server: start conversion of data structures to network friendly types
[fio.git] / fio.c
1 /*
2  * fio - the flexible io tester
3  *
4  * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
5  * Copyright (C) 2006 Jens Axboe <axboe@kernel.dk>
6  *
7  * The license below covers all files distributed with fio unless otherwise
8  * noted in the file itself.
9  *
10  *  This program is free software; you can redistribute it and/or modify
11  *  it under the terms of the GNU General Public License version 2 as
12  *  published by the Free Software Foundation.
13  *
14  *  This program is distributed in the hope that it will be useful,
15  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
16  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
17  *  GNU General Public License for more details.
18  *
19  *  You should have received a copy of the GNU General Public License
20  *  along with this program; if not, write to the Free Software
21  *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
22  *
23  */
24 #include <unistd.h>
25 #include <fcntl.h>
26 #include <string.h>
27 #include <limits.h>
28 #include <signal.h>
29 #include <time.h>
30 #include <locale.h>
31 #include <assert.h>
32 #include <time.h>
33 #include <sys/stat.h>
34 #include <sys/wait.h>
35 #include <sys/ipc.h>
36 #include <sys/shm.h>
37 #include <sys/mman.h>
38
39 #include "fio.h"
40 #include "hash.h"
41 #include "smalloc.h"
42 #include "verify.h"
43 #include "trim.h"
44 #include "diskutil.h"
45 #include "cgroup.h"
46 #include "profile.h"
47 #include "lib/rand.h"
48 #include "memalign.h"
49 #include "server.h"
50
51 unsigned long page_mask;
52 unsigned long page_size;
53
54 #define PAGE_ALIGN(buf) \
55         (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
56
57 int groupid = 0;
58 int thread_number = 0;
59 int nr_process = 0;
60 int nr_thread = 0;
61 int shm_id = 0;
62 int temp_stall_ts;
63 unsigned long done_secs = 0;
64
65 static struct fio_mutex *startup_mutex;
66 static struct fio_mutex *writeout_mutex;
67 static volatile int fio_abort;
68 static int exit_value;
69 static pthread_t gtod_thread;
70 static pthread_t disk_util_thread;
71 static struct flist_head *cgroup_list;
72 static char *cgroup_mnt;
73
74 unsigned long arch_flags = 0;
75
76 struct io_log *agg_io_log[2];
77
78 #define TERMINATE_ALL           (-1)
79 #define JOB_START_TIMEOUT       (5 * 1000)
80
81 void td_set_runstate(struct thread_data *td, int runstate)
82 {
83         if (td->runstate == runstate)
84                 return;
85
86         dprint(FD_PROCESS, "pid=%d: runstate %d -> %d\n", (int) td->pid,
87                                                 td->runstate, runstate);
88         td->runstate = runstate;
89 }
90
91 static void terminate_threads(int group_id)
92 {
93         struct thread_data *td;
94         int i;
95
96         dprint(FD_PROCESS, "terminate group_id=%d\n", group_id);
97
98         for_each_td(td, i) {
99                 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
100                         dprint(FD_PROCESS, "setting terminate on %s/%d\n",
101                                                 td->o.name, (int) td->pid);
102                         td->terminate = 1;
103                         td->o.start_delay = 0;
104
105                         /*
106                          * if the thread is running, just let it exit
107                          */
108                         if (!td->pid)
109                                 continue;
110                         else if (td->runstate < TD_RAMP)
111                                 kill(td->pid, SIGTERM);
112                         else {
113                                 struct ioengine_ops *ops = td->io_ops;
114
115                                 if (ops && (ops->flags & FIO_SIGTERM))
116                                         kill(td->pid, SIGTERM);
117                         }
118                 }
119         }
120 }
121
122 static void sig_int(int sig)
123 {
124         if (threads) {
125                 log_info("\nfio: terminating on signal %d\n", sig);
126                 exit_backend = 1;
127                 fflush(stdout);
128                 exit_value = 128;
129                 terminate_threads(TERMINATE_ALL);
130         }
131 }
132
133 static void *disk_thread_main(void *data)
134 {
135         fio_mutex_up(startup_mutex);
136
137         while (threads) {
138                 usleep(DISK_UTIL_MSEC * 1000);
139                 if (!threads)
140                         break;
141                 update_io_ticks();
142                 print_thread_status();
143         }
144
145         return NULL;
146 }
147
148 static int create_disk_util_thread(void)
149 {
150         int ret;
151
152         ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
153         if (ret) {
154                 log_err("Can't create disk util thread: %s\n", strerror(ret));
155                 return 1;
156         }
157
158         ret = pthread_detach(disk_util_thread);
159         if (ret) {
160                 log_err("Can't detatch disk util thread: %s\n", strerror(ret));
161                 return 1;
162         }
163
164         dprint(FD_MUTEX, "wait on startup_mutex\n");
165         fio_mutex_down(startup_mutex);
166         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
167         return 0;
168 }
169
170 static void set_sig_handlers(void)
171 {
172         struct sigaction act;
173
174         memset(&act, 0, sizeof(act));
175         act.sa_handler = sig_int;
176         act.sa_flags = SA_RESTART;
177         sigaction(SIGINT, &act, NULL);
178
179         memset(&act, 0, sizeof(act));
180         act.sa_handler = sig_int;
181         act.sa_flags = SA_RESTART;
182         sigaction(SIGTERM, &act, NULL);
183 }
184
185 /*
186  * Check if we are above the minimum rate given.
187  */
188 static int __check_min_rate(struct thread_data *td, struct timeval *now,
189                             enum fio_ddir ddir)
190 {
191         unsigned long long bytes = 0;
192         unsigned long iops = 0;
193         unsigned long spent;
194         unsigned long rate;
195         unsigned int ratemin = 0;
196         unsigned int rate_iops = 0;
197         unsigned int rate_iops_min = 0;
198
199         assert(ddir_rw(ddir));
200
201         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
202                 return 0;
203
204         /*
205          * allow a 2 second settle period in the beginning
206          */
207         if (mtime_since(&td->start, now) < 2000)
208                 return 0;
209
210         iops += td->io_blocks[ddir];
211         bytes += td->this_io_bytes[ddir];
212         ratemin += td->o.ratemin[ddir];
213         rate_iops += td->o.rate_iops[ddir];
214         rate_iops_min += td->o.rate_iops_min[ddir];
215
216         /*
217          * if rate blocks is set, sample is running
218          */
219         if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
220                 spent = mtime_since(&td->lastrate[ddir], now);
221                 if (spent < td->o.ratecycle)
222                         return 0;
223
224                 if (td->o.rate[ddir]) {
225                         /*
226                          * check bandwidth specified rate
227                          */
228                         if (bytes < td->rate_bytes[ddir]) {
229                                 log_err("%s: min rate %u not met\n", td->o.name,
230                                                                 ratemin);
231                                 return 1;
232                         } else {
233                                 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
234                                 if (rate < ratemin ||
235                                     bytes < td->rate_bytes[ddir]) {
236                                         log_err("%s: min rate %u not met, got"
237                                                 " %luKB/sec\n", td->o.name,
238                                                         ratemin, rate);
239                                         return 1;
240                                 }
241                         }
242                 } else {
243                         /*
244                          * checks iops specified rate
245                          */
246                         if (iops < rate_iops) {
247                                 log_err("%s: min iops rate %u not met\n",
248                                                 td->o.name, rate_iops);
249                                 return 1;
250                         } else {
251                                 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
252                                 if (rate < rate_iops_min ||
253                                     iops < td->rate_blocks[ddir]) {
254                                         log_err("%s: min iops rate %u not met,"
255                                                 " got %lu\n", td->o.name,
256                                                         rate_iops_min, rate);
257                                 }
258                         }
259                 }
260         }
261
262         td->rate_bytes[ddir] = bytes;
263         td->rate_blocks[ddir] = iops;
264         memcpy(&td->lastrate[ddir], now, sizeof(*now));
265         return 0;
266 }
267
268 static int check_min_rate(struct thread_data *td, struct timeval *now,
269                           unsigned long *bytes_done)
270 {
271         int ret = 0;
272
273         if (bytes_done[0])
274                 ret |= __check_min_rate(td, now, 0);
275         if (bytes_done[1])
276                 ret |= __check_min_rate(td, now, 1);
277
278         return ret;
279 }
280
281 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
282 {
283         if (!td->o.timeout)
284                 return 0;
285         if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
286                 return 1;
287
288         return 0;
289 }
290
291 /*
292  * When job exits, we can cancel the in-flight IO if we are using async
293  * io. Attempt to do so.
294  */
295 static void cleanup_pending_aio(struct thread_data *td)
296 {
297         struct flist_head *entry, *n;
298         struct io_u *io_u;
299         int r;
300
301         /*
302          * get immediately available events, if any
303          */
304         r = io_u_queued_complete(td, 0, NULL);
305         if (r < 0)
306                 return;
307
308         /*
309          * now cancel remaining active events
310          */
311         if (td->io_ops->cancel) {
312                 flist_for_each_safe(entry, n, &td->io_u_busylist) {
313                         io_u = flist_entry(entry, struct io_u, list);
314
315                         /*
316                          * if the io_u isn't in flight, then that generally
317                          * means someone leaked an io_u. complain but fix
318                          * it up, so we don't stall here.
319                          */
320                         if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
321                                 log_err("fio: non-busy IO on busy list\n");
322                                 put_io_u(td, io_u);
323                         } else {
324                                 r = td->io_ops->cancel(td, io_u);
325                                 if (!r)
326                                         put_io_u(td, io_u);
327                         }
328                 }
329         }
330
331         if (td->cur_depth)
332                 r = io_u_queued_complete(td, td->cur_depth, NULL);
333 }
334
335 /*
336  * Helper to handle the final sync of a file. Works just like the normal
337  * io path, just does everything sync.
338  */
339 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
340 {
341         struct io_u *io_u = __get_io_u(td);
342         int ret;
343
344         if (!io_u)
345                 return 1;
346
347         io_u->ddir = DDIR_SYNC;
348         io_u->file = f;
349
350         if (td_io_prep(td, io_u)) {
351                 put_io_u(td, io_u);
352                 return 1;
353         }
354
355 requeue:
356         ret = td_io_queue(td, io_u);
357         if (ret < 0) {
358                 td_verror(td, io_u->error, "td_io_queue");
359                 put_io_u(td, io_u);
360                 return 1;
361         } else if (ret == FIO_Q_QUEUED) {
362                 if (io_u_queued_complete(td, 1, NULL) < 0)
363                         return 1;
364         } else if (ret == FIO_Q_COMPLETED) {
365                 if (io_u->error) {
366                         td_verror(td, io_u->error, "td_io_queue");
367                         return 1;
368                 }
369
370                 if (io_u_sync_complete(td, io_u, NULL) < 0)
371                         return 1;
372         } else if (ret == FIO_Q_BUSY) {
373                 if (td_io_commit(td))
374                         return 1;
375                 goto requeue;
376         }
377
378         return 0;
379 }
380
381 static inline void __update_tv_cache(struct thread_data *td)
382 {
383         fio_gettime(&td->tv_cache, NULL);
384 }
385
386 static inline void update_tv_cache(struct thread_data *td)
387 {
388         if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
389                 __update_tv_cache(td);
390 }
391
392 static int break_on_this_error(struct thread_data *td, int *retptr)
393 {
394         int ret = *retptr;
395
396         if (ret < 0 || td->error) {
397                 int err;
398
399                 if (!td->o.continue_on_error)
400                         return 1;
401
402                 if (ret < 0)
403                         err = -ret;
404                 else
405                         err = td->error;
406
407                 if (td_non_fatal_error(err)) {
408                         /*
409                          * Continue with the I/Os in case of
410                          * a non fatal error.
411                          */
412                         update_error_count(td, err);
413                         td_clear_error(td);
414                         *retptr = 0;
415                         return 0;
416                 } else if (td->o.fill_device && err == ENOSPC) {
417                         /*
418                          * We expect to hit this error if
419                          * fill_device option is set.
420                          */
421                         td_clear_error(td);
422                         td->terminate = 1;
423                         return 1;
424                 } else {
425                         /*
426                          * Stop the I/O in case of a fatal
427                          * error.
428                          */
429                         update_error_count(td, err);
430                         return 1;
431                 }
432         }
433
434         return 0;
435 }
436
437 /*
438  * The main verify engine. Runs over the writes we previously submitted,
439  * reads the blocks back in, and checks the crc/md5 of the data.
440  */
441 static void do_verify(struct thread_data *td)
442 {
443         struct fio_file *f;
444         struct io_u *io_u;
445         int ret, min_events;
446         unsigned int i;
447
448         dprint(FD_VERIFY, "starting loop\n");
449
450         /*
451          * sync io first and invalidate cache, to make sure we really
452          * read from disk.
453          */
454         for_each_file(td, f, i) {
455                 if (!fio_file_open(f))
456                         continue;
457                 if (fio_io_sync(td, f))
458                         break;
459                 if (file_invalidate_cache(td, f))
460                         break;
461         }
462
463         if (td->error)
464                 return;
465
466         td_set_runstate(td, TD_VERIFYING);
467
468         io_u = NULL;
469         while (!td->terminate) {
470                 int ret2, full;
471
472                 update_tv_cache(td);
473
474                 if (runtime_exceeded(td, &td->tv_cache)) {
475                         __update_tv_cache(td);
476                         if (runtime_exceeded(td, &td->tv_cache)) {
477                                 td->terminate = 1;
478                                 break;
479                         }
480                 }
481
482                 io_u = __get_io_u(td);
483                 if (!io_u)
484                         break;
485
486                 if (get_next_verify(td, io_u)) {
487                         put_io_u(td, io_u);
488                         break;
489                 }
490
491                 if (td_io_prep(td, io_u)) {
492                         put_io_u(td, io_u);
493                         break;
494                 }
495
496                 if (td->o.verify_async)
497                         io_u->end_io = verify_io_u_async;
498                 else
499                         io_u->end_io = verify_io_u;
500
501                 ret = td_io_queue(td, io_u);
502                 switch (ret) {
503                 case FIO_Q_COMPLETED:
504                         if (io_u->error) {
505                                 ret = -io_u->error;
506                                 clear_io_u(td, io_u);
507                         } else if (io_u->resid) {
508                                 int bytes = io_u->xfer_buflen - io_u->resid;
509
510                                 /*
511                                  * zero read, fail
512                                  */
513                                 if (!bytes) {
514                                         td_verror(td, EIO, "full resid");
515                                         put_io_u(td, io_u);
516                                         break;
517                                 }
518
519                                 io_u->xfer_buflen = io_u->resid;
520                                 io_u->xfer_buf += bytes;
521                                 io_u->offset += bytes;
522
523                                 if (ddir_rw(io_u->ddir))
524                                         td->ts.short_io_u[io_u->ddir]++;
525
526                                 f = io_u->file;
527                                 if (io_u->offset == f->real_file_size)
528                                         goto sync_done;
529
530                                 requeue_io_u(td, &io_u);
531                         } else {
532 sync_done:
533                                 ret = io_u_sync_complete(td, io_u, NULL);
534                                 if (ret < 0)
535                                         break;
536                         }
537                         continue;
538                 case FIO_Q_QUEUED:
539                         break;
540                 case FIO_Q_BUSY:
541                         requeue_io_u(td, &io_u);
542                         ret2 = td_io_commit(td);
543                         if (ret2 < 0)
544                                 ret = ret2;
545                         break;
546                 default:
547                         assert(ret < 0);
548                         td_verror(td, -ret, "td_io_queue");
549                         break;
550                 }
551
552                 if (break_on_this_error(td, &ret))
553                         break;
554
555                 /*
556                  * if we can queue more, do so. but check if there are
557                  * completed io_u's first. Note that we can get BUSY even
558                  * without IO queued, if the system is resource starved.
559                  */
560                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
561                 if (full || !td->o.iodepth_batch_complete) {
562                         min_events = min(td->o.iodepth_batch_complete,
563                                          td->cur_depth);
564                         if (full && !min_events && td->o.iodepth_batch_complete != 0)
565                                 min_events = 1;
566
567                         do {
568                                 /*
569                                  * Reap required number of io units, if any,
570                                  * and do the verification on them through
571                                  * the callback handler
572                                  */
573                                 if (io_u_queued_complete(td, min_events, NULL) < 0) {
574                                         ret = -1;
575                                         break;
576                                 }
577                         } while (full && (td->cur_depth > td->o.iodepth_low));
578                 }
579                 if (ret < 0)
580                         break;
581         }
582
583         if (!td->error) {
584                 min_events = td->cur_depth;
585
586                 if (min_events)
587                         ret = io_u_queued_complete(td, min_events, NULL);
588         } else
589                 cleanup_pending_aio(td);
590
591         td_set_runstate(td, TD_RUNNING);
592
593         dprint(FD_VERIFY, "exiting loop\n");
594 }
595
596 /*
597  * Main IO worker function. It retrieves io_u's to process and queues
598  * and reaps them, checking for rate and errors along the way.
599  */
600 static void do_io(struct thread_data *td)
601 {
602         unsigned int i;
603         int ret = 0;
604
605         if (in_ramp_time(td))
606                 td_set_runstate(td, TD_RAMP);
607         else
608                 td_set_runstate(td, TD_RUNNING);
609
610         while ( (td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
611                 (!flist_empty(&td->trim_list)) ||
612                 ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) ) {
613                 struct timeval comp_time;
614                 unsigned long bytes_done[2] = { 0, 0 };
615                 int min_evts = 0;
616                 struct io_u *io_u;
617                 int ret2, full;
618
619                 if (td->terminate)
620                         break;
621
622                 update_tv_cache(td);
623
624                 if (runtime_exceeded(td, &td->tv_cache)) {
625                         __update_tv_cache(td);
626                         if (runtime_exceeded(td, &td->tv_cache)) {
627                                 td->terminate = 1;
628                                 break;
629                         }
630                 }
631
632                 io_u = get_io_u(td);
633                 if (!io_u)
634                         break;
635
636                 /*
637                  * Add verification end_io handler, if asked to verify
638                  * a previously written file.
639                  */
640                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
641                     !td_rw(td)) {
642                         if (td->o.verify_async)
643                                 io_u->end_io = verify_io_u_async;
644                         else
645                                 io_u->end_io = verify_io_u;
646                         td_set_runstate(td, TD_VERIFYING);
647                 } else if (in_ramp_time(td))
648                         td_set_runstate(td, TD_RAMP);
649                 else
650                         td_set_runstate(td, TD_RUNNING);
651
652                 ret = td_io_queue(td, io_u);
653                 switch (ret) {
654                 case FIO_Q_COMPLETED:
655                         if (io_u->error) {
656                                 ret = -io_u->error;
657                                 clear_io_u(td, io_u);
658                         } else if (io_u->resid) {
659                                 int bytes = io_u->xfer_buflen - io_u->resid;
660                                 struct fio_file *f = io_u->file;
661
662                                 /*
663                                  * zero read, fail
664                                  */
665                                 if (!bytes) {
666                                         td_verror(td, EIO, "full resid");
667                                         put_io_u(td, io_u);
668                                         break;
669                                 }
670
671                                 io_u->xfer_buflen = io_u->resid;
672                                 io_u->xfer_buf += bytes;
673                                 io_u->offset += bytes;
674
675                                 if (ddir_rw(io_u->ddir))
676                                         td->ts.short_io_u[io_u->ddir]++;
677
678                                 if (io_u->offset == f->real_file_size)
679                                         goto sync_done;
680
681                                 requeue_io_u(td, &io_u);
682                         } else {
683 sync_done:
684                                 if (__should_check_rate(td, 0) ||
685                                     __should_check_rate(td, 1))
686                                         fio_gettime(&comp_time, NULL);
687
688                                 ret = io_u_sync_complete(td, io_u, bytes_done);
689                                 if (ret < 0)
690                                         break;
691                         }
692                         break;
693                 case FIO_Q_QUEUED:
694                         /*
695                          * if the engine doesn't have a commit hook,
696                          * the io_u is really queued. if it does have such
697                          * a hook, it has to call io_u_queued() itself.
698                          */
699                         if (td->io_ops->commit == NULL)
700                                 io_u_queued(td, io_u);
701                         break;
702                 case FIO_Q_BUSY:
703                         requeue_io_u(td, &io_u);
704                         ret2 = td_io_commit(td);
705                         if (ret2 < 0)
706                                 ret = ret2;
707                         break;
708                 default:
709                         assert(ret < 0);
710                         put_io_u(td, io_u);
711                         break;
712                 }
713
714                 if (break_on_this_error(td, &ret))
715                         break;
716
717                 /*
718                  * See if we need to complete some commands. Note that we
719                  * can get BUSY even without IO queued, if the system is
720                  * resource starved.
721                  */
722                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
723                 if (full || !td->o.iodepth_batch_complete) {
724                         min_evts = min(td->o.iodepth_batch_complete,
725                                         td->cur_depth);
726                         if (full && !min_evts && td->o.iodepth_batch_complete != 0)
727                                 min_evts = 1;
728
729                         if (__should_check_rate(td, 0) ||
730                             __should_check_rate(td, 1))
731                                 fio_gettime(&comp_time, NULL);
732
733                         do {
734                                 ret = io_u_queued_complete(td, min_evts, bytes_done);
735                                 if (ret < 0)
736                                         break;
737
738                         } while (full && (td->cur_depth > td->o.iodepth_low));
739                 }
740
741                 if (ret < 0)
742                         break;
743                 if (!(bytes_done[0] + bytes_done[1]))
744                         continue;
745
746                 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
747                         if (check_min_rate(td, &comp_time, bytes_done)) {
748                                 if (exitall_on_terminate)
749                                         terminate_threads(td->groupid);
750                                 td_verror(td, EIO, "check_min_rate");
751                                 break;
752                         }
753                 }
754
755                 if (td->o.thinktime) {
756                         unsigned long long b;
757
758                         b = td->io_blocks[0] + td->io_blocks[1];
759                         if (!(b % td->o.thinktime_blocks)) {
760                                 int left;
761
762                                 if (td->o.thinktime_spin)
763                                         usec_spin(td->o.thinktime_spin);
764
765                                 left = td->o.thinktime - td->o.thinktime_spin;
766                                 if (left)
767                                         usec_sleep(td, left);
768                         }
769                 }
770         }
771
772         if (td->trim_entries)
773                 printf("trim entries %ld\n", td->trim_entries);
774
775         if (td->o.fill_device && td->error == ENOSPC) {
776                 td->error = 0;
777                 td->terminate = 1;
778         }
779         if (!td->error) {
780                 struct fio_file *f;
781
782                 i = td->cur_depth;
783                 if (i) {
784                         ret = io_u_queued_complete(td, i, NULL);
785                         if (td->o.fill_device && td->error == ENOSPC)
786                                 td->error = 0;
787                 }
788
789                 if (should_fsync(td) && td->o.end_fsync) {
790                         td_set_runstate(td, TD_FSYNCING);
791
792                         for_each_file(td, f, i) {
793                                 if (!fio_file_open(f))
794                                         continue;
795                                 fio_io_sync(td, f);
796                         }
797                 }
798         } else
799                 cleanup_pending_aio(td);
800
801         /*
802          * stop job if we failed doing any IO
803          */
804         if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
805                 td->done = 1;
806 }
807
808 static void cleanup_io_u(struct thread_data *td)
809 {
810         struct flist_head *entry, *n;
811         struct io_u *io_u;
812
813         flist_for_each_safe(entry, n, &td->io_u_freelist) {
814                 io_u = flist_entry(entry, struct io_u, list);
815
816                 flist_del(&io_u->list);
817                 fio_memfree(io_u, sizeof(*io_u));
818         }
819
820         free_io_mem(td);
821 }
822
823 static int init_io_u(struct thread_data *td)
824 {
825         struct io_u *io_u;
826         unsigned int max_bs;
827         int cl_align, i, max_units;
828         char *p;
829
830         max_units = td->o.iodepth;
831         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
832         td->orig_buffer_size = (unsigned long long) max_bs
833                                         * (unsigned long long) max_units;
834
835         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
836                 unsigned long bs;
837
838                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
839                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
840         }
841
842         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
843                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
844                 return 1;
845         }
846
847         if (allocate_io_mem(td))
848                 return 1;
849
850         if (td->o.odirect || td->o.mem_align ||
851             (td->io_ops->flags & FIO_RAWIO))
852                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
853         else
854                 p = td->orig_buffer;
855
856         cl_align = os_cache_line_size();
857
858         for (i = 0; i < max_units; i++) {
859                 void *ptr;
860
861                 if (td->terminate)
862                         return 1;
863
864                 ptr = fio_memalign(cl_align, sizeof(*io_u));
865                 if (!ptr) {
866                         log_err("fio: unable to allocate aligned memory\n");
867                         break;
868                 }
869
870                 io_u = ptr;
871                 memset(io_u, 0, sizeof(*io_u));
872                 INIT_FLIST_HEAD(&io_u->list);
873                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
874
875                 if (!(td->io_ops->flags & FIO_NOIO)) {
876                         io_u->buf = p + max_bs * i;
877                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
878
879                         if (td_write(td))
880                                 io_u_fill_buffer(td, io_u, max_bs);
881                         if (td_write(td) && td->o.verify_pattern_bytes) {
882                                 /*
883                                  * Fill the buffer with the pattern if we are
884                                  * going to be doing writes.
885                                  */
886                                 fill_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
887                         }
888                 }
889
890                 io_u->index = i;
891                 io_u->flags = IO_U_F_FREE;
892                 flist_add(&io_u->list, &td->io_u_freelist);
893         }
894
895         return 0;
896 }
897
898 static int switch_ioscheduler(struct thread_data *td)
899 {
900         char tmp[256], tmp2[128];
901         FILE *f;
902         int ret;
903
904         if (td->io_ops->flags & FIO_DISKLESSIO)
905                 return 0;
906
907         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
908
909         f = fopen(tmp, "r+");
910         if (!f) {
911                 if (errno == ENOENT) {
912                         log_err("fio: os or kernel doesn't support IO scheduler"
913                                 " switching\n");
914                         return 0;
915                 }
916                 td_verror(td, errno, "fopen iosched");
917                 return 1;
918         }
919
920         /*
921          * Set io scheduler.
922          */
923         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
924         if (ferror(f) || ret != 1) {
925                 td_verror(td, errno, "fwrite");
926                 fclose(f);
927                 return 1;
928         }
929
930         rewind(f);
931
932         /*
933          * Read back and check that the selected scheduler is now the default.
934          */
935         ret = fread(tmp, 1, sizeof(tmp), f);
936         if (ferror(f) || ret < 0) {
937                 td_verror(td, errno, "fread");
938                 fclose(f);
939                 return 1;
940         }
941
942         sprintf(tmp2, "[%s]", td->o.ioscheduler);
943         if (!strstr(tmp, tmp2)) {
944                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
945                 td_verror(td, EINVAL, "iosched_switch");
946                 fclose(f);
947                 return 1;
948         }
949
950         fclose(f);
951         return 0;
952 }
953
954 static int keep_running(struct thread_data *td)
955 {
956         unsigned long long io_done;
957
958         if (td->done)
959                 return 0;
960         if (td->o.time_based)
961                 return 1;
962         if (td->o.loops) {
963                 td->o.loops--;
964                 return 1;
965         }
966
967         io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
968                         + td->io_skip_bytes;
969         if (io_done < td->o.size)
970                 return 1;
971
972         return 0;
973 }
974
975 static void reset_io_counters(struct thread_data *td)
976 {
977         td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
978         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
979         td->zone_bytes = 0;
980         td->rate_bytes[0] = td->rate_bytes[1] = 0;
981         td->rate_blocks[0] = td->rate_blocks[1] = 0;
982
983         td->last_was_sync = 0;
984
985         /*
986          * reset file done count if we are to start over
987          */
988         if (td->o.time_based || td->o.loops)
989                 td->nr_done_files = 0;
990 }
991
992 void reset_all_stats(struct thread_data *td)
993 {
994         struct timeval tv;
995         int i;
996
997         reset_io_counters(td);
998
999         for (i = 0; i < 2; i++) {
1000                 td->io_bytes[i] = 0;
1001                 td->io_blocks[i] = 0;
1002                 td->io_issues[i] = 0;
1003                 td->ts.total_io_u[i] = 0;
1004         }
1005
1006         fio_gettime(&tv, NULL);
1007         td->ts.runtime[0] = 0;
1008         td->ts.runtime[1] = 0;
1009         memcpy(&td->epoch, &tv, sizeof(tv));
1010         memcpy(&td->start, &tv, sizeof(tv));
1011 }
1012
1013 static void clear_io_state(struct thread_data *td)
1014 {
1015         struct fio_file *f;
1016         unsigned int i;
1017
1018         reset_io_counters(td);
1019
1020         close_files(td);
1021         for_each_file(td, f, i)
1022                 fio_file_clear_done(f);
1023
1024         /*
1025          * Set the same seed to get repeatable runs
1026          */
1027         td_fill_rand_seeds(td);
1028 }
1029
1030 static int exec_string(const char *string)
1031 {
1032         int ret, newlen = strlen(string) + 1 + 8;
1033         char *str;
1034
1035         str = malloc(newlen);
1036         sprintf(str, "sh -c %s", string);
1037
1038         ret = system(str);
1039         if (ret == -1)
1040                 log_err("fio: exec of cmd <%s> failed\n", str);
1041
1042         free(str);
1043         return ret;
1044 }
1045
1046 /*
1047  * Entry point for the thread based jobs. The process based jobs end up
1048  * here as well, after a little setup.
1049  */
1050 static void *thread_main(void *data)
1051 {
1052         unsigned long long elapsed;
1053         struct thread_data *td = data;
1054         pthread_condattr_t attr;
1055         int clear_state;
1056
1057         if (!td->o.use_thread) {
1058                 setsid();
1059                 td->pid = getpid();
1060         } else
1061                 td->pid = gettid();
1062
1063         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1064
1065         INIT_FLIST_HEAD(&td->io_u_freelist);
1066         INIT_FLIST_HEAD(&td->io_u_busylist);
1067         INIT_FLIST_HEAD(&td->io_u_requeues);
1068         INIT_FLIST_HEAD(&td->io_log_list);
1069         INIT_FLIST_HEAD(&td->io_hist_list);
1070         INIT_FLIST_HEAD(&td->verify_list);
1071         INIT_FLIST_HEAD(&td->trim_list);
1072         pthread_mutex_init(&td->io_u_lock, NULL);
1073         td->io_hist_tree = RB_ROOT;
1074
1075         pthread_condattr_init(&attr);
1076         pthread_cond_init(&td->verify_cond, &attr);
1077         pthread_cond_init(&td->free_cond, &attr);
1078
1079         td_set_runstate(td, TD_INITIALIZED);
1080         dprint(FD_MUTEX, "up startup_mutex\n");
1081         fio_mutex_up(startup_mutex);
1082         dprint(FD_MUTEX, "wait on td->mutex\n");
1083         fio_mutex_down(td->mutex);
1084         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1085
1086         /*
1087          * the ->mutex mutex is now no longer used, close it to avoid
1088          * eating a file descriptor
1089          */
1090         fio_mutex_remove(td->mutex);
1091
1092         /*
1093          * A new gid requires privilege, so we need to do this before setting
1094          * the uid.
1095          */
1096         if (td->o.gid != -1U && setgid(td->o.gid)) {
1097                 td_verror(td, errno, "setgid");
1098                 goto err;
1099         }
1100         if (td->o.uid != -1U && setuid(td->o.uid)) {
1101                 td_verror(td, errno, "setuid");
1102                 goto err;
1103         }
1104
1105         /*
1106          * If we have a gettimeofday() thread, make sure we exclude that
1107          * thread from this job
1108          */
1109         if (td->o.gtod_cpu)
1110                 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1111
1112         /*
1113          * Set affinity first, in case it has an impact on the memory
1114          * allocations.
1115          */
1116         if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1117                 td_verror(td, errno, "cpu_set_affinity");
1118                 goto err;
1119         }
1120
1121         /*
1122          * May alter parameters that init_io_u() will use, so we need to
1123          * do this first.
1124          */
1125         if (init_iolog(td))
1126                 goto err;
1127
1128         if (init_io_u(td))
1129                 goto err;
1130
1131         if (td->o.verify_async && verify_async_init(td))
1132                 goto err;
1133
1134         if (td->ioprio_set) {
1135                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1136                         td_verror(td, errno, "ioprio_set");
1137                         goto err;
1138                 }
1139         }
1140
1141         if (td->o.cgroup_weight && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1142                 goto err;
1143
1144         if (nice(td->o.nice) == -1) {
1145                 td_verror(td, errno, "nice");
1146                 goto err;
1147         }
1148
1149         if (td->o.ioscheduler && switch_ioscheduler(td))
1150                 goto err;
1151
1152         if (!td->o.create_serialize && setup_files(td))
1153                 goto err;
1154
1155         if (td_io_init(td))
1156                 goto err;
1157
1158         if (init_random_map(td))
1159                 goto err;
1160
1161         if (td->o.exec_prerun) {
1162                 if (exec_string(td->o.exec_prerun))
1163                         goto err;
1164         }
1165
1166         if (td->o.pre_read) {
1167                 if (pre_read_files(td) < 0)
1168                         goto err;
1169         }
1170
1171         fio_gettime(&td->epoch, NULL);
1172         getrusage(RUSAGE_SELF, &td->ts.ru_start);
1173
1174         clear_state = 0;
1175         while (keep_running(td)) {
1176                 fio_gettime(&td->start, NULL);
1177                 memcpy(&td->ts.stat_sample_time[0], &td->start,
1178                                 sizeof(td->start));
1179                 memcpy(&td->ts.stat_sample_time[1], &td->start,
1180                                 sizeof(td->start));
1181                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1182
1183                 if (td->o.ratemin[0] || td->o.ratemin[1])
1184                         memcpy(&td->lastrate, &td->ts.stat_sample_time,
1185                                                         sizeof(td->lastrate));
1186
1187                 if (clear_state)
1188                         clear_io_state(td);
1189
1190                 prune_io_piece_log(td);
1191
1192                 do_io(td);
1193
1194                 clear_state = 1;
1195
1196                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1197                         elapsed = utime_since_now(&td->start);
1198                         td->ts.runtime[DDIR_READ] += elapsed;
1199                 }
1200                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1201                         elapsed = utime_since_now(&td->start);
1202                         td->ts.runtime[DDIR_WRITE] += elapsed;
1203                 }
1204
1205                 if (td->error || td->terminate)
1206                         break;
1207
1208                 if (!td->o.do_verify ||
1209                     td->o.verify == VERIFY_NONE ||
1210                     (td->io_ops->flags & FIO_UNIDIR))
1211                         continue;
1212
1213                 clear_io_state(td);
1214
1215                 fio_gettime(&td->start, NULL);
1216
1217                 do_verify(td);
1218
1219                 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1220
1221                 if (td->error || td->terminate)
1222                         break;
1223         }
1224
1225         update_rusage_stat(td);
1226         td->ts.runtime[0] = (td->ts.runtime[0] + 999) / 1000;
1227         td->ts.runtime[1] = (td->ts.runtime[1] + 999) / 1000;
1228         td->ts.total_run_time = mtime_since_now(&td->epoch);
1229         td->ts.io_bytes[0] = td->io_bytes[0];
1230         td->ts.io_bytes[1] = td->io_bytes[1];
1231
1232         fio_mutex_down(writeout_mutex);
1233         if (td->bw_log) {
1234                 if (td->o.bw_log_file) {
1235                         finish_log_named(td, td->bw_log,
1236                                                 td->o.bw_log_file, "bw");
1237                 } else
1238                         finish_log(td, td->bw_log, "bw");
1239         }
1240         if (td->lat_log) {
1241                 if (td->o.lat_log_file) {
1242                         finish_log_named(td, td->lat_log,
1243                                                 td->o.lat_log_file, "lat");
1244                 } else
1245                         finish_log(td, td->lat_log, "lat");
1246         }
1247         if (td->slat_log) {
1248                 if (td->o.lat_log_file) {
1249                         finish_log_named(td, td->slat_log,
1250                                                 td->o.lat_log_file, "slat");
1251                 } else
1252                         finish_log(td, td->slat_log, "slat");
1253         }
1254         if (td->clat_log) {
1255                 if (td->o.lat_log_file) {
1256                         finish_log_named(td, td->clat_log,
1257                                                 td->o.lat_log_file, "clat");
1258                 } else
1259                         finish_log(td, td->clat_log, "clat");
1260         }
1261         fio_mutex_up(writeout_mutex);
1262         if (td->o.exec_postrun)
1263                 exec_string(td->o.exec_postrun);
1264
1265         if (exitall_on_terminate)
1266                 terminate_threads(td->groupid);
1267
1268 err:
1269         if (td->error)
1270                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1271                                                         td->verror);
1272
1273         if (td->o.verify_async)
1274                 verify_async_exit(td);
1275
1276         close_and_free_files(td);
1277         close_ioengine(td);
1278         cleanup_io_u(td);
1279         cgroup_shutdown(td, &cgroup_mnt);
1280
1281         if (td->o.cpumask_set) {
1282                 int ret = fio_cpuset_exit(&td->o.cpumask);
1283
1284                 td_verror(td, ret, "fio_cpuset_exit");
1285         }
1286
1287         /*
1288          * do this very late, it will log file closing as well
1289          */
1290         if (td->o.write_iolog_file)
1291                 write_iolog_close(td);
1292
1293         options_mem_free(td);
1294         td_set_runstate(td, TD_EXITED);
1295         return (void *) (unsigned long) td->error;
1296 }
1297
1298 /*
1299  * We cannot pass the td data into a forked process, so attach the td and
1300  * pass it to the thread worker.
1301  */
1302 static int fork_main(int shmid, int offset)
1303 {
1304         struct thread_data *td;
1305         void *data, *ret;
1306
1307 #ifndef __hpux
1308         data = shmat(shmid, NULL, 0);
1309         if (data == (void *) -1) {
1310                 int __err = errno;
1311
1312                 perror("shmat");
1313                 return __err;
1314         }
1315 #else
1316         /*
1317          * HP-UX inherits shm mappings?
1318          */
1319         data = threads;
1320 #endif
1321
1322         td = data + offset * sizeof(struct thread_data);
1323         ret = thread_main(td);
1324         shmdt(data);
1325         return (int) (unsigned long) ret;
1326 }
1327
1328 /*
1329  * Run over the job map and reap the threads that have exited, if any.
1330  */
1331 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1332 {
1333         struct thread_data *td;
1334         int i, cputhreads, realthreads, pending, status, ret;
1335
1336         /*
1337          * reap exited threads (TD_EXITED -> TD_REAPED)
1338          */
1339         realthreads = pending = cputhreads = 0;
1340         for_each_td(td, i) {
1341                 int flags = 0;
1342
1343                 /*
1344                  * ->io_ops is NULL for a thread that has closed its
1345                  * io engine
1346                  */
1347                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1348                         cputhreads++;
1349                 else
1350                         realthreads++;
1351
1352                 if (!td->pid) {
1353                         pending++;
1354                         continue;
1355                 }
1356                 if (td->runstate == TD_REAPED)
1357                         continue;
1358                 if (td->o.use_thread) {
1359                         if (td->runstate == TD_EXITED) {
1360                                 td_set_runstate(td, TD_REAPED);
1361                                 goto reaped;
1362                         }
1363                         continue;
1364                 }
1365
1366                 flags = WNOHANG;
1367                 if (td->runstate == TD_EXITED)
1368                         flags = 0;
1369
1370                 /*
1371                  * check if someone quit or got killed in an unusual way
1372                  */
1373                 ret = waitpid(td->pid, &status, flags);
1374                 if (ret < 0) {
1375                         if (errno == ECHILD) {
1376                                 log_err("fio: pid=%d disappeared %d\n",
1377                                                 (int) td->pid, td->runstate);
1378                                 td_set_runstate(td, TD_REAPED);
1379                                 goto reaped;
1380                         }
1381                         perror("waitpid");
1382                 } else if (ret == td->pid) {
1383                         if (WIFSIGNALED(status)) {
1384                                 int sig = WTERMSIG(status);
1385
1386                                 if (sig != SIGTERM)
1387                                         log_err("fio: pid=%d, got signal=%d\n",
1388                                                         (int) td->pid, sig);
1389                                 td_set_runstate(td, TD_REAPED);
1390                                 goto reaped;
1391                         }
1392                         if (WIFEXITED(status)) {
1393                                 if (WEXITSTATUS(status) && !td->error)
1394                                         td->error = WEXITSTATUS(status);
1395
1396                                 td_set_runstate(td, TD_REAPED);
1397                                 goto reaped;
1398                         }
1399                 }
1400
1401                 /*
1402                  * thread is not dead, continue
1403                  */
1404                 pending++;
1405                 continue;
1406 reaped:
1407                 (*nr_running)--;
1408                 (*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
1409                 (*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
1410                 if (!td->pid)
1411                         pending--;
1412
1413                 if (td->error)
1414                         exit_value++;
1415
1416                 done_secs += mtime_since_now(&td->epoch) / 1000;
1417         }
1418
1419         if (*nr_running == cputhreads && !pending && realthreads)
1420                 terminate_threads(TERMINATE_ALL);
1421 }
1422
1423 static void *gtod_thread_main(void *data)
1424 {
1425         fio_mutex_up(startup_mutex);
1426
1427         /*
1428          * As long as we have jobs around, update the clock. It would be nice
1429          * to have some way of NOT hammering that CPU with gettimeofday(),
1430          * but I'm not sure what to use outside of a simple CPU nop to relax
1431          * it - we don't want to lose precision.
1432          */
1433         while (threads) {
1434                 fio_gtod_update();
1435                 nop;
1436         }
1437
1438         return NULL;
1439 }
1440
1441 static int fio_start_gtod_thread(void)
1442 {
1443         pthread_attr_t attr;
1444         int ret;
1445
1446         pthread_attr_init(&attr);
1447         pthread_attr_setstacksize(&attr, PTHREAD_STACK_MIN);
1448         ret = pthread_create(&gtod_thread, &attr, gtod_thread_main, NULL);
1449         pthread_attr_destroy(&attr);
1450         if (ret) {
1451                 log_err("Can't create gtod thread: %s\n", strerror(ret));
1452                 return 1;
1453         }
1454
1455         ret = pthread_detach(gtod_thread);
1456         if (ret) {
1457                 log_err("Can't detatch gtod thread: %s\n", strerror(ret));
1458                 return 1;
1459         }
1460
1461         dprint(FD_MUTEX, "wait on startup_mutex\n");
1462         fio_mutex_down(startup_mutex);
1463         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1464         return 0;
1465 }
1466
1467 /*
1468  * Main function for kicking off and reaping jobs, as needed.
1469  */
1470 static void run_threads(void)
1471 {
1472         struct thread_data *td;
1473         unsigned long spent;
1474         int i, todo, nr_running, m_rate, t_rate, nr_started;
1475
1476         if (fio_pin_memory())
1477                 return;
1478
1479         if (fio_gtod_offload && fio_start_gtod_thread())
1480                 return;
1481
1482         if (!terse_output) {
1483                 log_info("Starting ");
1484                 if (nr_thread)
1485                         log_info("%d thread%s", nr_thread,
1486                                                 nr_thread > 1 ? "s" : "");
1487                 if (nr_process) {
1488                         if (nr_thread)
1489                                 printf(" and ");
1490                         log_info("%d process%s", nr_process,
1491                                                 nr_process > 1 ? "es" : "");
1492                 }
1493                 log_info("\n");
1494                 fflush(stdout);
1495         }
1496
1497         set_sig_handlers();
1498
1499         todo = thread_number;
1500         nr_running = 0;
1501         nr_started = 0;
1502         m_rate = t_rate = 0;
1503
1504         for_each_td(td, i) {
1505                 print_status_init(td->thread_number - 1);
1506
1507                 if (!td->o.create_serialize)
1508                         continue;
1509
1510                 /*
1511                  * do file setup here so it happens sequentially,
1512                  * we don't want X number of threads getting their
1513                  * client data interspersed on disk
1514                  */
1515                 if (setup_files(td)) {
1516                         exit_value++;
1517                         if (td->error)
1518                                 log_err("fio: pid=%d, err=%d/%s\n",
1519                                         (int) td->pid, td->error, td->verror);
1520                         td_set_runstate(td, TD_REAPED);
1521                         todo--;
1522                 } else {
1523                         struct fio_file *f;
1524                         unsigned int j;
1525
1526                         /*
1527                          * for sharing to work, each job must always open
1528                          * its own files. so close them, if we opened them
1529                          * for creation
1530                          */
1531                         for_each_file(td, f, j) {
1532                                 if (fio_file_open(f))
1533                                         td_io_close_file(td, f);
1534                         }
1535                 }
1536         }
1537
1538         set_genesis_time();
1539
1540         while (todo) {
1541                 struct thread_data *map[REAL_MAX_JOBS];
1542                 struct timeval this_start;
1543                 int this_jobs = 0, left;
1544
1545                 /*
1546                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1547                  */
1548                 for_each_td(td, i) {
1549                         if (td->runstate != TD_NOT_CREATED)
1550                                 continue;
1551
1552                         /*
1553                          * never got a chance to start, killed by other
1554                          * thread for some reason
1555                          */
1556                         if (td->terminate) {
1557                                 todo--;
1558                                 continue;
1559                         }
1560
1561                         if (td->o.start_delay) {
1562                                 spent = mtime_since_genesis();
1563
1564                                 if (td->o.start_delay * 1000 > spent)
1565                                         continue;
1566                         }
1567
1568                         if (td->o.stonewall && (nr_started || nr_running)) {
1569                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1570                                                         td->o.name);
1571                                 break;
1572                         }
1573
1574                         init_disk_util(td);
1575
1576                         /*
1577                          * Set state to created. Thread will transition
1578                          * to TD_INITIALIZED when it's done setting up.
1579                          */
1580                         td_set_runstate(td, TD_CREATED);
1581                         map[this_jobs++] = td;
1582                         nr_started++;
1583
1584                         if (td->o.use_thread) {
1585                                 int ret;
1586
1587                                 dprint(FD_PROCESS, "will pthread_create\n");
1588                                 ret = pthread_create(&td->thread, NULL,
1589                                                         thread_main, td);
1590                                 if (ret) {
1591                                         log_err("pthread_create: %s\n",
1592                                                         strerror(ret));
1593                                         nr_started--;
1594                                         break;
1595                                 }
1596                                 ret = pthread_detach(td->thread);
1597                                 if (ret)
1598                                         log_err("pthread_detach: %s",
1599                                                         strerror(ret));
1600                         } else {
1601                                 pid_t pid;
1602                                 dprint(FD_PROCESS, "will fork\n");
1603                                 pid = fork();
1604                                 if (!pid) {
1605                                         int ret = fork_main(shm_id, i);
1606
1607                                         _exit(ret);
1608                                 } else if (i == fio_debug_jobno)
1609                                         *fio_debug_jobp = pid;
1610                         }
1611                         dprint(FD_MUTEX, "wait on startup_mutex\n");
1612                         if (fio_mutex_down_timeout(startup_mutex, 10)) {
1613                                 log_err("fio: job startup hung? exiting.\n");
1614                                 terminate_threads(TERMINATE_ALL);
1615                                 fio_abort = 1;
1616                                 nr_started--;
1617                                 break;
1618                         }
1619                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1620                 }
1621
1622                 /*
1623                  * Wait for the started threads to transition to
1624                  * TD_INITIALIZED.
1625                  */
1626                 fio_gettime(&this_start, NULL);
1627                 left = this_jobs;
1628                 while (left && !fio_abort) {
1629                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1630                                 break;
1631
1632                         usleep(100000);
1633
1634                         for (i = 0; i < this_jobs; i++) {
1635                                 td = map[i];
1636                                 if (!td)
1637                                         continue;
1638                                 if (td->runstate == TD_INITIALIZED) {
1639                                         map[i] = NULL;
1640                                         left--;
1641                                 } else if (td->runstate >= TD_EXITED) {
1642                                         map[i] = NULL;
1643                                         left--;
1644                                         todo--;
1645                                         nr_running++; /* work-around... */
1646                                 }
1647                         }
1648                 }
1649
1650                 if (left) {
1651                         log_err("fio: %d jobs failed to start\n", left);
1652                         for (i = 0; i < this_jobs; i++) {
1653                                 td = map[i];
1654                                 if (!td)
1655                                         continue;
1656                                 kill(td->pid, SIGTERM);
1657                         }
1658                         break;
1659                 }
1660
1661                 /*
1662                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1663                  */
1664                 for_each_td(td, i) {
1665                         if (td->runstate != TD_INITIALIZED)
1666                                 continue;
1667
1668                         if (in_ramp_time(td))
1669                                 td_set_runstate(td, TD_RAMP);
1670                         else
1671                                 td_set_runstate(td, TD_RUNNING);
1672                         nr_running++;
1673                         nr_started--;
1674                         m_rate += td->o.ratemin[0] + td->o.ratemin[1];
1675                         t_rate += td->o.rate[0] + td->o.rate[1];
1676                         todo--;
1677                         fio_mutex_up(td->mutex);
1678                 }
1679
1680                 reap_threads(&nr_running, &t_rate, &m_rate);
1681
1682                 if (todo)
1683                         usleep(100000);
1684         }
1685
1686         while (nr_running) {
1687                 reap_threads(&nr_running, &t_rate, &m_rate);
1688                 usleep(10000);
1689         }
1690
1691         update_io_ticks();
1692         fio_unpin_memory();
1693 }
1694
1695 int exec_run(void)
1696 {
1697         if (nr_clients)
1698                 return fio_handle_clients();
1699         if (exec_profile && load_profile(exec_profile))
1700                 return 1;
1701
1702         if (!thread_number)
1703                 return 0;
1704
1705         if (write_bw_log) {
1706                 setup_log(&agg_io_log[DDIR_READ]);
1707                 setup_log(&agg_io_log[DDIR_WRITE]);
1708         }
1709
1710         startup_mutex = fio_mutex_init(0);
1711         if (startup_mutex == NULL)
1712                 return 1;
1713         writeout_mutex = fio_mutex_init(1);
1714         if (writeout_mutex == NULL)
1715                 return 1;
1716
1717         set_genesis_time();
1718         create_disk_util_thread();
1719
1720         cgroup_list = smalloc(sizeof(*cgroup_list));
1721         INIT_FLIST_HEAD(cgroup_list);
1722
1723         run_threads();
1724
1725         if (!fio_abort) {
1726                 show_run_stats();
1727                 if (write_bw_log) {
1728                         __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1729                         __finish_log(agg_io_log[DDIR_WRITE],
1730                                         "agg-write_bw.log");
1731                 }
1732         }
1733
1734         cgroup_kill(cgroup_list);
1735         sfree(cgroup_list);
1736         sfree(cgroup_mnt);
1737
1738         fio_mutex_remove(startup_mutex);
1739         fio_mutex_remove(writeout_mutex);
1740         return exit_value;
1741 }
1742
1743 void reset_fio_state(void)
1744 {
1745         groupid = 0;
1746         thread_number = 0;
1747         nr_process = 0;
1748         nr_thread = 0;
1749         done_secs = 0;
1750 }
1751
1752 int main(int argc, char *argv[], char *envp[])
1753 {
1754         long ps;
1755
1756         arch_init(envp);
1757
1758         sinit();
1759
1760         /*
1761          * We need locale for number printing, if it isn't set then just
1762          * go with the US format.
1763          */
1764         if (!getenv("LC_NUMERIC"))
1765                 setlocale(LC_NUMERIC, "en_US");
1766
1767         ps = sysconf(_SC_PAGESIZE);
1768         if (ps < 0) {
1769                 log_err("Failed to get page size\n");
1770                 return 1;
1771         }
1772
1773         page_size = ps;
1774         page_mask = ps - 1;
1775
1776         fio_keywords_init();
1777
1778         if (parse_options(argc, argv))
1779                 return 1;
1780
1781         return exec_run();
1782 }