Document that write_iolog is unsafe for concurrent jobs
[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 <time.h>
32 #include <sys/stat.h>
33 #include <sys/wait.h>
34 #include <sys/ipc.h>
35 #include <sys/shm.h>
36 #include <sys/mman.h>
37
38 #include "fio.h"
39 #include "hash.h"
40 #include "smalloc.h"
41 #include "verify.h"
42 #include "trim.h"
43 #include "diskutil.h"
44 #include "cgroup.h"
45 #include "profile.h"
46 #include "lib/rand.h"
47
48 unsigned long page_mask;
49 unsigned long page_size;
50
51 #define PAGE_ALIGN(buf) \
52         (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
53
54 int groupid = 0;
55 int thread_number = 0;
56 int nr_process = 0;
57 int nr_thread = 0;
58 int shm_id = 0;
59 int temp_stall_ts;
60 unsigned long done_secs = 0;
61
62 static struct fio_mutex *startup_mutex;
63 static struct fio_mutex *writeout_mutex;
64 static volatile int fio_abort;
65 static int exit_value;
66 static timer_t ival_timer;
67 static pthread_t gtod_thread;
68 static struct flist_head *cgroup_list;
69 static char *cgroup_mnt;
70
71 struct io_log *agg_io_log[2];
72
73 #define TERMINATE_ALL           (-1)
74 #define JOB_START_TIMEOUT       (5 * 1000)
75
76 void td_set_runstate(struct thread_data *td, int runstate)
77 {
78         if (td->runstate == runstate)
79                 return;
80
81         dprint(FD_PROCESS, "pid=%d: runstate %d -> %d\n", (int) td->pid,
82                                                 td->runstate, runstate);
83         td->runstate = runstate;
84 }
85
86 static void terminate_threads(int group_id)
87 {
88         struct thread_data *td;
89         int i;
90
91         dprint(FD_PROCESS, "terminate group_id=%d\n", group_id);
92
93         for_each_td(td, i) {
94                 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
95                         dprint(FD_PROCESS, "setting terminate on %s/%d\n",
96                                                 td->o.name, (int) td->pid);
97                         td->terminate = 1;
98                         td->o.start_delay = 0;
99
100                         /*
101                          * if the thread is running, just let it exit
102                          */
103                         if (td->runstate < TD_RUNNING)
104                                 kill(td->pid, SIGQUIT);
105                         else {
106                                 struct ioengine_ops *ops = td->io_ops;
107
108                                 if (ops && (ops->flags & FIO_SIGQUIT))
109                                         kill(td->pid, SIGQUIT);
110                         }
111                 }
112         }
113 }
114
115 static void status_timer_arm(void)
116 {
117         struct itimerspec value;
118
119         value.it_value.tv_sec = 0;
120         value.it_value.tv_nsec = DISK_UTIL_MSEC * 1000000;
121         value.it_interval.tv_sec = 0;
122         value.it_interval.tv_nsec = DISK_UTIL_MSEC * 1000000;
123
124         timer_settime(ival_timer, 0, &value, NULL);
125 }
126
127 static void ival_fn(union sigval sig)
128 {
129         if (threads) {
130                 update_io_ticks();
131                 print_thread_status();
132         }
133 }
134
135 /*
136  * Happens on thread runs with ctrl-c, ignore our own SIGQUIT
137  */
138 static void sig_quit(int sig)
139 {
140 }
141
142 static void sig_int(int sig)
143 {
144         if (threads) {
145                 log_info("\nfio: terminating on signal %d\n", sig);
146                 fflush(stdout);
147                 terminate_threads(TERMINATE_ALL);
148         }
149 }
150
151 static void posix_timer_teardown(void)
152 {
153         timer_delete(ival_timer);
154 }
155
156 static void posix_timer_setup(void)
157 {
158         struct sigevent evt;
159
160         memset(&evt, 0, sizeof(evt));
161         evt.sigev_notify = SIGEV_THREAD;
162         evt.sigev_notify_function = ival_fn;
163
164         if (timer_create(CLOCK_MONOTONIC, &evt, &ival_timer) < 0)
165                 perror("timer_create");
166 }
167
168 static void set_sig_handlers(void)
169 {
170         struct sigaction act;
171
172         memset(&act, 0, sizeof(act));
173         act.sa_handler = sig_int;
174         act.sa_flags = SA_RESTART;
175         sigaction(SIGINT, &act, NULL);
176
177         memset(&act, 0, sizeof(act));
178         act.sa_handler = sig_quit;
179         act.sa_flags = SA_RESTART;
180         sigaction(SIGQUIT, &act, NULL);
181 }
182
183 /*
184  * Check if we are above the minimum rate given.
185  */
186 static int __check_min_rate(struct thread_data *td, struct timeval *now,
187                             enum td_ddir ddir)
188 {
189         unsigned long long bytes = 0;
190         unsigned long iops = 0;
191         unsigned long spent;
192         unsigned long rate;
193         unsigned int ratemin = 0;
194         unsigned int rate_iops = 0;
195         unsigned int rate_iops_min = 0;
196
197         assert(ddir_rw(ddir));
198
199         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
200                 return 0;
201
202         /*
203          * allow a 2 second settle period in the beginning
204          */
205         if (mtime_since(&td->start, now) < 2000)
206                 return 0;
207
208         iops += td->io_blocks[ddir];
209         bytes += td->this_io_bytes[ddir];
210         ratemin += td->o.ratemin[ddir];
211         rate_iops += td->o.rate_iops[ddir];
212         rate_iops_min += td->o.rate_iops_min[ddir];
213
214         /*
215          * if rate blocks is set, sample is running
216          */
217         if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
218                 spent = mtime_since(&td->lastrate[ddir], now);
219                 if (spent < td->o.ratecycle)
220                         return 0;
221
222                 if (td->o.rate[ddir]) {
223                         /*
224                          * check bandwidth specified rate
225                          */
226                         if (bytes < td->rate_bytes[ddir]) {
227                                 log_err("%s: min rate %u not met\n", td->o.name,
228                                                                 ratemin);
229                                 return 1;
230                         } else {
231                                 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
232                                 if (rate < ratemin ||
233                                     bytes < td->rate_bytes[ddir]) {
234                                         log_err("%s: min rate %u not met, got"
235                                                 " %luKB/sec\n", td->o.name,
236                                                         ratemin, rate);
237                                         return 1;
238                                 }
239                         }
240                 } else {
241                         /*
242                          * checks iops specified rate
243                          */
244                         if (iops < rate_iops) {
245                                 log_err("%s: min iops rate %u not met\n",
246                                                 td->o.name, rate_iops);
247                                 return 1;
248                         } else {
249                                 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
250                                 if (rate < rate_iops_min ||
251                                     iops < td->rate_blocks[ddir]) {
252                                         log_err("%s: min iops rate %u not met,"
253                                                 " got %lu\n", td->o.name,
254                                                         rate_iops_min, rate);
255                                 }
256                         }
257                 }
258         }
259
260         td->rate_bytes[ddir] = bytes;
261         td->rate_blocks[ddir] = iops;
262         memcpy(&td->lastrate[ddir], now, sizeof(*now));
263         return 0;
264 }
265
266 static int check_min_rate(struct thread_data *td, struct timeval *now,
267                           unsigned long *bytes_done)
268 {
269         int ret = 0;
270
271         if (bytes_done[0])
272                 ret |= __check_min_rate(td, now, 0);
273         if (bytes_done[1])
274                 ret |= __check_min_rate(td, now, 1);
275
276         return ret;
277 }
278
279 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
280 {
281         if (!td->o.timeout)
282                 return 0;
283         if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
284                 return 1;
285
286         return 0;
287 }
288
289 /*
290  * When job exits, we can cancel the in-flight IO if we are using async
291  * io. Attempt to do so.
292  */
293 static void cleanup_pending_aio(struct thread_data *td)
294 {
295         struct flist_head *entry, *n;
296         struct io_u *io_u;
297         int r;
298
299         /*
300          * get immediately available events, if any
301          */
302         r = io_u_queued_complete(td, 0, NULL);
303         if (r < 0)
304                 return;
305
306         /*
307          * now cancel remaining active events
308          */
309         if (td->io_ops->cancel) {
310                 flist_for_each_safe(entry, n, &td->io_u_busylist) {
311                         io_u = flist_entry(entry, struct io_u, list);
312
313                         /*
314                          * if the io_u isn't in flight, then that generally
315                          * means someone leaked an io_u. complain but fix
316                          * it up, so we don't stall here.
317                          */
318                         if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
319                                 log_err("fio: non-busy IO on busy list\n");
320                                 put_io_u(td, io_u);
321                         } else {
322                                 r = td->io_ops->cancel(td, io_u);
323                                 if (!r)
324                                         put_io_u(td, io_u);
325                         }
326                 }
327         }
328
329         if (td->cur_depth)
330                 r = io_u_queued_complete(td, td->cur_depth, NULL);
331 }
332
333 /*
334  * Helper to handle the final sync of a file. Works just like the normal
335  * io path, just does everything sync.
336  */
337 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
338 {
339         struct io_u *io_u = __get_io_u(td);
340         int ret;
341
342         if (!io_u)
343                 return 1;
344
345         io_u->ddir = DDIR_SYNC;
346         io_u->file = f;
347
348         if (td_io_prep(td, io_u)) {
349                 put_io_u(td, io_u);
350                 return 1;
351         }
352
353 requeue:
354         ret = td_io_queue(td, io_u);
355         if (ret < 0) {
356                 td_verror(td, io_u->error, "td_io_queue");
357                 put_io_u(td, io_u);
358                 return 1;
359         } else if (ret == FIO_Q_QUEUED) {
360                 if (io_u_queued_complete(td, 1, NULL) < 0)
361                         return 1;
362         } else if (ret == FIO_Q_COMPLETED) {
363                 if (io_u->error) {
364                         td_verror(td, io_u->error, "td_io_queue");
365                         return 1;
366                 }
367
368                 if (io_u_sync_complete(td, io_u, NULL) < 0)
369                         return 1;
370         } else if (ret == FIO_Q_BUSY) {
371                 if (td_io_commit(td))
372                         return 1;
373                 goto requeue;
374         }
375
376         return 0;
377 }
378
379 static inline void update_tv_cache(struct thread_data *td)
380 {
381         if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
382                 fio_gettime(&td->tv_cache, NULL);
383 }
384
385 static int break_on_this_error(struct thread_data *td, int *retptr)
386 {
387         int ret = *retptr;
388
389         if (ret < 0 || td->error) {
390                 int err;
391
392                 if (!td->o.continue_on_error)
393                         return 1;
394
395                 if (ret < 0)
396                         err = -ret;
397                 else
398                         err = td->error;
399
400                 if (td_non_fatal_error(err)) {
401                         /*
402                          * Continue with the I/Os in case of
403                          * a non fatal error.
404                          */
405                         update_error_count(td, err);
406                         td_clear_error(td);
407                         *retptr = 0;
408                         return 0;
409                 } else if (td->o.fill_device && err == ENOSPC) {
410                         /*
411                          * We expect to hit this error if
412                          * fill_device option is set.
413                          */
414                         td_clear_error(td);
415                         td->terminate = 1;
416                         return 1;
417                 } else {
418                         /*
419                          * Stop the I/O in case of a fatal
420                          * error.
421                          */
422                         update_error_count(td, err);
423                         return 1;
424                 }
425         }
426
427         return 0;
428 }
429
430 /*
431  * The main verify engine. Runs over the writes we previously submitted,
432  * reads the blocks back in, and checks the crc/md5 of the data.
433  */
434 static void do_verify(struct thread_data *td)
435 {
436         struct fio_file *f;
437         struct io_u *io_u;
438         int ret, min_events;
439         unsigned int i;
440
441         dprint(FD_VERIFY, "starting loop\n");
442
443         /*
444          * sync io first and invalidate cache, to make sure we really
445          * read from disk.
446          */
447         for_each_file(td, f, i) {
448                 if (!fio_file_open(f))
449                         continue;
450                 if (fio_io_sync(td, f))
451                         break;
452                 if (file_invalidate_cache(td, f))
453                         break;
454         }
455
456         if (td->error)
457                 return;
458
459         td_set_runstate(td, TD_VERIFYING);
460
461         io_u = NULL;
462         while (!td->terminate) {
463                 int ret2, full;
464
465                 update_tv_cache(td);
466
467                 if (runtime_exceeded(td, &td->tv_cache)) {
468                         td->terminate = 1;
469                         break;
470                 }
471
472                 io_u = __get_io_u(td);
473                 if (!io_u)
474                         break;
475
476                 if (get_next_verify(td, io_u)) {
477                         put_io_u(td, io_u);
478                         break;
479                 }
480
481                 if (td_io_prep(td, io_u)) {
482                         put_io_u(td, io_u);
483                         break;
484                 }
485
486                 if (td->o.verify_async)
487                         io_u->end_io = verify_io_u_async;
488                 else
489                         io_u->end_io = verify_io_u;
490
491                 ret = td_io_queue(td, io_u);
492                 switch (ret) {
493                 case FIO_Q_COMPLETED:
494                         if (io_u->error) {
495                                 ret = -io_u->error;
496                                 clear_io_u(td, io_u);
497                         } else if (io_u->resid) {
498                                 int bytes = io_u->xfer_buflen - io_u->resid;
499                                 struct fio_file *f = io_u->file;
500
501                                 /*
502                                  * zero read, fail
503                                  */
504                                 if (!bytes) {
505                                         td_verror(td, EIO, "full resid");
506                                         put_io_u(td, io_u);
507                                         break;
508                                 }
509
510                                 io_u->xfer_buflen = io_u->resid;
511                                 io_u->xfer_buf += bytes;
512                                 io_u->offset += bytes;
513
514                                 if (ddir_rw(io_u->ddir))
515                                         td->ts.short_io_u[io_u->ddir]++;
516
517                                 if (io_u->offset == f->real_file_size)
518                                         goto sync_done;
519
520                                 requeue_io_u(td, &io_u);
521                         } else {
522 sync_done:
523                                 ret = io_u_sync_complete(td, io_u, NULL);
524                                 if (ret < 0)
525                                         break;
526                         }
527                         continue;
528                 case FIO_Q_QUEUED:
529                         break;
530                 case FIO_Q_BUSY:
531                         requeue_io_u(td, &io_u);
532                         ret2 = td_io_commit(td);
533                         if (ret2 < 0)
534                                 ret = ret2;
535                         break;
536                 default:
537                         assert(ret < 0);
538                         td_verror(td, -ret, "td_io_queue");
539                         break;
540                 }
541
542                 if (break_on_this_error(td, &ret))
543                         break;
544
545                 /*
546                  * if we can queue more, do so. but check if there are
547                  * completed io_u's first.
548                  */
549                 full = queue_full(td) || ret == FIO_Q_BUSY;
550                 if (full || !td->o.iodepth_batch_complete) {
551                         min_events = min(td->o.iodepth_batch_complete,
552                                          td->cur_depth);
553                         if (full && !min_events)
554                                 min_events = 1;
555
556                         do {
557                                 /*
558                                  * Reap required number of io units, if any,
559                                  * and do the verification on them through
560                                  * the callback handler
561                                  */
562                                 if (io_u_queued_complete(td, min_events, NULL) < 0) {
563                                         ret = -1;
564                                         break;
565                                 }
566                         } while (full && (td->cur_depth > td->o.iodepth_low));
567                 }
568                 if (ret < 0)
569                         break;
570         }
571
572         if (!td->error) {
573                 min_events = td->cur_depth;
574
575                 if (min_events)
576                         ret = io_u_queued_complete(td, min_events, NULL);
577         } else
578                 cleanup_pending_aio(td);
579
580         td_set_runstate(td, TD_RUNNING);
581
582         dprint(FD_VERIFY, "exiting loop\n");
583 }
584
585 /*
586  * Main IO worker function. It retrieves io_u's to process and queues
587  * and reaps them, checking for rate and errors along the way.
588  */
589 static void do_io(struct thread_data *td)
590 {
591         unsigned int i;
592         int ret = 0;
593
594         if (in_ramp_time(td))
595                 td_set_runstate(td, TD_RAMP);
596         else
597                 td_set_runstate(td, TD_RUNNING);
598
599         while ( (td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
600                 (!flist_empty(&td->trim_list)) ||
601                 ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) ) {
602                 struct timeval comp_time;
603                 unsigned long bytes_done[2] = { 0, 0 };
604                 int min_evts = 0;
605                 struct io_u *io_u;
606                 int ret2, full;
607
608                 if (td->terminate)
609                         break;
610
611                 update_tv_cache(td);
612
613                 if (runtime_exceeded(td, &td->tv_cache)) {
614                         td->terminate = 1;
615                         break;
616                 }
617
618                 io_u = get_io_u(td);
619                 if (!io_u)
620                         break;
621
622                 /*
623                  * Add verification end_io handler, if asked to verify
624                  * a previously written file.
625                  */
626                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
627                     !td_rw(td)) {
628                         if (td->o.verify_async)
629                                 io_u->end_io = verify_io_u_async;
630                         else
631                                 io_u->end_io = verify_io_u;
632                         td_set_runstate(td, TD_VERIFYING);
633                 } else if (in_ramp_time(td))
634                         td_set_runstate(td, TD_RAMP);
635                 else
636                         td_set_runstate(td, TD_RUNNING);
637
638                 ret = td_io_queue(td, io_u);
639                 switch (ret) {
640                 case FIO_Q_COMPLETED:
641                         if (io_u->error) {
642                                 ret = -io_u->error;
643                                 clear_io_u(td, io_u);
644                         } else if (io_u->resid) {
645                                 int bytes = io_u->xfer_buflen - io_u->resid;
646                                 struct fio_file *f = io_u->file;
647
648                                 /*
649                                  * zero read, fail
650                                  */
651                                 if (!bytes) {
652                                         td_verror(td, EIO, "full resid");
653                                         put_io_u(td, io_u);
654                                         break;
655                                 }
656
657                                 io_u->xfer_buflen = io_u->resid;
658                                 io_u->xfer_buf += bytes;
659                                 io_u->offset += bytes;
660
661                                 if (ddir_rw(io_u->ddir))
662                                         td->ts.short_io_u[io_u->ddir]++;
663
664                                 if (io_u->offset == f->real_file_size)
665                                         goto sync_done;
666
667                                 requeue_io_u(td, &io_u);
668                         } else {
669 sync_done:
670                                 if (__should_check_rate(td, 0) ||
671                                     __should_check_rate(td, 1))
672                                         fio_gettime(&comp_time, NULL);
673
674                                 ret = io_u_sync_complete(td, io_u, bytes_done);
675                                 if (ret < 0)
676                                         break;
677                         }
678                         break;
679                 case FIO_Q_QUEUED:
680                         /*
681                          * if the engine doesn't have a commit hook,
682                          * the io_u is really queued. if it does have such
683                          * a hook, it has to call io_u_queued() itself.
684                          */
685                         if (td->io_ops->commit == NULL)
686                                 io_u_queued(td, io_u);
687                         break;
688                 case FIO_Q_BUSY:
689                         requeue_io_u(td, &io_u);
690                         ret2 = td_io_commit(td);
691                         if (ret2 < 0)
692                                 ret = ret2;
693                         break;
694                 default:
695                         assert(ret < 0);
696                         put_io_u(td, io_u);
697                         break;
698                 }
699
700                 if (break_on_this_error(td, &ret))
701                         break;
702
703                 /*
704                  * See if we need to complete some commands
705                  */
706                 full = queue_full(td) || ret == FIO_Q_BUSY;
707                 if (full || !td->o.iodepth_batch_complete) {
708                         min_evts = min(td->o.iodepth_batch_complete,
709                                         td->cur_depth);
710                         if (full && !min_evts)
711                                 min_evts = 1;
712
713                         if (__should_check_rate(td, 0) ||
714                             __should_check_rate(td, 1))
715                                 fio_gettime(&comp_time, NULL);
716
717                         do {
718                                 ret = io_u_queued_complete(td, min_evts, bytes_done);
719                                 if (ret < 0)
720                                         break;
721
722                         } while (full && (td->cur_depth > td->o.iodepth_low));
723                 }
724
725                 if (ret < 0)
726                         break;
727                 if (!(bytes_done[0] + bytes_done[1]))
728                         continue;
729
730                 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
731                         if (check_min_rate(td, &comp_time, bytes_done)) {
732                                 if (exitall_on_terminate)
733                                         terminate_threads(td->groupid);
734                                 td_verror(td, EIO, "check_min_rate");
735                                 break;
736                         }
737                 }
738
739                 if (td->o.thinktime) {
740                         unsigned long long b;
741
742                         b = td->io_blocks[0] + td->io_blocks[1];
743                         if (!(b % td->o.thinktime_blocks)) {
744                                 int left;
745
746                                 if (td->o.thinktime_spin)
747                                         usec_spin(td->o.thinktime_spin);
748
749                                 left = td->o.thinktime - td->o.thinktime_spin;
750                                 if (left)
751                                         usec_sleep(td, left);
752                         }
753                 }
754         }
755
756         if (td->trim_entries)
757                 printf("trim entries %ld\n", td->trim_entries);
758
759         if (td->o.fill_device && td->error == ENOSPC) {
760                 td->error = 0;
761                 td->terminate = 1;
762         }
763         if (!td->error) {
764                 struct fio_file *f;
765
766                 i = td->cur_depth;
767                 if (i)
768                         ret = io_u_queued_complete(td, i, NULL);
769
770                 if (should_fsync(td) && td->o.end_fsync) {
771                         td_set_runstate(td, TD_FSYNCING);
772
773                         for_each_file(td, f, i) {
774                                 if (!fio_file_open(f))
775                                         continue;
776                                 fio_io_sync(td, f);
777                         }
778                 }
779         } else
780                 cleanup_pending_aio(td);
781
782         /*
783          * stop job if we failed doing any IO
784          */
785         if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
786                 td->done = 1;
787 }
788
789 static void cleanup_io_u(struct thread_data *td)
790 {
791         struct flist_head *entry, *n;
792         struct io_u *io_u;
793
794         flist_for_each_safe(entry, n, &td->io_u_freelist) {
795                 io_u = flist_entry(entry, struct io_u, list);
796
797                 flist_del(&io_u->list);
798                 free(io_u);
799         }
800
801         free_io_mem(td);
802 }
803
804 static int init_io_u(struct thread_data *td)
805 {
806         struct io_u *io_u;
807         unsigned int max_bs;
808         int cl_align, i, max_units;
809         char *p;
810
811         max_units = td->o.iodepth;
812         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
813         td->orig_buffer_size = (unsigned long long) max_bs
814                                         * (unsigned long long) max_units;
815
816         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
817                 unsigned long bs;
818
819                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
820                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
821         }
822
823         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
824                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
825                 return 1;
826         }
827
828         if (allocate_io_mem(td))
829                 return 1;
830
831         if (td->o.odirect || td->o.mem_align ||
832             (td->io_ops->flags & FIO_RAWIO))
833                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
834         else
835                 p = td->orig_buffer;
836
837         cl_align = os_cache_line_size();
838
839         for (i = 0; i < max_units; i++) {
840                 void *ptr;
841
842                 if (td->terminate)
843                         return 1;
844
845                 if (posix_memalign(&ptr, cl_align, sizeof(*io_u))) {
846                         log_err("fio: posix_memalign=%s\n", strerror(errno));
847                         break;
848                 }
849
850                 io_u = ptr;
851                 memset(io_u, 0, sizeof(*io_u));
852                 INIT_FLIST_HEAD(&io_u->list);
853                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
854
855                 if (!(td->io_ops->flags & FIO_NOIO)) {
856                         io_u->buf = p + max_bs * i;
857                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
858
859                         if (td_write(td) && !td->o.refill_buffers)
860                                 io_u_fill_buffer(td, io_u, max_bs);
861                         else if (td_write(td) && td->o.verify_pattern_bytes) {
862                                 /*
863                                  * Fill the buffer with the pattern if we are
864                                  * going to be doing writes.
865                                  */
866                                 fill_pattern(td, io_u->buf, max_bs, io_u);
867                         }
868                 }
869
870                 io_u->index = i;
871                 io_u->flags = IO_U_F_FREE;
872                 flist_add(&io_u->list, &td->io_u_freelist);
873         }
874
875         return 0;
876 }
877
878 static int switch_ioscheduler(struct thread_data *td)
879 {
880         char tmp[256], tmp2[128];
881         FILE *f;
882         int ret;
883
884         if (td->io_ops->flags & FIO_DISKLESSIO)
885                 return 0;
886
887         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
888
889         f = fopen(tmp, "r+");
890         if (!f) {
891                 if (errno == ENOENT) {
892                         log_err("fio: os or kernel doesn't support IO scheduler"
893                                 " switching\n");
894                         return 0;
895                 }
896                 td_verror(td, errno, "fopen iosched");
897                 return 1;
898         }
899
900         /*
901          * Set io scheduler.
902          */
903         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
904         if (ferror(f) || ret != 1) {
905                 td_verror(td, errno, "fwrite");
906                 fclose(f);
907                 return 1;
908         }
909
910         rewind(f);
911
912         /*
913          * Read back and check that the selected scheduler is now the default.
914          */
915         ret = fread(tmp, 1, sizeof(tmp), f);
916         if (ferror(f) || ret < 0) {
917                 td_verror(td, errno, "fread");
918                 fclose(f);
919                 return 1;
920         }
921
922         sprintf(tmp2, "[%s]", td->o.ioscheduler);
923         if (!strstr(tmp, tmp2)) {
924                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
925                 td_verror(td, EINVAL, "iosched_switch");
926                 fclose(f);
927                 return 1;
928         }
929
930         fclose(f);
931         return 0;
932 }
933
934 static int keep_running(struct thread_data *td)
935 {
936         unsigned long long io_done;
937
938         if (td->done)
939                 return 0;
940         if (td->o.time_based)
941                 return 1;
942         if (td->o.loops) {
943                 td->o.loops--;
944                 return 1;
945         }
946
947         io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
948                         + td->io_skip_bytes;
949         if (io_done < td->o.size)
950                 return 1;
951
952         return 0;
953 }
954
955 static void reset_io_counters(struct thread_data *td)
956 {
957         td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
958         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
959         td->zone_bytes = 0;
960         td->rate_bytes[0] = td->rate_bytes[1] = 0;
961         td->rate_blocks[0] = td->rate_blocks[1] = 0;
962
963         td->last_was_sync = 0;
964
965         /*
966          * reset file done count if we are to start over
967          */
968         if (td->o.time_based || td->o.loops)
969                 td->nr_done_files = 0;
970
971         /*
972          * Set the same seed to get repeatable runs
973          */
974         td_fill_rand_seeds(td);
975 }
976
977 void reset_all_stats(struct thread_data *td)
978 {
979         struct timeval tv;
980         int i;
981
982         reset_io_counters(td);
983
984         for (i = 0; i < 2; i++) {
985                 td->io_bytes[i] = 0;
986                 td->io_blocks[i] = 0;
987                 td->io_issues[i] = 0;
988                 td->ts.total_io_u[i] = 0;
989         }
990         
991         fio_gettime(&tv, NULL);
992         td->ts.runtime[0] = 0;
993         td->ts.runtime[1] = 0;
994         memcpy(&td->epoch, &tv, sizeof(tv));
995         memcpy(&td->start, &tv, sizeof(tv));
996 }
997
998 static void clear_io_state(struct thread_data *td)
999 {
1000         struct fio_file *f;
1001         unsigned int i;
1002
1003         reset_io_counters(td);
1004
1005         close_files(td);
1006         for_each_file(td, f, i)
1007                 fio_file_clear_done(f);
1008 }
1009
1010 static int exec_string(const char *string)
1011 {
1012         int ret, newlen = strlen(string) + 1 + 8;
1013         char *str;
1014
1015         str = malloc(newlen);
1016         sprintf(str, "sh -c %s", string);
1017
1018         ret = system(str);
1019         if (ret == -1)
1020                 log_err("fio: exec of cmd <%s> failed\n", str);
1021
1022         free(str);
1023         return ret;
1024 }
1025
1026 /*
1027  * Entry point for the thread based jobs. The process based jobs end up
1028  * here as well, after a little setup.
1029  */
1030 static void *thread_main(void *data)
1031 {
1032         unsigned long long elapsed;
1033         struct thread_data *td = data;
1034         pthread_condattr_t attr;
1035         int clear_state;
1036
1037         if (!td->o.use_thread)
1038                 setsid();
1039
1040         td->pid = getpid();
1041
1042         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1043
1044         INIT_FLIST_HEAD(&td->io_u_freelist);
1045         INIT_FLIST_HEAD(&td->io_u_busylist);
1046         INIT_FLIST_HEAD(&td->io_u_requeues);
1047         INIT_FLIST_HEAD(&td->io_log_list);
1048         INIT_FLIST_HEAD(&td->io_hist_list);
1049         INIT_FLIST_HEAD(&td->verify_list);
1050         INIT_FLIST_HEAD(&td->trim_list);
1051         pthread_mutex_init(&td->io_u_lock, NULL);
1052         td->io_hist_tree = RB_ROOT;
1053
1054         pthread_condattr_init(&attr);
1055         pthread_cond_init(&td->verify_cond, &attr);
1056         pthread_cond_init(&td->free_cond, &attr);
1057
1058         td_set_runstate(td, TD_INITIALIZED);
1059         dprint(FD_MUTEX, "up startup_mutex\n");
1060         fio_mutex_up(startup_mutex);
1061         dprint(FD_MUTEX, "wait on td->mutex\n");
1062         fio_mutex_down(td->mutex);
1063         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1064
1065         /*
1066          * the ->mutex mutex is now no longer used, close it to avoid
1067          * eating a file descriptor
1068          */
1069         fio_mutex_remove(td->mutex);
1070
1071         if (td->o.uid != -1U && setuid(td->o.uid)) {
1072                 td_verror(td, errno, "setuid");
1073                 goto err;
1074         }
1075         if (td->o.gid != -1U && setgid(td->o.gid)) {
1076                 td_verror(td, errno, "setgid");
1077                 goto err;
1078         }
1079
1080         /*
1081          * May alter parameters that init_io_u() will use, so we need to
1082          * do this first.
1083          */
1084         if (init_iolog(td))
1085                 goto err;
1086
1087         if (init_io_u(td))
1088                 goto err;
1089
1090         if (td->o.verify_async && verify_async_init(td))
1091                 goto err;
1092
1093         if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1094                 td_verror(td, errno, "cpu_set_affinity");
1095                 goto err;
1096         }
1097
1098         /*
1099          * If we have a gettimeofday() thread, make sure we exclude that
1100          * thread from this job
1101          */
1102         if (td->o.gtod_cpu) {
1103                 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1104                 if (fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1105                         td_verror(td, errno, "cpu_set_affinity");
1106                         goto err;
1107                 }
1108         }
1109
1110         if (td->ioprio_set) {
1111                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1112                         td_verror(td, errno, "ioprio_set");
1113                         goto err;
1114                 }
1115         }
1116
1117         if (td->o.cgroup_weight && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1118                 goto err;
1119
1120         if (nice(td->o.nice) == -1) {
1121                 td_verror(td, errno, "nice");
1122                 goto err;
1123         }
1124
1125         if (td->o.ioscheduler && switch_ioscheduler(td))
1126                 goto err;
1127
1128         if (!td->o.create_serialize && setup_files(td))
1129                 goto err;
1130
1131         if (td_io_init(td))
1132                 goto err;
1133
1134         if (init_random_map(td))
1135                 goto err;
1136
1137         if (td->o.exec_prerun) {
1138                 if (exec_string(td->o.exec_prerun))
1139                         goto err;
1140         }
1141
1142         if (td->o.pre_read) {
1143                 if (pre_read_files(td) < 0)
1144                         goto err;
1145         }
1146
1147         fio_gettime(&td->epoch, NULL);
1148         getrusage(RUSAGE_SELF, &td->ts.ru_start);
1149
1150         clear_state = 0;
1151         while (keep_running(td)) {
1152                 fio_gettime(&td->start, NULL);
1153                 memcpy(&td->ts.stat_sample_time[0], &td->start,
1154                                 sizeof(td->start));
1155                 memcpy(&td->ts.stat_sample_time[1], &td->start,
1156                                 sizeof(td->start));
1157                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1158
1159                 if (td->o.ratemin[0] || td->o.ratemin[1])
1160                         memcpy(&td->lastrate, &td->ts.stat_sample_time,
1161                                                         sizeof(td->lastrate));
1162
1163                 if (clear_state)
1164                         clear_io_state(td);
1165
1166                 prune_io_piece_log(td);
1167
1168                 do_io(td);
1169
1170                 clear_state = 1;
1171
1172                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1173                         elapsed = utime_since_now(&td->start);
1174                         td->ts.runtime[DDIR_READ] += elapsed;
1175                 }
1176                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1177                         elapsed = utime_since_now(&td->start);
1178                         td->ts.runtime[DDIR_WRITE] += elapsed;
1179                 }
1180
1181                 if (td->error || td->terminate)
1182                         break;
1183
1184                 if (!td->o.do_verify ||
1185                     td->o.verify == VERIFY_NONE ||
1186                     (td->io_ops->flags & FIO_UNIDIR))
1187                         continue;
1188
1189                 clear_io_state(td);
1190
1191                 fio_gettime(&td->start, NULL);
1192
1193                 do_verify(td);
1194
1195                 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1196
1197                 if (td->error || td->terminate)
1198                         break;
1199         }
1200
1201         update_rusage_stat(td);
1202         td->ts.runtime[0] = (td->ts.runtime[0] + 999) / 1000;
1203         td->ts.runtime[1] = (td->ts.runtime[1] + 999) / 1000;
1204         td->ts.total_run_time = mtime_since_now(&td->epoch);
1205         td->ts.io_bytes[0] = td->io_bytes[0];
1206         td->ts.io_bytes[1] = td->io_bytes[1];
1207
1208         fio_mutex_down(writeout_mutex);
1209         if (td->ts.bw_log) {
1210                 if (td->o.bw_log_file) {
1211                         finish_log_named(td, td->ts.bw_log,
1212                                                 td->o.bw_log_file, "bw");
1213                 } else
1214                         finish_log(td, td->ts.bw_log, "bw");
1215         }
1216         if (td->ts.lat_log) {
1217                 if (td->o.lat_log_file) {
1218                         finish_log_named(td, td->ts.lat_log,
1219                                                 td->o.lat_log_file, "lat");
1220                 } else
1221                         finish_log(td, td->ts.lat_log, "lat");
1222         }
1223         if (td->ts.slat_log) {
1224                 if (td->o.lat_log_file) {
1225                         finish_log_named(td, td->ts.slat_log,
1226                                                 td->o.lat_log_file, "slat");
1227                 } else
1228                         finish_log(td, td->ts.slat_log, "slat");
1229         }
1230         if (td->ts.clat_log) {
1231                 if (td->o.lat_log_file) {
1232                         finish_log_named(td, td->ts.clat_log,
1233                                                 td->o.lat_log_file, "clat");
1234                 } else
1235                         finish_log(td, td->ts.clat_log, "clat");
1236         }
1237         fio_mutex_up(writeout_mutex);
1238         if (td->o.exec_postrun)
1239                 exec_string(td->o.exec_postrun);
1240
1241         if (exitall_on_terminate)
1242                 terminate_threads(td->groupid);
1243
1244 err:
1245         if (td->error)
1246                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1247                                                         td->verror);
1248
1249         if (td->o.verify_async)
1250                 verify_async_exit(td);
1251
1252         close_and_free_files(td);
1253         close_ioengine(td);
1254         cleanup_io_u(td);
1255         cgroup_shutdown(td, &cgroup_mnt);
1256
1257         if (td->o.cpumask_set) {
1258                 int ret = fio_cpuset_exit(&td->o.cpumask);
1259
1260                 td_verror(td, ret, "fio_cpuset_exit");
1261         }
1262
1263         /*
1264          * do this very late, it will log file closing as well
1265          */
1266         if (td->o.write_iolog_file)
1267                 write_iolog_close(td);
1268
1269         options_mem_free(td);
1270         td_set_runstate(td, TD_EXITED);
1271         return (void *) (unsigned long) td->error;
1272 }
1273
1274 /*
1275  * We cannot pass the td data into a forked process, so attach the td and
1276  * pass it to the thread worker.
1277  */
1278 static int fork_main(int shmid, int offset)
1279 {
1280         struct thread_data *td;
1281         void *data, *ret;
1282
1283         data = shmat(shmid, NULL, 0);
1284         if (data == (void *) -1) {
1285                 int __err = errno;
1286
1287                 perror("shmat");
1288                 return __err;
1289         }
1290
1291         td = data + offset * sizeof(struct thread_data);
1292         ret = thread_main(td);
1293         shmdt(data);
1294         return (int) (unsigned long) ret;
1295 }
1296
1297 /*
1298  * Run over the job map and reap the threads that have exited, if any.
1299  */
1300 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1301 {
1302         struct thread_data *td;
1303         int i, cputhreads, realthreads, pending, status, ret;
1304
1305         /*
1306          * reap exited threads (TD_EXITED -> TD_REAPED)
1307          */
1308         realthreads = pending = cputhreads = 0;
1309         for_each_td(td, i) {
1310                 int flags = 0;
1311
1312                 /*
1313                  * ->io_ops is NULL for a thread that has closed its
1314                  * io engine
1315                  */
1316                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1317                         cputhreads++;
1318                 else
1319                         realthreads++;
1320
1321                 if (!td->pid) {
1322                         pending++;
1323                         continue;
1324                 }
1325                 if (td->runstate == TD_REAPED)
1326                         continue;
1327                 if (td->o.use_thread) {
1328                         if (td->runstate == TD_EXITED) {
1329                                 td_set_runstate(td, TD_REAPED);
1330                                 goto reaped;
1331                         }
1332                         continue;
1333                 }
1334
1335                 flags = WNOHANG;
1336                 if (td->runstate == TD_EXITED)
1337                         flags = 0;
1338
1339                 /*
1340                  * check if someone quit or got killed in an unusual way
1341                  */
1342                 ret = waitpid(td->pid, &status, flags);
1343                 if (ret < 0) {
1344                         if (errno == ECHILD) {
1345                                 log_err("fio: pid=%d disappeared %d\n",
1346                                                 (int) td->pid, td->runstate);
1347                                 td_set_runstate(td, TD_REAPED);
1348                                 goto reaped;
1349                         }
1350                         perror("waitpid");
1351                 } else if (ret == td->pid) {
1352                         if (WIFSIGNALED(status)) {
1353                                 int sig = WTERMSIG(status);
1354
1355                                 if (sig != SIGQUIT)
1356                                         log_err("fio: pid=%d, got signal=%d\n",
1357                                                         (int) td->pid, sig);
1358                                 td_set_runstate(td, TD_REAPED);
1359                                 goto reaped;
1360                         }
1361                         if (WIFEXITED(status)) {
1362                                 if (WEXITSTATUS(status) && !td->error)
1363                                         td->error = WEXITSTATUS(status);
1364
1365                                 td_set_runstate(td, TD_REAPED);
1366                                 goto reaped;
1367                         }
1368                 }
1369
1370                 /*
1371                  * thread is not dead, continue
1372                  */
1373                 pending++;
1374                 continue;
1375 reaped:
1376                 (*nr_running)--;
1377                 (*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
1378                 (*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
1379                 if (!td->pid)
1380                         pending--;
1381
1382                 if (td->error)
1383                         exit_value++;
1384
1385                 done_secs += mtime_since_now(&td->epoch) / 1000;
1386         }
1387
1388         if (*nr_running == cputhreads && !pending && realthreads)
1389                 terminate_threads(TERMINATE_ALL);
1390 }
1391
1392 static void *gtod_thread_main(void *data)
1393 {
1394         fio_mutex_up(startup_mutex);
1395
1396         /*
1397          * As long as we have jobs around, update the clock. It would be nice
1398          * to have some way of NOT hammering that CPU with gettimeofday(),
1399          * but I'm not sure what to use outside of a simple CPU nop to relax
1400          * it - we don't want to lose precision.
1401          */
1402         while (threads) {
1403                 fio_gtod_update();
1404                 nop;
1405         }
1406
1407         return NULL;
1408 }
1409
1410 static int fio_start_gtod_thread(void)
1411 {
1412         pthread_attr_t attr;
1413         int ret;
1414
1415         pthread_attr_init(&attr);
1416         pthread_attr_setstacksize(&attr, PTHREAD_STACK_MIN);
1417         ret = pthread_create(&gtod_thread, &attr, gtod_thread_main, NULL);
1418         pthread_attr_destroy(&attr);
1419         if (ret) {
1420                 log_err("Can't create gtod thread: %s\n", strerror(ret));
1421                 return 1;
1422         }
1423
1424         ret = pthread_detach(gtod_thread);
1425         if (ret) {
1426                 log_err("Can't detatch gtod thread: %s\n", strerror(ret));
1427                 return 1;
1428         }
1429
1430         dprint(FD_MUTEX, "wait on startup_mutex\n");
1431         fio_mutex_down(startup_mutex);
1432         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1433         return 0;
1434 }
1435
1436 /*
1437  * Main function for kicking off and reaping jobs, as needed.
1438  */
1439 static void run_threads(void)
1440 {
1441         struct thread_data *td;
1442         unsigned long spent;
1443         int i, todo, nr_running, m_rate, t_rate, nr_started;
1444
1445         if (fio_pin_memory())
1446                 return;
1447
1448         if (fio_gtod_offload && fio_start_gtod_thread())
1449                 return;
1450
1451         if (!terse_output) {
1452                 log_info("Starting ");
1453                 if (nr_thread)
1454                         log_info("%d thread%s", nr_thread,
1455                                                 nr_thread > 1 ? "s" : "");
1456                 if (nr_process) {
1457                         if (nr_thread)
1458                                 printf(" and ");
1459                         log_info("%d process%s", nr_process,
1460                                                 nr_process > 1 ? "es" : "");
1461                 }
1462                 log_info("\n");
1463                 fflush(stdout);
1464         }
1465
1466         set_sig_handlers();
1467
1468         todo = thread_number;
1469         nr_running = 0;
1470         nr_started = 0;
1471         m_rate = t_rate = 0;
1472
1473         for_each_td(td, i) {
1474                 print_status_init(td->thread_number - 1);
1475
1476                 if (!td->o.create_serialize) {
1477                         init_disk_util(td);
1478                         continue;
1479                 }
1480
1481                 /*
1482                  * do file setup here so it happens sequentially,
1483                  * we don't want X number of threads getting their
1484                  * client data interspersed on disk
1485                  */
1486                 if (setup_files(td)) {
1487                         exit_value++;
1488                         if (td->error)
1489                                 log_err("fio: pid=%d, err=%d/%s\n",
1490                                         (int) td->pid, td->error, td->verror);
1491                         td_set_runstate(td, TD_REAPED);
1492                         todo--;
1493                 } else {
1494                         struct fio_file *f;
1495                         unsigned int i;
1496
1497                         /*
1498                          * for sharing to work, each job must always open
1499                          * its own files. so close them, if we opened them
1500                          * for creation
1501                          */
1502                         for_each_file(td, f, i) {
1503                                 if (fio_file_open(f))
1504                                         td_io_close_file(td, f);
1505                         }
1506                 }
1507
1508                 init_disk_util(td);
1509         }
1510
1511         set_genesis_time();
1512
1513         while (todo) {
1514                 struct thread_data *map[MAX_JOBS];
1515                 struct timeval this_start;
1516                 int this_jobs = 0, left;
1517
1518                 /*
1519                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1520                  */
1521                 for_each_td(td, i) {
1522                         if (td->runstate != TD_NOT_CREATED)
1523                                 continue;
1524
1525                         /*
1526                          * never got a chance to start, killed by other
1527                          * thread for some reason
1528                          */
1529                         if (td->terminate) {
1530                                 todo--;
1531                                 continue;
1532                         }
1533
1534                         if (td->o.start_delay) {
1535                                 spent = mtime_since_genesis();
1536
1537                                 if (td->o.start_delay * 1000 > spent)
1538                                         continue;
1539                         }
1540
1541                         if (td->o.stonewall && (nr_started || nr_running)) {
1542                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1543                                                         td->o.name);
1544                                 break;
1545                         }
1546
1547                         /*
1548                          * Set state to created. Thread will transition
1549                          * to TD_INITIALIZED when it's done setting up.
1550                          */
1551                         td_set_runstate(td, TD_CREATED);
1552                         map[this_jobs++] = td;
1553                         nr_started++;
1554
1555                         if (td->o.use_thread) {
1556                                 int ret;
1557
1558                                 dprint(FD_PROCESS, "will pthread_create\n");
1559                                 ret = pthread_create(&td->thread, NULL,
1560                                                         thread_main, td);
1561                                 if (ret) {
1562                                         log_err("pthread_create: %s\n",
1563                                                         strerror(ret));
1564                                         nr_started--;
1565                                         break;
1566                                 }
1567                                 ret = pthread_detach(td->thread);
1568                                 if (ret)
1569                                         log_err("pthread_detach: %s",
1570                                                         strerror(ret));
1571                         } else {
1572                                 pid_t pid;
1573                                 dprint(FD_PROCESS, "will fork\n");
1574                                 pid = fork();
1575                                 if (!pid) {
1576                                         int ret = fork_main(shm_id, i);
1577
1578                                         _exit(ret);
1579                                 } else if (i == fio_debug_jobno)
1580                                         *fio_debug_jobp = pid;
1581                         }
1582                         dprint(FD_MUTEX, "wait on startup_mutex\n");
1583                         if (fio_mutex_down_timeout(startup_mutex, 10)) {
1584                                 log_err("fio: job startup hung? exiting.\n");
1585                                 terminate_threads(TERMINATE_ALL);
1586                                 fio_abort = 1;
1587                                 nr_started--;
1588                                 break;
1589                         }
1590                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1591                 }
1592
1593                 /*
1594                  * Wait for the started threads to transition to
1595                  * TD_INITIALIZED.
1596                  */
1597                 fio_gettime(&this_start, NULL);
1598                 left = this_jobs;
1599                 while (left && !fio_abort) {
1600                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1601                                 break;
1602
1603                         usleep(100000);
1604
1605                         for (i = 0; i < this_jobs; i++) {
1606                                 td = map[i];
1607                                 if (!td)
1608                                         continue;
1609                                 if (td->runstate == TD_INITIALIZED) {
1610                                         map[i] = NULL;
1611                                         left--;
1612                                 } else if (td->runstate >= TD_EXITED) {
1613                                         map[i] = NULL;
1614                                         left--;
1615                                         todo--;
1616                                         nr_running++; /* work-around... */
1617                                 }
1618                         }
1619                 }
1620
1621                 if (left) {
1622                         log_err("fio: %d jobs failed to start\n", left);
1623                         for (i = 0; i < this_jobs; i++) {
1624                                 td = map[i];
1625                                 if (!td)
1626                                         continue;
1627                                 kill(td->pid, SIGTERM);
1628                         }
1629                         break;
1630                 }
1631
1632                 /*
1633                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1634                  */
1635                 for_each_td(td, i) {
1636                         if (td->runstate != TD_INITIALIZED)
1637                                 continue;
1638
1639                         if (in_ramp_time(td))
1640                                 td_set_runstate(td, TD_RAMP);
1641                         else
1642                                 td_set_runstate(td, TD_RUNNING);
1643                         nr_running++;
1644                         nr_started--;
1645                         m_rate += td->o.ratemin[0] + td->o.ratemin[1];
1646                         t_rate += td->o.rate[0] + td->o.rate[1];
1647                         todo--;
1648                         fio_mutex_up(td->mutex);
1649                 }
1650
1651                 reap_threads(&nr_running, &t_rate, &m_rate);
1652
1653                 if (todo)
1654                         usleep(100000);
1655         }
1656
1657         while (nr_running) {
1658                 reap_threads(&nr_running, &t_rate, &m_rate);
1659                 usleep(10000);
1660         }
1661
1662         update_io_ticks();
1663         fio_unpin_memory();
1664 }
1665
1666 int main(int argc, char *argv[])
1667 {
1668         long ps;
1669
1670         sinit();
1671         init_rand(&__fio_rand_state);
1672
1673         /*
1674          * We need locale for number printing, if it isn't set then just
1675          * go with the US format.
1676          */
1677         if (!getenv("LC_NUMERIC"))
1678                 setlocale(LC_NUMERIC, "en_US");
1679
1680         ps = sysconf(_SC_PAGESIZE);
1681         if (ps < 0) {
1682                 log_err("Failed to get page size\n");
1683                 return 1;
1684         }
1685
1686         page_size = ps;
1687         page_mask = ps - 1;
1688
1689         fio_keywords_init();
1690
1691         if (parse_options(argc, argv))
1692                 return 1;
1693
1694         if (exec_profile && load_profile(exec_profile))
1695                 return 1;
1696
1697         if (!thread_number)
1698                 return 0;
1699
1700         if (write_bw_log) {
1701                 setup_log(&agg_io_log[DDIR_READ]);
1702                 setup_log(&agg_io_log[DDIR_WRITE]);
1703         }
1704
1705         startup_mutex = fio_mutex_init(0);
1706         writeout_mutex = fio_mutex_init(1);
1707
1708         set_genesis_time();
1709
1710         posix_timer_setup();
1711         status_timer_arm();
1712
1713         cgroup_list = smalloc(sizeof(*cgroup_list));
1714         INIT_FLIST_HEAD(cgroup_list);
1715
1716         run_threads();
1717
1718         if (!fio_abort) {
1719                 show_run_stats();
1720                 if (write_bw_log) {
1721                         __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1722                         __finish_log(agg_io_log[DDIR_WRITE],
1723                                         "agg-write_bw.log");
1724                 }
1725         }
1726
1727         cgroup_kill(cgroup_list);
1728         sfree(cgroup_list);
1729         sfree(cgroup_mnt);
1730
1731         posix_timer_teardown();
1732         fio_mutex_remove(startup_mutex);
1733         fio_mutex_remove(writeout_mutex);
1734         return exit_value;
1735 }