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