716
1717                         if (td->runstate != TD_CREATED)
1718                                 continue;
1719
1720                         td->runstate = TD_STARTED;
1721                         run_str[td->thread_number - 1] = '+';
1722                         nr_running++;
1723                         m_rate += td->ratemin;
1724                         t_rate += td->rate;
1725                         sem_post(&td->mutex);
1726
1727                         print_thread_status(td, nr_running, t_rate, m_rate);
1728                 }
1729
1730                 reap_threads(&nr_running, &t_rate, &m_rate);
1731
1732                 if (todo)
1733                         usleep(100000);
1734         }
1735
1736         while (nr_running) {
1737                 reap_threads(&nr_running, &t_rate, &m_rate);
1738                 usleep(10000);
1739         }
1740 }
1741
1742 int setup_thread_area(void)
1743 {
1744         /*
1745          * 1024 is too much on some machines, scale max_jobs if
1746          * we get a failure that looks like too large a shm segment
1747          */
1748         do {
1749                 int s = max_jobs * sizeof(struct thread_data);
1750
1751                 shm_id = shmget(0, s, IPC_CREAT | 0600);
1752                 if (shm_id != -1)
1753                         break;
1754                 if (errno != EINVAL) {
1755                         perror("shmget");
1756                         break;
1757                 }
1758
1759                 max_jobs >>= 1;
1760         } while (max_jobs);
1761
1762         if (shm_id == -1)
1763                 return 1;
1764
1765         threads = shmat(shm_id, NULL, 0);
1766         if (threads == (void *) -1) {
1767                 perror("shmat");
1768                 return 1;
1769         }
1770
1771         atexit(free_shm);
1772         return 0;
1773 }
1774
1775 int main(int argc, char *argv[])
1776 {
1777         static unsigned long max_run[2], min_run[2], total_blocks[2];
1778         static unsigned long max_bw[2], min_bw[2];
1779         static unsigned long read_mb, write_mb, read_agg, write_agg;
1780         int i;
1781
1782         if (setup_thread_area())
1783                 return 1;
1784
1785         if (sched_getaffinity(getpid(), sizeof(cpu_set_t), &def_thread.cpumask) == -1) {
1786                 perror("sched_getaffinity");
1787                 return 1;
1788         }
1789
1790         /*
1791          * fill globals
1792          */
1793         def_thread.ddir = DDIR_READ;
1794         def_thread.bs = DEF_BS;
1795         def_thread.odirect = DEF_ODIRECT;
1796         def_thread.ratecycle = DEF_RATE_CYCLE;
1797         def_thread.sequential = DEF_SEQUENTIAL;
1798         def_thread.timeout = DEF_TIMEOUT;
1799         def_thread.create_file = DEF_CREATE;
1800         def_thread.overwrite = DEF_OVERWRITE;
1801         def_thread.invalidate_cache = DEF_INVALIDATE;
1802
1803         i = parse_options(argc, argv);
1804
1805         if (ini_file) {
1806                 if (parse_jobs_ini(ini_file))
1807                         return 1;
1808         } else
1809                 parse_jobs_cmd(argc, argv, i);
1810
1811         if (!thread_number) {
1812                 printf("Nothing to do\n");
1813                 return 1;
1814         }
1815
1816         run_threads(argv);
1817
1818         min_bw[0] = min_run[0] = ~0UL;
1819         min_bw[1] = min_run[1] = ~0UL;
1820         for (i = 0; i < thread_number; i++) {
1821                 struct thread_data *td = &threads[i];
1822                 unsigned long bw = 0;
1823
1824                 if (td->error)
1825                         goto show_stat;
1826
1827                 if (td->runtime < min_run[td->ddir])
1828                         min_run[td->ddir] = td->runtime;
1829                 if (td->runtime > max_run[td->ddir])
1830                         max_run[td->ddir] = td->runtime;
1831
1832                 if (td->runtime)
1833                         bw = (td->io_blocks * td->bs) / td->runtime;
1834                 if (bw < min_bw[td->ddir])
1835                         min_bw[td->ddir] = bw;
1836                 if (bw > max_bw[td->ddir])
1837                         max_bw[td->ddir] = bw;
1838
1839                 total_blocks[td->ddir] += td->io_blocks;
1840
1841                 if (td_read(td)) {
1842                         read_mb += (td->bs * td->io_blocks) >> 20;
1843                         if (td->runtime)
1844                                 read_agg += (td->io_blocks * td->bs) / td->runtime;
1845                 } else {
1846                         write_mb += (td->bs * td->io_blocks) >> 20;
1847                         if (td->runtime)
1848                                 write_agg += (td->io_blocks * td->bs) / td->runtime;
1849                 }
1850
1851 show_stat:
1852                 show_thread_status(td);
1853         }
1854
1855         printf("\nRun status:\n");
1856         if (max_run[DDIR_READ])
1857                 printf("   READ: io=%luMiB, aggrb=%lu, minb=%lu, maxb=%lu, mint=%lumsec, maxt=%lumsec\n", read_mb, read_agg, min_bw[0], max_bw[0], min_run[0], max_run[0]);
1858         if (max_run[DDIR_WRITE])
1859                 printf("  WRITE: io=%luMiB, aggrb=%lu, minb=%lu, maxb=%lu, mint=%lumsec, maxt=%lumsec\n", write_mb, write_agg, min_bw[1], max_bw[1], min_run[1], max_run[1]);
1860
1861         return 0;
1862 }