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