When verify fails, dump the good/bad blocks to files
[fio.git] / fio.c
1 /*
2  * fio - the flexible io tester
3  *
4  * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
5  * Copyright (C) 2006 Jens Axboe <axboe@kernel.dk>
6  *
7  * The license below covers all files distributed with fio unless otherwise
8  * noted in the file itself.
9  *
10  *  This program is free software; you can redistribute it and/or modify
11  *  it under the terms of the GNU General Public License version 2 as
12  *  published by the Free Software Foundation.
13  *
14  *  This program is distributed in the hope that it will be useful,
15  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
16  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
17  *  GNU General Public License for more details.
18  *
19  *  You should have received a copy of the GNU General Public License
20  *  along with this program; if not, write to the Free Software
21  *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
22  *
23  */
24 #include <unistd.h>
25 #include <fcntl.h>
26 #include <string.h>
27 #include <limits.h>
28 #include <signal.h>
29 #include <time.h>
30 #include <locale.h>
31 #include <assert.h>
32 #include <time.h>
33 #include <sys/stat.h>
34 #include <sys/wait.h>
35 #include <sys/ipc.h>
36 #include <sys/shm.h>
37 #include <sys/mman.h>
38
39 #include "fio.h"
40 #include "hash.h"
41 #include "smalloc.h"
42 #include "verify.h"
43 #include "trim.h"
44 #include "diskutil.h"
45 #include "cgroup.h"
46 #include "profile.h"
47 #include "lib/rand.h"
48
49 unsigned long page_mask;
50 unsigned long page_size;
51
52 #define PAGE_ALIGN(buf) \
53         (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
54
55 int groupid = 0;
56 int thread_number = 0;
57 int nr_process = 0;
58 int nr_thread = 0;
59 int shm_id = 0;
60 int temp_stall_ts;
61 unsigned long done_secs = 0;
62
63 static struct fio_mutex *startup_mutex;
64 static struct fio_mutex *writeout_mutex;
65 static volatile int fio_abort;
66 static int exit_value;
67 static timer_t ival_timer;
68 static pthread_t gtod_thread;
69 static struct flist_head *cgroup_list;
70 static char *cgroup_mnt;
71
72 struct io_log *agg_io_log[2];
73
74 #define TERMINATE_ALL           (-1)
75 #define JOB_START_TIMEOUT       (5 * 1000)
76
77 void td_set_runstate(struct thread_data *td, int runstate)
78 {
79         if (td->runstate == runstate)
80                 return;
81
82         dprint(FD_PROCESS, "pid=%d: runstate %d -> %d\n", (int) td->pid,
83                                                 td->runstate, runstate);
84         td->runstate = runstate;
85 }
86
87 static void terminate_threads(int group_id)
88 {
89         struct thread_data *td;
90         int i;
91
92         dprint(FD_PROCESS, "terminate group_id=%d\n", group_id);
93
94         for_each_td(td, i) {
95                 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
96                         dprint(FD_PROCESS, "setting terminate on %s/%d\n",
97                                                 td->o.name, (int) td->pid);
98                         td->terminate = 1;
99                         td->o.start_delay = 0;
100
101                         /*
102                          * if the thread is running, just let it exit
103                          */
104                         if (td->runstate < TD_RUNNING)
105                                 kill(td->pid, SIGTERM);
106                         else {
107                                 struct ioengine_ops *ops = td->io_ops;
108
109                                 if (ops && (ops->flags & FIO_SIGTERM))
110                                         kill(td->pid, SIGTERM);
111                         }
112                 }
113         }
114 }
115
116 static void status_timer_arm(void)
117 {
118         struct itimerspec value;
119
120         value.it_value.tv_sec = 0;
121         value.it_value.tv_nsec = DISK_UTIL_MSEC * 1000000;
122         value.it_interval.tv_sec = 0;
123         value.it_interval.tv_nsec = DISK_UTIL_MSEC * 1000000;
124
125         timer_settime(ival_timer, 0, &value, NULL);
126 }
127
128 static void ival_fn(union sigval sig)
129 {
130         if (threads) {
131                 update_io_ticks();
132                 print_thread_status();
133         }
134 }
135
136 /*
137  * Happens on thread runs with ctrl-c, ignore our own SIGQUIT
138  */
139 static void sig_quit(int sig)
140 {
141 }
142
143 static void sig_int(int sig)
144 {
145         if (threads) {
146                 log_info("\nfio: terminating on signal %d\n", sig);
147                 fflush(stdout);
148                 terminate_threads(TERMINATE_ALL);
149         }
150 }
151
152 static void posix_timer_teardown(void)
153 {
154         timer_delete(ival_timer);
155 }
156
157 static void posix_timer_setup(void)
158 {
159         struct sigevent evt;
160
161         memset(&evt, 0, sizeof(evt));
162         evt.sigev_notify = SIGEV_THREAD;
163         evt.sigev_notify_function = ival_fn;
164
165         if (timer_create(FIO_TIMER_CLOCK, &evt, &ival_timer) < 0)
166                 perror("timer_create");
167
168 }
169
170 static void set_sig_handlers(void)
171 {
172         struct sigaction act;
173
174         memset(&act, 0, sizeof(act));
175         act.sa_handler = sig_int;
176         act.sa_flags = SA_RESTART;
177         sigaction(SIGINT, &act, NULL);
178
179         memset(&act, 0, sizeof(act));
180         act.sa_handler = sig_quit;
181         act.sa_flags = SA_RESTART;
182         sigaction(SIGTERM, &act, NULL);
183 }
184
185 /*
186  * Check if we are above the minimum rate given.
187  */
188 static int __check_min_rate(struct thread_data *td, struct timeval *now,
189                             enum td_ddir ddir)
190 {
191         unsigned long long bytes = 0;
192         unsigned long iops = 0;
193         unsigned long spent;
194         unsigned long rate;
195         unsigned int ratemin = 0;
196         unsigned int rate_iops = 0;
197         unsigned int rate_iops_min = 0;
198
199         assert(ddir_rw(ddir));
200
201         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
202                 return 0;
203
204         /*
205          * allow a 2 second settle period in the beginning
206          */
207         if (mtime_since(&td->start, now) < 2000)
208                 return 0;
209
210         iops += td->io_blocks[ddir];
211         bytes += td->this_io_bytes[ddir];
212         ratemin += td->o.ratemin[ddir];
213         rate_iops += td->o.rate_iops[ddir];
214         rate_iops_min += td->o.rate_iops_min[ddir];
215
216         /*
217          * if rate blocks is set, sample is running
218          */
219         if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
220                 spent = mtime_since(&td->lastrate[ddir], now);
221                 if (spent < td->o.ratecycle)
222                         return 0;
223
224                 if (td->o.rate[ddir]) {
225                         /*
226                          * check bandwidth specified rate
227                          */
228                         if (bytes < td->rate_bytes[ddir]) {
229                                 log_err("%s: min rate %u not met\n", td->o.name,
230                                                                 ratemin);
231                                 return 1;
232                         } else {
233                                 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
234                                 if (rate < ratemin ||
235                                     bytes < td->rate_bytes[ddir]) {
236                                         log_err("%s: min rate %u not met, got"
237                                                 " %luKB/sec\n", td->o.name,
238                                                         ratemin, rate);
239                                         return 1;
240                                 }
241                         }
242                 } else {
243                         /*
244                          * checks iops specified rate
245                          */
246                         if (iops < rate_iops) {
247                                 log_err("%s: min iops rate %u not met\n",
248                                                 td->o.name, rate_iops);
249                                 return 1;
250                         } else {
251                                 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
252                                 if (rate < rate_iops_min ||
253                                     iops < td->rate_blocks[ddir]) {
254                                         log_err("%s: min iops rate %u not met,"
255                                                 " got %lu\n", td->o.name,
256                                                         rate_iops_min, rate);
257                                 }
258                         }
259                 }
260         }
261
262         td->rate_bytes[ddir] = bytes;
263         td->rate_blocks[ddir] = iops;
264         memcpy(&td->lastrate[ddir], now, sizeof(*now));
265         return 0;
266 }
267
268 static int check_min_rate(struct thread_data *td, struct timeval *now,
269                           unsigned long *bytes_done)
270 {
271         int ret = 0;
272
273         if (bytes_done[0])
274                 ret |= __check_min_rate(td, now, 0);
275         if (bytes_done[1])
276                 ret |= __check_min_rate(td, now, 1);
277
278         return ret;
279 }
280
281 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
282 {
283         if (!td->o.timeout)
284                 return 0;
285         if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
286                 return 1;
287
288         return 0;
289 }
290
291 /*
292  * When job exits, we can cancel the in-flight IO if we are using async
293  * io. Attempt to do so.
294  */
295 static void cleanup_pending_aio(struct thread_data *td)
296 {
297         struct flist_head *entry, *n;
298         struct io_u *io_u;
299         int r;
300
301         /*
302          * get immediately available events, if any
303          */
304         r = io_u_queued_complete(td, 0, NULL);
305         if (r < 0)
306                 return;
307
308         /*
309          * now cancel remaining active events
310          */
311         if (td->io_ops->cancel) {
312                 flist_for_each_safe(entry, n, &td->io_u_busylist) {
313                         io_u = flist_entry(entry, struct io_u, list);
314
315                         /*
316                          * if the io_u isn't in flight, then that generally
317                          * means someone leaked an io_u. complain but fix
318                          * it up, so we don't stall here.
319                          */
320                         if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
321                                 log_err("fio: non-busy IO on busy list\n");
322                                 put_io_u(td, io_u);
323                         } else {
324                                 r = td->io_ops->cancel(td, io_u);
325                                 if (!r)
326                                         put_io_u(td, io_u);
327                         }
328                 }
329         }
330
331         if (td->cur_depth)
332                 r = io_u_queued_complete(td, td->cur_depth, NULL);
333 }
334
335 /*
336  * Helper to handle the final sync of a file. Works just like the normal
337  * io path, just does everything sync.
338  */
339 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
340 {
341         struct io_u *io_u = __get_io_u(td);
342         int ret;
343
344         if (!io_u)
345                 return 1;
346
347         io_u->ddir = DDIR_SYNC;
348         io_u->file = f;
349
350         if (td_io_prep(td, io_u)) {
351                 put_io_u(td, io_u);
352                 return 1;
353         }
354
355 requeue:
356         ret = td_io_queue(td, io_u);
357         if (ret < 0) {
358                 td_verror(td, io_u->error, "td_io_queue");
359                 put_io_u(td, io_u);
360                 return 1;
361         } else if (ret == FIO_Q_QUEUED) {
362                 if (io_u_queued_complete(td, 1, NULL) < 0)
363                         return 1;
364         } else if (ret == FIO_Q_COMPLETED) {
365                 if (io_u->error) {
366                         td_verror(td, io_u->error, "td_io_queue");
367                         return 1;
368                 }
369
370                 if (io_u_sync_complete(td, io_u, NULL) < 0)
371                         return 1;
372         } else if (ret == FIO_Q_BUSY) {
373                 if (td_io_commit(td))
374                         return 1;
375                 goto requeue;
376         }
377
378         return 0;
379 }
380
381 static inline void update_tv_cache(struct thread_data *td)
382 {
383         if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
384                 fio_gettime(&td->tv_cache, NULL);
385 }
386
387 static int break_on_this_error(struct thread_data *td, int *retptr)
388 {
389         int ret = *retptr;
390
391         if (ret < 0 || td->error) {
392                 int err;
393
394                 if (!td->o.continue_on_error)
395                         return 1;
396
397                 if (ret < 0)
398                         err = -ret;
399                 else
400                         err = td->error;
401
402                 if (td_non_fatal_error(err)) {
403                         /*
404                          * Continue with the I/Os in case of
405                          * a non fatal error.
406                          */
407                         update_error_count(td, err);
408                         td_clear_error(td);
409                         *retptr = 0;
410                         return 0;
411                 } else if (td->o.fill_device && err == ENOSPC) {
412                         /*
413                          * We expect to hit this error if
414                          * fill_device option is set.
415                          */
416                         td_clear_error(td);
417                         td->terminate = 1;
418                         return 1;
419                 } else {
420                         /*
421                          * Stop the I/O in case of a fatal
422                          * error.
423                          */
424                         update_error_count(td, err);
425                         return 1;
426                 }
427         }
428
429         return 0;
430 }
431
432 /*
433  * The main verify engine. Runs over the writes we previously submitted,
434  * reads the blocks back in, and checks the crc/md5 of the data.
435  */
436 static void do_verify(struct thread_data *td)
437 {
438         struct fio_file *f;
439         struct io_u *io_u;
440         int ret, min_events;
441         unsigned int i;
442
443         dprint(FD_VERIFY, "starting loop\n");
444
445         /*
446          * sync io first and invalidate cache, to make sure we really
447          * read from disk.
448          */
449         for_each_file(td, f, i) {
450                 if (!fio_file_open(f))
451                         continue;
452                 if (fio_io_sync(td, f))
453                         break;
454                 if (file_invalidate_cache(td, f))
455                         break;
456         }
457
458         if (td->error)
459                 return;
460
461         td_set_runstate(td, TD_VERIFYING);
462
463         io_u = NULL;
464         while (!td->terminate) {
465                 int ret2, full;
466
467                 update_tv_cache(td);
468
469                 if (runtime_exceeded(td, &td->tv_cache)) {
470                         td->terminate = 1;
471                         break;
472                 }
473
474                 io_u = __get_io_u(td);
475                 if (!io_u)
476                         break;
477
478                 if (get_next_verify(td, io_u)) {
479                         put_io_u(td, io_u);
480                         break;
481                 }
482
483                 if (td_io_prep(td, io_u)) {
484                         put_io_u(td, io_u);
485                         break;
486                 }
487
488                 if (td->o.verify_async)
489                         io_u->end_io = verify_io_u_async;
490                 else
491                         io_u->end_io = verify_io_u;
492
493                 ret = td_io_queue(td, io_u);
494                 switch (ret) {
495                 case FIO_Q_COMPLETED:
496                         if (io_u->error) {
497                                 ret = -io_u->error;
498                                 clear_io_u(td, io_u);
499                         } else if (io_u->resid) {
500                                 int bytes = io_u->xfer_buflen - io_u->resid;
501                                 struct fio_file *f = io_u->file;
502
503                                 /*
504                                  * zero read, fail
505                                  */
506                                 if (!bytes) {
507                                         td_verror(td, EIO, "full resid");
508                                         put_io_u(td, io_u);
509                                         break;
510                                 }
511
512                                 io_u->xfer_buflen = io_u->resid;
513                                 io_u->xfer_buf += bytes;
514                                 io_u->offset += bytes;
515
516                                 if (ddir_rw(io_u->ddir))
517                                         td->ts.short_io_u[io_u->ddir]++;
518
519                                 if (io_u->offset == f->real_file_size)
520                                         goto sync_done;
521
522                                 requeue_io_u(td, &io_u);
523                         } else {
524 sync_done:
525                                 ret = io_u_sync_complete(td, io_u, NULL);
526                                 if (ret < 0)
527                                         break;
528                         }
529                         continue;
530                 case FIO_Q_QUEUED:
531                         break;
532                 case FIO_Q_BUSY:
533                         requeue_io_u(td, &io_u);
534                         ret2 = td_io_commit(td);
535                         if (ret2 < 0)
536                                 ret = ret2;
537                         break;
538                 default:
539                         assert(ret < 0);
540                         td_verror(td, -ret, "td_io_queue");
541                         break;
542                 }
543
544                 if (break_on_this_error(td, &ret))
545                         break;
546
547                 /*
548                  * if we can queue more, do so. but check if there are
549                  * completed io_u's first.
550                  */
551                 full = queue_full(td) || ret == FIO_Q_BUSY;
552                 if (full || !td->o.iodepth_batch_complete) {
553                         min_events = min(td->o.iodepth_batch_complete,
554                                          td->cur_depth);
555                         if (full && !min_events)
556                                 min_events = 1;
557
558                         do {
559                                 /*
560                                  * Reap required number of io units, if any,
561                                  * and do the verification on them through
562                                  * the callback handler
563                                  */
564                                 if (io_u_queued_complete(td, min_events, NULL) < 0) {
565                                         ret = -1;
566                                         break;
567                                 }
568                         } while (full && (td->cur_depth > td->o.iodepth_low));
569                 }
570                 if (ret < 0)
571                         break;
572         }
573
574         if (!td->error) {
575                 min_events = td->cur_depth;
576
577                 if (min_events)
578                         ret = io_u_queued_complete(td, min_events, NULL);
579         } else
580                 cleanup_pending_aio(td);
581
582         td_set_runstate(td, TD_RUNNING);
583
584         dprint(FD_VERIFY, "exiting loop\n");
585 }
586
587 /*
588  * Main IO worker function. It retrieves io_u's to process and queues
589  * and reaps them, checking for rate and errors along the way.
590  */
591 static void do_io(struct thread_data *td)
592 {
593         unsigned int i;
594         int ret = 0;
595
596         if (in_ramp_time(td))
597                 td_set_runstate(td, TD_RAMP);
598         else
599                 td_set_runstate(td, TD_RUNNING);
600
601         while ( (td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
602                 (!flist_empty(&td->trim_list)) ||
603                 ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) ) {
604                 struct timeval comp_time;
605                 unsigned long bytes_done[2] = { 0, 0 };
606                 int min_evts = 0;
607                 struct io_u *io_u;
608                 int ret2, full;
609
610                 if (td->terminate)
611                         break;
612
613                 update_tv_cache(td);
614
615                 if (runtime_exceeded(td, &td->tv_cache)) {
616                         td->terminate = 1;
617                         break;
618                 }
619
620                 io_u = get_io_u(td);
621                 if (!io_u)
622                         break;
623
624                 /*
625                  * Add verification end_io handler, if asked to verify
626                  * a previously written file.
627                  */
628                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
629                     !td_rw(td)) {
630                         if (td->o.verify_async)
631                                 io_u->end_io = verify_io_u_async;
632                         else
633                                 io_u->end_io = verify_io_u;
634                         td_set_runstate(td, TD_VERIFYING);
635                 } else if (in_ramp_time(td))
636                         td_set_runstate(td, TD_RAMP);
637                 else
638                         td_set_runstate(td, TD_RUNNING);
639
640                 ret = td_io_queue(td, io_u);
641                 switch (ret) {
642                 case FIO_Q_COMPLETED:
643                         if (io_u->error) {
644                                 ret = -io_u->error;
645                                 clear_io_u(td, io_u);
646                         } else if (io_u->resid) {
647                                 int bytes = io_u->xfer_buflen - io_u->resid;
648                                 struct fio_file *f = io_u->file;
649
650                                 /*
651                                  * zero read, fail
652                                  */
653                                 if (!bytes) {
654                                         td_verror(td, EIO, "full resid");
655                                         put_io_u(td, io_u);
656                                         break;
657                                 }
658
659                                 io_u->xfer_buflen = io_u->resid;
660                                 io_u->xfer_buf += bytes;
661                                 io_u->offset += bytes;
662
663                                 if (ddir_rw(io_u->ddir))
664                                         td->ts.short_io_u[io_u->ddir]++;
665
666                                 if (io_u->offset == f->real_file_size)
667                                         goto sync_done;
668
669                                 requeue_io_u(td, &io_u);
670                         } else {
671 sync_done:
672                                 if (__should_check_rate(td, 0) ||
673                                     __should_check_rate(td, 1))
674                                         fio_gettime(&comp_time, NULL);
675
676                                 ret = io_u_sync_complete(td, io_u, bytes_done);
677                                 if (ret < 0)
678                                         break;
679                         }
680                         break;
681                 case FIO_Q_QUEUED:
682                         /*
683                          * if the engine doesn't have a commit hook,
684                          * the io_u is really queued. if it does have such
685                          * a hook, it has to call io_u_queued() itself.
686                          */
687                         if (td->io_ops->commit == NULL)
688                                 io_u_queued(td, io_u);
689                         break;
690                 case FIO_Q_BUSY:
691                         requeue_io_u(td, &io_u);
692                         ret2 = td_io_commit(td);
693                         if (ret2 < 0)
694                                 ret = ret2;
695                         break;
696                 default:
697                         assert(ret < 0);
698                         put_io_u(td, io_u);
699                         break;
700                 }
701
702                 if (break_on_this_error(td, &ret))
703                         break;
704
705                 /*
706                  * See if we need to complete some commands
707                  */
708                 full = queue_full(td) || ret == FIO_Q_BUSY;
709                 if (full || !td->o.iodepth_batch_complete) {
710                         min_evts = min(td->o.iodepth_batch_complete,
711                                         td->cur_depth);
712                         if (full && !min_evts)
713                                 min_evts = 1;
714
715                         if (__should_check_rate(td, 0) ||
716                             __should_check_rate(td, 1))
717                                 fio_gettime(&comp_time, NULL);
718
719                         do {
720                                 ret = io_u_queued_complete(td, min_evts, bytes_done);
721                                 if (ret < 0)
722                                         break;
723
724                         } while (full && (td->cur_depth > td->o.iodepth_low));
725                 }
726
727                 if (ret < 0)
728                         break;
729                 if (!(bytes_done[0] + bytes_done[1]))
730                         continue;
731
732                 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
733                         if (check_min_rate(td, &comp_time, bytes_done)) {
734                                 if (exitall_on_terminate)
735                                         terminate_threads(td->groupid);
736                                 td_verror(td, EIO, "check_min_rate");
737                                 break;
738                         }
739                 }
740
741                 if (td->o.thinktime) {
742                         unsigned long long b;
743
744                         b = td->io_blocks[0] + td->io_blocks[1];
745                         if (!(b % td->o.thinktime_blocks)) {
746                                 int left;
747
748                                 if (td->o.thinktime_spin)
749                                         usec_spin(td->o.thinktime_spin);
750
751                                 left = td->o.thinktime - td->o.thinktime_spin;
752                                 if (left)
753                                         usec_sleep(td, left);
754                         }
755                 }
756         }
757
758         if (td->trim_entries)
759                 printf("trim entries %ld\n", td->trim_entries);
760
761         if (td->o.fill_device && td->error == ENOSPC) {
762                 td->error = 0;
763                 td->terminate = 1;
764         }
765         if (!td->error) {
766                 struct fio_file *f;
767
768                 i = td->cur_depth;
769                 if (i)
770                         ret = io_u_queued_complete(td, i, NULL);
771
772                 if (should_fsync(td) && td->o.end_fsync) {
773                         td_set_runstate(td, TD_FSYNCING);
774
775                         for_each_file(td, f, i) {
776                                 if (!fio_file_open(f))
777                                         continue;
778                                 fio_io_sync(td, f);
779                         }
780                 }
781         } else
782                 cleanup_pending_aio(td);
783
784         /*
785          * stop job if we failed doing any IO
786          */
787         if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
788                 td->done = 1;
789 }
790
791 static void cleanup_io_u(struct thread_data *td)
792 {
793         struct flist_head *entry, *n;
794         struct io_u *io_u;
795
796         flist_for_each_safe(entry, n, &td->io_u_freelist) {
797                 io_u = flist_entry(entry, struct io_u, list);
798
799                 flist_del(&io_u->list);
800                 free(io_u);
801         }
802
803         free_io_mem(td);
804 }
805
806 static int init_io_u(struct thread_data *td)
807 {
808         struct io_u *io_u;
809         unsigned int max_bs;
810         int cl_align, i, max_units;
811         char *p;
812
813         max_units = td->o.iodepth;
814         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
815         td->orig_buffer_size = (unsigned long long) max_bs
816                                         * (unsigned long long) max_units;
817
818         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
819                 unsigned long bs;
820
821                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
822                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
823         }
824
825         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
826                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
827                 return 1;
828         }
829
830         if (allocate_io_mem(td))
831                 return 1;
832
833         if (td->o.odirect || td->o.mem_align ||
834             (td->io_ops->flags & FIO_RAWIO))
835                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
836         else
837                 p = td->orig_buffer;
838
839         cl_align = os_cache_line_size();
840
841         for (i = 0; i < max_units; i++) {
842                 void *ptr;
843
844                 if (td->terminate)
845                         return 1;
846
847                 if (posix_memalign(&ptr, cl_align, sizeof(*io_u))) {
848                         log_err("fio: posix_memalign=%s\n", strerror(errno));
849                         break;
850                 }
851
852                 io_u = ptr;
853                 memset(io_u, 0, sizeof(*io_u));
854                 INIT_FLIST_HEAD(&io_u->list);
855                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
856
857                 if (!(td->io_ops->flags & FIO_NOIO)) {
858                         io_u->buf = p + max_bs * i;
859                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
860
861                         if (td_write(td) && !td->o.refill_buffers)
862                                 io_u_fill_buffer(td, io_u, max_bs);
863                         else if (td_write(td) && td->o.verify_pattern_bytes) {
864                                 /*
865                                  * Fill the buffer with the pattern if we are
866                                  * going to be doing writes.
867                                  */
868                                 fill_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
869                         }
870                 }
871
872                 io_u->index = i;
873                 io_u->flags = IO_U_F_FREE;
874                 flist_add(&io_u->list, &td->io_u_freelist);
875         }
876
877         return 0;
878 }
879
880 static int switch_ioscheduler(struct thread_data *td)
881 {
882         char tmp[256], tmp2[128];
883         FILE *f;
884         int ret;
885
886         if (td->io_ops->flags & FIO_DISKLESSIO)
887                 return 0;
888
889         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
890
891         f = fopen(tmp, "r+");
892         if (!f) {
893                 if (errno == ENOENT) {
894                         log_err("fio: os or kernel doesn't support IO scheduler"
895                                 " switching\n");
896                         return 0;
897                 }
898                 td_verror(td, errno, "fopen iosched");
899                 return 1;
900         }
901
902         /*
903          * Set io scheduler.
904          */
905         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
906         if (ferror(f) || ret != 1) {
907                 td_verror(td, errno, "fwrite");
908                 fclose(f);
909                 return 1;
910         }
911
912         rewind(f);
913
914         /*
915          * Read back and check that the selected scheduler is now the default.
916          */
917         ret = fread(tmp, 1, sizeof(tmp), f);
918         if (ferror(f) || ret < 0) {
919                 td_verror(td, errno, "fread");
920                 fclose(f);
921                 return 1;
922         }
923
924         sprintf(tmp2, "[%s]", td->o.ioscheduler);
925         if (!strstr(tmp, tmp2)) {
926                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
927                 td_verror(td, EINVAL, "iosched_switch");
928                 fclose(f);
929                 return 1;
930         }
931
932         fclose(f);
933         return 0;
934 }
935
936 static int keep_running(struct thread_data *td)
937 {
938         unsigned long long io_done;
939
940         if (td->done)
941                 return 0;
942         if (td->o.time_based)
943                 return 1;
944         if (td->o.loops) {
945                 td->o.loops--;
946                 return 1;
947         }
948
949         io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
950                         + td->io_skip_bytes;
951         if (io_done < td->o.size)
952                 return 1;
953
954         return 0;
955 }
956
957 static void reset_io_counters(struct thread_data *td)
958 {
959         td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
960         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
961         td->zone_bytes = 0;
962         td->rate_bytes[0] = td->rate_bytes[1] = 0;
963         td->rate_blocks[0] = td->rate_blocks[1] = 0;
964
965         td->last_was_sync = 0;
966
967         /*
968          * reset file done count if we are to start over
969          */
970         if (td->o.time_based || td->o.loops)
971                 td->nr_done_files = 0;
972
973         /*
974          * Set the same seed to get repeatable runs
975          */
976         td_fill_rand_seeds(td);
977 }
978
979 void reset_all_stats(struct thread_data *td)
980 {
981         struct timeval tv;
982         int i;
983
984         reset_io_counters(td);
985
986         for (i = 0; i < 2; i++) {
987                 td->io_bytes[i] = 0;
988                 td->io_blocks[i] = 0;
989                 td->io_issues[i] = 0;
990                 td->ts.total_io_u[i] = 0;
991         }
992
993         fio_gettime(&tv, NULL);
994         td->ts.runtime[0] = 0;
995         td->ts.runtime[1] = 0;
996         memcpy(&td->epoch, &tv, sizeof(tv));
997         memcpy(&td->start, &tv, sizeof(tv));
998 }
999
1000 static void clear_io_state(struct thread_data *td)
1001 {
1002         struct fio_file *f;
1003         unsigned int i;
1004
1005         reset_io_counters(td);
1006
1007         close_files(td);
1008         for_each_file(td, f, i)
1009                 fio_file_clear_done(f);
1010 }
1011
1012 static int exec_string(const char *string)
1013 {
1014         int ret, newlen = strlen(string) + 1 + 8;
1015         char *str;
1016
1017         str = malloc(newlen);
1018         sprintf(str, "sh -c %s", string);
1019
1020         ret = system(str);
1021         if (ret == -1)
1022                 log_err("fio: exec of cmd <%s> failed\n", str);
1023
1024         free(str);
1025         return ret;
1026 }
1027
1028 /*
1029  * Entry point for the thread based jobs. The process based jobs end up
1030  * here as well, after a little setup.
1031  */
1032 static void *thread_main(void *data)
1033 {
1034         unsigned long long elapsed;
1035         struct thread_data *td = data;
1036         pthread_condattr_t attr;
1037         int clear_state;
1038
1039         if (!td->o.use_thread)
1040                 setsid();
1041
1042         td->pid = getpid();
1043
1044         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1045
1046         INIT_FLIST_HEAD(&td->io_u_freelist);
1047         INIT_FLIST_HEAD(&td->io_u_busylist);
1048         INIT_FLIST_HEAD(&td->io_u_requeues);
1049         INIT_FLIST_HEAD(&td->io_log_list);
1050         INIT_FLIST_HEAD(&td->io_hist_list);
1051         INIT_FLIST_HEAD(&td->verify_list);
1052         INIT_FLIST_HEAD(&td->trim_list);
1053         pthread_mutex_init(&td->io_u_lock, NULL);
1054         td->io_hist_tree = RB_ROOT;
1055
1056         pthread_condattr_init(&attr);
1057         pthread_cond_init(&td->verify_cond, &attr);
1058         pthread_cond_init(&td->free_cond, &attr);
1059
1060         td_set_runstate(td, TD_INITIALIZED);
1061         dprint(FD_MUTEX, "up startup_mutex\n");
1062         fio_mutex_up(startup_mutex);
1063         dprint(FD_MUTEX, "wait on td->mutex\n");
1064         fio_mutex_down(td->mutex);
1065         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1066
1067         /*
1068          * the ->mutex mutex is now no longer used, close it to avoid
1069          * eating a file descriptor
1070          */
1071         fio_mutex_remove(td->mutex);
1072
1073         /*
1074          * A new gid requires privilege, so we need to do this before setting
1075          * the uid.
1076          */
1077         if (td->o.gid != -1U && setgid(td->o.gid)) {
1078                 td_verror(td, errno, "setgid");
1079                 goto err;
1080         }
1081         if (td->o.uid != -1U && setuid(td->o.uid)) {
1082                 td_verror(td, errno, "setuid");
1083                 goto err;
1084         }
1085
1086         /*
1087          * May alter parameters that init_io_u() will use, so we need to
1088          * do this first.
1089          */
1090         if (init_iolog(td))
1091                 goto err;
1092
1093         if (init_io_u(td))
1094                 goto err;
1095
1096         if (td->o.verify_async && verify_async_init(td))
1097                 goto err;
1098
1099         if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1100                 td_verror(td, errno, "cpu_set_affinity");
1101                 goto err;
1102         }
1103
1104         /*
1105          * If we have a gettimeofday() thread, make sure we exclude that
1106          * thread from this job
1107          */
1108         if (td->o.gtod_cpu) {
1109                 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1110                 if (fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1111                         td_verror(td, errno, "cpu_set_affinity");
1112                         goto err;
1113                 }
1114         }
1115
1116         if (td->ioprio_set) {
1117                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1118                         td_verror(td, errno, "ioprio_set");
1119                         goto err;
1120                 }
1121         }
1122
1123         if (td->o.cgroup_weight && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1124                 goto err;
1125
1126         if (nice(td->o.nice) == -1) {
1127                 td_verror(td, errno, "nice");
1128                 goto err;
1129         }
1130
1131         if (td->o.ioscheduler && switch_ioscheduler(td))
1132                 goto err;
1133
1134         if (!td->o.create_serialize && setup_files(td))
1135                 goto err;
1136
1137         if (td_io_init(td))
1138                 goto err;
1139
1140         if (init_random_map(td))
1141                 goto err;
1142
1143         if (td->o.exec_prerun) {
1144                 if (exec_string(td->o.exec_prerun))
1145                         goto err;
1146         }
1147
1148         if (td->o.pre_read) {
1149                 if (pre_read_files(td) < 0)
1150                         goto err;
1151         }
1152
1153         fio_gettime(&td->epoch, NULL);
1154         getrusage(RUSAGE_SELF, &td->ts.ru_start);
1155
1156         clear_state = 0;
1157         while (keep_running(td)) {
1158                 fio_gettime(&td->start, NULL);
1159                 memcpy(&td->ts.stat_sample_time[0], &td->start,
1160                                 sizeof(td->start));
1161                 memcpy(&td->ts.stat_sample_time[1], &td->start,
1162                                 sizeof(td->start));
1163                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1164
1165                 if (td->o.ratemin[0] || td->o.ratemin[1])
1166                         memcpy(&td->lastrate, &td->ts.stat_sample_time,
1167                                                         sizeof(td->lastrate));
1168
1169                 if (clear_state)
1170                         clear_io_state(td);
1171
1172                 prune_io_piece_log(td);
1173
1174                 do_io(td);
1175
1176                 clear_state = 1;
1177
1178                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1179                         elapsed = utime_since_now(&td->start);
1180                         td->ts.runtime[DDIR_READ] += elapsed;
1181                 }
1182                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1183                         elapsed = utime_since_now(&td->start);
1184                         td->ts.runtime[DDIR_WRITE] += elapsed;
1185                 }
1186
1187                 if (td->error || td->terminate)
1188                         break;
1189
1190                 if (!td->o.do_verify ||
1191                     td->o.verify == VERIFY_NONE ||
1192                     (td->io_ops->flags & FIO_UNIDIR))
1193                         continue;
1194
1195                 clear_io_state(td);
1196
1197                 fio_gettime(&td->start, NULL);
1198
1199                 do_verify(td);
1200
1201                 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1202
1203                 if (td->error || td->terminate)
1204                         break;
1205         }
1206
1207         update_rusage_stat(td);
1208         td->ts.runtime[0] = (td->ts.runtime[0] + 999) / 1000;
1209         td->ts.runtime[1] = (td->ts.runtime[1] + 999) / 1000;
1210         td->ts.total_run_time = mtime_since_now(&td->epoch);
1211         td->ts.io_bytes[0] = td->io_bytes[0];
1212         td->ts.io_bytes[1] = td->io_bytes[1];
1213
1214         fio_mutex_down(writeout_mutex);
1215         if (td->ts.bw_log) {
1216                 if (td->o.bw_log_file) {
1217                         finish_log_named(td, td->ts.bw_log,
1218                                                 td->o.bw_log_file, "bw");
1219                 } else
1220                         finish_log(td, td->ts.bw_log, "bw");
1221         }
1222         if (td->ts.lat_log) {
1223                 if (td->o.lat_log_file) {
1224                         finish_log_named(td, td->ts.lat_log,
1225                                                 td->o.lat_log_file, "lat");
1226                 } else
1227                         finish_log(td, td->ts.lat_log, "lat");
1228         }
1229         if (td->ts.slat_log) {
1230                 if (td->o.lat_log_file) {
1231                         finish_log_named(td, td->ts.slat_log,
1232                                                 td->o.lat_log_file, "slat");
1233                 } else
1234                         finish_log(td, td->ts.slat_log, "slat");
1235         }
1236         if (td->ts.clat_log) {
1237                 if (td->o.lat_log_file) {
1238                         finish_log_named(td, td->ts.clat_log,
1239                                                 td->o.lat_log_file, "clat");
1240                 } else
1241                         finish_log(td, td->ts.clat_log, "clat");
1242         }
1243         fio_mutex_up(writeout_mutex);
1244         if (td->o.exec_postrun)
1245                 exec_string(td->o.exec_postrun);
1246
1247         if (exitall_on_terminate)
1248                 terminate_threads(td->groupid);
1249
1250 err:
1251         if (td->error)
1252                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1253                                                         td->verror);
1254
1255         if (td->o.verify_async)
1256                 verify_async_exit(td);
1257
1258         close_and_free_files(td);
1259         close_ioengine(td);
1260         cleanup_io_u(td);
1261         cgroup_shutdown(td, &cgroup_mnt);
1262
1263         if (td->o.cpumask_set) {
1264                 int ret = fio_cpuset_exit(&td->o.cpumask);
1265
1266                 td_verror(td, ret, "fio_cpuset_exit");
1267         }
1268
1269         /*
1270          * do this very late, it will log file closing as well
1271          */
1272         if (td->o.write_iolog_file)
1273                 write_iolog_close(td);
1274
1275         options_mem_free(td);
1276         td_set_runstate(td, TD_EXITED);
1277         return (void *) (unsigned long) td->error;
1278 }
1279
1280 /*
1281  * We cannot pass the td data into a forked process, so attach the td and
1282  * pass it to the thread worker.
1283  */
1284 static int fork_main(int shmid, int offset)
1285 {
1286         struct thread_data *td;
1287         void *data, *ret;
1288
1289         data = shmat(shmid, NULL, 0);
1290         if (data == (void *) -1) {
1291                 int __err = errno;
1292
1293                 perror("shmat");
1294                 return __err;
1295         }
1296
1297         td = data + offset * sizeof(struct thread_data);
1298         ret = thread_main(td);
1299         shmdt(data);
1300         return (int) (unsigned long) ret;
1301 }
1302
1303 /*
1304  * Run over the job map and reap the threads that have exited, if any.
1305  */
1306 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1307 {
1308         struct thread_data *td;
1309         int i, cputhreads, realthreads, pending, status, ret;
1310
1311         /*
1312          * reap exited threads (TD_EXITED -> TD_REAPED)
1313          */
1314         realthreads = pending = cputhreads = 0;
1315         for_each_td(td, i) {
1316                 int flags = 0;
1317
1318                 /*
1319                  * ->io_ops is NULL for a thread that has closed its
1320                  * io engine
1321                  */
1322                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1323                         cputhreads++;
1324                 else
1325                         realthreads++;
1326
1327                 if (!td->pid) {
1328                         pending++;
1329                         continue;
1330                 }
1331                 if (td->runstate == TD_REAPED)
1332                         continue;
1333                 if (td->o.use_thread) {
1334                         if (td->runstate == TD_EXITED) {
1335                                 td_set_runstate(td, TD_REAPED);
1336                                 goto reaped;
1337                         }
1338                         continue;
1339                 }
1340
1341                 flags = WNOHANG;
1342                 if (td->runstate == TD_EXITED)
1343                         flags = 0;
1344
1345                 /*
1346                  * check if someone quit or got killed in an unusual way
1347                  */
1348                 ret = waitpid(td->pid, &status, flags);
1349                 if (ret < 0) {
1350                         if (errno == ECHILD) {
1351                                 log_err("fio: pid=%d disappeared %d\n",
1352                                                 (int) td->pid, td->runstate);
1353                                 td_set_runstate(td, TD_REAPED);
1354                                 goto reaped;
1355                         }
1356                         perror("waitpid");
1357                 } else if (ret == td->pid) {
1358                         if (WIFSIGNALED(status)) {
1359                                 int sig = WTERMSIG(status);
1360
1361                                 if (sig != SIGTERM)
1362                                         log_err("fio: pid=%d, got signal=%d\n",
1363                                                         (int) td->pid, sig);
1364                                 td_set_runstate(td, TD_REAPED);
1365                                 goto reaped;
1366                         }
1367                         if (WIFEXITED(status)) {
1368                                 if (WEXITSTATUS(status) && !td->error)
1369                                         td->error = WEXITSTATUS(status);
1370
1371                                 td_set_runstate(td, TD_REAPED);
1372                                 goto reaped;
1373                         }
1374                 }
1375
1376                 /*
1377                  * thread is not dead, continue
1378                  */
1379                 pending++;
1380                 continue;
1381 reaped:
1382                 (*nr_running)--;
1383                 (*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
1384                 (*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
1385                 if (!td->pid)
1386                         pending--;
1387
1388                 if (td->error)
1389                         exit_value++;
1390
1391                 done_secs += mtime_since_now(&td->epoch) / 1000;
1392         }
1393
1394         if (*nr_running == cputhreads && !pending && realthreads)
1395                 terminate_threads(TERMINATE_ALL);
1396 }
1397
1398 static void *gtod_thread_main(void *data)
1399 {
1400         fio_mutex_up(startup_mutex);
1401
1402         /*
1403          * As long as we have jobs around, update the clock. It would be nice
1404          * to have some way of NOT hammering that CPU with gettimeofday(),
1405          * but I'm not sure what to use outside of a simple CPU nop to relax
1406          * it - we don't want to lose precision.
1407          */
1408         while (threads) {
1409                 fio_gtod_update();
1410                 nop;
1411         }
1412
1413         return NULL;
1414 }
1415
1416 static int fio_start_gtod_thread(void)
1417 {
1418         pthread_attr_t attr;
1419         int ret;
1420
1421         pthread_attr_init(&attr);
1422         pthread_attr_setstacksize(&attr, PTHREAD_STACK_MIN);
1423         ret = pthread_create(&gtod_thread, &attr, gtod_thread_main, NULL);
1424         pthread_attr_destroy(&attr);
1425         if (ret) {
1426                 log_err("Can't create gtod thread: %s\n", strerror(ret));
1427                 return 1;
1428         }
1429
1430         ret = pthread_detach(gtod_thread);
1431         if (ret) {
1432                 log_err("Can't detatch gtod thread: %s\n", strerror(ret));
1433                 return 1;
1434         }
1435
1436         dprint(FD_MUTEX, "wait on startup_mutex\n");
1437         fio_mutex_down(startup_mutex);
1438         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1439         return 0;
1440 }
1441
1442 /*
1443  * Main function for kicking off and reaping jobs, as needed.
1444  */
1445 static void run_threads(void)
1446 {
1447         struct thread_data *td;
1448         unsigned long spent;
1449         int i, todo, nr_running, m_rate, t_rate, nr_started;
1450
1451         if (fio_pin_memory())
1452                 return;
1453
1454         if (fio_gtod_offload && fio_start_gtod_thread())
1455                 return;
1456
1457         if (!terse_output) {
1458                 log_info("Starting ");
1459                 if (nr_thread)
1460                         log_info("%d thread%s", nr_thread,
1461                                                 nr_thread > 1 ? "s" : "");
1462                 if (nr_process) {
1463                         if (nr_thread)
1464                                 printf(" and ");
1465                         log_info("%d process%s", nr_process,
1466                                                 nr_process > 1 ? "es" : "");
1467                 }
1468                 log_info("\n");
1469                 fflush(stdout);
1470         }
1471
1472         set_sig_handlers();
1473
1474         todo = thread_number;
1475         nr_running = 0;
1476         nr_started = 0;
1477         m_rate = t_rate = 0;
1478
1479         for_each_td(td, i) {
1480                 print_status_init(td->thread_number - 1);
1481
1482                 if (!td->o.create_serialize) {
1483                         init_disk_util(td);
1484                         continue;
1485                 }
1486
1487                 /*
1488                  * do file setup here so it happens sequentially,
1489                  * we don't want X number of threads getting their
1490                  * client data interspersed on disk
1491                  */
1492                 if (setup_files(td)) {
1493                         exit_value++;
1494                         if (td->error)
1495                                 log_err("fio: pid=%d, err=%d/%s\n",
1496                                         (int) td->pid, td->error, td->verror);
1497                         td_set_runstate(td, TD_REAPED);
1498                         todo--;
1499                 } else {
1500                         struct fio_file *f;
1501                         unsigned int i;
1502
1503                         /*
1504                          * for sharing to work, each job must always open
1505                          * its own files. so close them, if we opened them
1506                          * for creation
1507                          */
1508                         for_each_file(td, f, i) {
1509                                 if (fio_file_open(f))
1510                                         td_io_close_file(td, f);
1511                         }
1512                 }
1513
1514                 init_disk_util(td);
1515         }
1516
1517         set_genesis_time();
1518
1519         while (todo) {
1520                 struct thread_data *map[MAX_JOBS];
1521                 struct timeval this_start;
1522                 int this_jobs = 0, left;
1523
1524                 /*
1525                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1526                  */
1527                 for_each_td(td, i) {
1528                         if (td->runstate != TD_NOT_CREATED)
1529                                 continue;
1530
1531                         /*
1532                          * never got a chance to start, killed by other
1533                          * thread for some reason
1534                          */
1535                         if (td->terminate) {
1536                                 todo--;
1537                                 continue;
1538                         }
1539
1540                         if (td->o.start_delay) {
1541                                 spent = mtime_since_genesis();
1542
1543                                 if (td->o.start_delay * 1000 > spent)
1544                                         continue;
1545                         }
1546
1547                         if (td->o.stonewall && (nr_started || nr_running)) {
1548                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1549                                                         td->o.name);
1550                                 break;
1551                         }
1552
1553                         /*
1554                          * Set state to created. Thread will transition
1555                          * to TD_INITIALIZED when it's done setting up.
1556                          */
1557                         td_set_runstate(td, TD_CREATED);
1558                         map[this_jobs++] = td;
1559                         nr_started++;
1560
1561                         if (td->o.use_thread) {
1562                                 int ret;
1563
1564                                 dprint(FD_PROCESS, "will pthread_create\n");
1565                                 ret = pthread_create(&td->thread, NULL,
1566                                                         thread_main, td);
1567                                 if (ret) {
1568                                         log_err("pthread_create: %s\n",
1569                                                         strerror(ret));
1570                                         nr_started--;
1571                                         break;
1572                                 }
1573                                 ret = pthread_detach(td->thread);
1574                                 if (ret)
1575                                         log_err("pthread_detach: %s",
1576                                                         strerror(ret));
1577                         } else {
1578                                 pid_t pid;
1579                                 dprint(FD_PROCESS, "will fork\n");
1580                                 pid = fork();
1581                                 if (!pid) {
1582                                         int ret = fork_main(shm_id, i);
1583
1584                                         _exit(ret);
1585                                 } else if (i == fio_debug_jobno)
1586                                         *fio_debug_jobp = pid;
1587                         }
1588                         dprint(FD_MUTEX, "wait on startup_mutex\n");
1589                         if (fio_mutex_down_timeout(startup_mutex, 10)) {
1590                                 log_err("fio: job startup hung? exiting.\n");
1591                                 terminate_threads(TERMINATE_ALL);
1592                                 fio_abort = 1;
1593                                 nr_started--;
1594                                 break;
1595                         }
1596                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1597                 }
1598
1599                 /*
1600                  * Wait for the started threads to transition to
1601                  * TD_INITIALIZED.
1602                  */
1603                 fio_gettime(&this_start, NULL);
1604                 left = this_jobs;
1605                 while (left && !fio_abort) {
1606                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1607                                 break;
1608
1609                         usleep(100000);
1610
1611                         for (i = 0; i < this_jobs; i++) {
1612                                 td = map[i];
1613                                 if (!td)
1614                                         continue;
1615                                 if (td->runstate == TD_INITIALIZED) {
1616                                         map[i] = NULL;
1617                                         left--;
1618                                 } else if (td->runstate >= TD_EXITED) {
1619                                         map[i] = NULL;
1620                                         left--;
1621                                         todo--;
1622                                         nr_running++; /* work-around... */
1623                                 }
1624                         }
1625                 }
1626
1627                 if (left) {
1628                         log_err("fio: %d jobs failed to start\n", left);
1629                         for (i = 0; i < this_jobs; i++) {
1630                                 td = map[i];
1631                                 if (!td)
1632                                         continue;
1633                                 kill(td->pid, SIGTERM);
1634                         }
1635                         break;
1636                 }
1637
1638                 /*
1639                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1640                  */
1641                 for_each_td(td, i) {
1642                         if (td->runstate != TD_INITIALIZED)
1643                                 continue;
1644
1645                         if (in_ramp_time(td))
1646                                 td_set_runstate(td, TD_RAMP);
1647                         else
1648                                 td_set_runstate(td, TD_RUNNING);
1649                         nr_running++;
1650                         nr_started--;
1651                         m_rate += td->o.ratemin[0] + td->o.ratemin[1];
1652                         t_rate += td->o.rate[0] + td->o.rate[1];
1653                         todo--;
1654                         fio_mutex_up(td->mutex);
1655                 }
1656
1657                 reap_threads(&nr_running, &t_rate, &m_rate);
1658
1659                 if (todo)
1660                         usleep(100000);
1661         }
1662
1663         while (nr_running) {
1664                 reap_threads(&nr_running, &t_rate, &m_rate);
1665                 usleep(10000);
1666         }
1667
1668         update_io_ticks();
1669         fio_unpin_memory();
1670 }
1671
1672 int main(int argc, char *argv[])
1673 {
1674         long ps;
1675
1676         sinit();
1677         init_rand(&__fio_rand_state);
1678
1679         /*
1680          * We need locale for number printing, if it isn't set then just
1681          * go with the US format.
1682          */
1683         if (!getenv("LC_NUMERIC"))
1684                 setlocale(LC_NUMERIC, "en_US");
1685
1686         ps = sysconf(_SC_PAGESIZE);
1687         if (ps < 0) {
1688                 log_err("Failed to get page size\n");
1689                 return 1;
1690         }
1691
1692         page_size = ps;
1693         page_mask = ps - 1;
1694
1695         fio_keywords_init();
1696
1697         if (parse_options(argc, argv))
1698                 return 1;
1699
1700         if (exec_profile && load_profile(exec_profile))
1701                 return 1;
1702
1703         if (!thread_number)
1704                 return 0;
1705
1706         if (write_bw_log) {
1707                 setup_log(&agg_io_log[DDIR_READ]);
1708                 setup_log(&agg_io_log[DDIR_WRITE]);
1709         }
1710
1711         startup_mutex = fio_mutex_init(0);
1712         if (startup_mutex == NULL)
1713                 return 1;
1714         writeout_mutex = fio_mutex_init(1);
1715         if (writeout_mutex == NULL)
1716                 return 1;
1717
1718         set_genesis_time();
1719
1720         posix_timer_setup();
1721         status_timer_arm();
1722
1723         cgroup_list = smalloc(sizeof(*cgroup_list));
1724         INIT_FLIST_HEAD(cgroup_list);
1725
1726         run_threads();
1727
1728         if (!fio_abort) {
1729                 show_run_stats();
1730                 if (write_bw_log) {
1731                         __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1732                         __finish_log(agg_io_log[DDIR_WRITE],
1733                                         "agg-write_bw.log");
1734                 }
1735         }
1736
1737         cgroup_kill(cgroup_list);
1738         sfree(cgroup_list);
1739         sfree(cgroup_mnt);
1740
1741         posix_timer_teardown();
1742         fio_mutex_remove(startup_mutex);
1743         fio_mutex_remove(writeout_mutex);
1744         return exit_value;
1745 }