Add CPU affinity support to Windows
[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_RAMP)
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         fio_gettime(&td->tv_cache, NULL);
378 }
379
380 static inline void update_tv_cache(struct thread_data *td)
381 {
382         if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
383                 __update_tv_cache(td);
384 }
385
386 static int break_on_this_error(struct thread_data *td, int *retptr)
387 {
388         int ret = *retptr;
389
390         if (ret < 0 || td->error) {
391                 int err;
392
393                 if (!td->o.continue_on_error)
394                         return 1;
395
396                 if (ret < 0)
397                         err = -ret;
398                 else
399                         err = td->error;
400
401                 if (td_non_fatal_error(err)) {
402                         /*
403                          * Continue with the I/Os in case of
404                          * a non fatal error.
405                          */
406                         update_error_count(td, err);
407                         td_clear_error(td);
408                         *retptr = 0;
409                         return 0;
410                 } else if (td->o.fill_device && err == ENOSPC) {
411                         /*
412                          * We expect to hit this error if
413                          * fill_device option is set.
414                          */
415                         td_clear_error(td);
416                         td->terminate = 1;
417                         return 1;
418                 } else {
419                         /*
420                          * Stop the I/O in case of a fatal
421                          * error.
422                          */
423                         update_error_count(td, err);
424                         return 1;
425                 }
426         }
427
428         return 0;
429 }
430
431 /*
432  * The main verify engine. Runs over the writes we previously submitted,
433  * reads the blocks back in, and checks the crc/md5 of the data.
434  */
435 static void do_verify(struct thread_data *td)
436 {
437         struct fio_file *f;
438         struct io_u *io_u;
439         int ret, min_events;
440         unsigned int i;
441
442         dprint(FD_VERIFY, "starting loop\n");
443
444         /*
445          * sync io first and invalidate cache, to make sure we really
446          * read from disk.
447          */
448         for_each_file(td, f, i) {
449                 if (!fio_file_open(f))
450                         continue;
451                 if (fio_io_sync(td, f))
452                         break;
453                 if (file_invalidate_cache(td, f))
454                         break;
455         }
456
457         if (td->error)
458                 return;
459
460         td_set_runstate(td, TD_VERIFYING);
461
462         io_u = NULL;
463         while (!td->terminate) {
464                 int ret2, full;
465
466                 update_tv_cache(td);
467
468                 if (runtime_exceeded(td, &td->tv_cache)) {
469                         __update_tv_cache(td);
470                         if (runtime_exceeded(td, &td->tv_cache)) {
471                                 td->terminate = 1;
472                                 break;
473                         }
474                 }
475
476                 io_u = __get_io_u(td);
477                 if (!io_u)
478                         break;
479
480                 if (get_next_verify(td, io_u)) {
481                         put_io_u(td, io_u);
482                         break;
483                 }
484
485                 if (td_io_prep(td, io_u)) {
486                         put_io_u(td, io_u);
487                         break;
488                 }
489
490                 if (td->o.verify_async)
491                         io_u->end_io = verify_io_u_async;
492                 else
493                         io_u->end_io = verify_io_u;
494
495                 ret = td_io_queue(td, io_u);
496                 switch (ret) {
497                 case FIO_Q_COMPLETED:
498                         if (io_u->error) {
499                                 ret = -io_u->error;
500                                 clear_io_u(td, io_u);
501                         } else if (io_u->resid) {
502                                 int bytes = io_u->xfer_buflen - io_u->resid;
503
504                                 /*
505                                  * zero read, fail
506                                  */
507                                 if (!bytes) {
508                                         td_verror(td, EIO, "full resid");
509                                         put_io_u(td, io_u);
510                                         break;
511                                 }
512
513                                 io_u->xfer_buflen = io_u->resid;
514                                 io_u->xfer_buf += bytes;
515                                 io_u->offset += bytes;
516
517                                 if (ddir_rw(io_u->ddir))
518                                         td->ts.short_io_u[io_u->ddir]++;
519
520                                 f = io_u->file;
521                                 if (io_u->offset == f->real_file_size)
522                                         goto sync_done;
523
524                                 requeue_io_u(td, &io_u);
525                         } else {
526 sync_done:
527                                 ret = io_u_sync_complete(td, io_u, NULL);
528                                 if (ret < 0)
529                                         break;
530                         }
531                         continue;
532                 case FIO_Q_QUEUED:
533                         break;
534                 case FIO_Q_BUSY:
535                         requeue_io_u(td, &io_u);
536                         ret2 = td_io_commit(td);
537                         if (ret2 < 0)
538                                 ret = ret2;
539                         break;
540                 default:
541                         assert(ret < 0);
542                         td_verror(td, -ret, "td_io_queue");
543                         break;
544                 }
545
546                 if (break_on_this_error(td, &ret))
547                         break;
548
549                 /*
550                  * if we can queue more, do so. but check if there are
551                  * completed io_u's first.
552                  */
553                 full = queue_full(td) || ret == FIO_Q_BUSY;
554                 if (full || !td->o.iodepth_batch_complete) {
555                         min_events = min(td->o.iodepth_batch_complete,
556                                          td->cur_depth);
557                         if (full && !min_events)
558                                 min_events = 1;
559
560                         do {
561                                 /*
562                                  * Reap required number of io units, if any,
563                                  * and do the verification on them through
564                                  * the callback handler
565                                  */
566                                 if (io_u_queued_complete(td, min_events, NULL) < 0) {
567                                         ret = -1;
568                                         break;
569                                 }
570                         } while (full && (td->cur_depth > td->o.iodepth_low));
571                 }
572                 if (ret < 0)
573                         break;
574         }
575
576         if (!td->error) {
577                 min_events = td->cur_depth;
578
579                 if (min_events)
580                         ret = io_u_queued_complete(td, min_events, NULL);
581         } else
582                 cleanup_pending_aio(td);
583
584         td_set_runstate(td, TD_RUNNING);
585
586         dprint(FD_VERIFY, "exiting loop\n");
587 }
588
589 /*
590  * Main IO worker function. It retrieves io_u's to process and queues
591  * and reaps them, checking for rate and errors along the way.
592  */
593 static void do_io(struct thread_data *td)
594 {
595         unsigned int i;
596         int ret = 0;
597
598         if (in_ramp_time(td))
599                 td_set_runstate(td, TD_RAMP);
600         else
601                 td_set_runstate(td, TD_RUNNING);
602
603         while ( (td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
604                 (!flist_empty(&td->trim_list)) ||
605                 ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) ) {
606                 struct timeval comp_time;
607                 unsigned long bytes_done[2] = { 0, 0 };
608                 int min_evts = 0;
609                 struct io_u *io_u;
610                 int ret2, full;
611
612                 if (td->terminate)
613                         break;
614
615                 update_tv_cache(td);
616
617                 if (runtime_exceeded(td, &td->tv_cache)) {
618                         __update_tv_cache(td);
619                         if (runtime_exceeded(td, &td->tv_cache)) {
620                                 td->terminate = 1;
621                                 break;
622                         }
623                 }
624
625                 io_u = get_io_u(td);
626                 if (!io_u)
627                         break;
628
629                 /*
630                  * Add verification end_io handler, if asked to verify
631                  * a previously written file.
632                  */
633                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
634                     !td_rw(td)) {
635                         if (td->o.verify_async)
636                                 io_u->end_io = verify_io_u_async;
637                         else
638                                 io_u->end_io = verify_io_u;
639                         td_set_runstate(td, TD_VERIFYING);
640                 } else if (in_ramp_time(td))
641                         td_set_runstate(td, TD_RAMP);
642                 else
643                         td_set_runstate(td, TD_RUNNING);
644
645                 ret = td_io_queue(td, io_u);
646                 switch (ret) {
647                 case FIO_Q_COMPLETED:
648                         if (io_u->error) {
649                                 ret = -io_u->error;
650                                 clear_io_u(td, io_u);
651                         } else if (io_u->resid) {
652                                 int bytes = io_u->xfer_buflen - io_u->resid;
653                                 struct fio_file *f = io_u->file;
654
655                                 /*
656                                  * zero read, fail
657                                  */
658                                 if (!bytes) {
659                                         td_verror(td, EIO, "full resid");
660                                         put_io_u(td, io_u);
661                                         break;
662                                 }
663
664                                 io_u->xfer_buflen = io_u->resid;
665                                 io_u->xfer_buf += bytes;
666                                 io_u->offset += bytes;
667
668                                 if (ddir_rw(io_u->ddir))
669                                         td->ts.short_io_u[io_u->ddir]++;
670
671                                 if (io_u->offset == f->real_file_size)
672                                         goto sync_done;
673
674                                 requeue_io_u(td, &io_u);
675                         } else {
676 sync_done:
677                                 if (__should_check_rate(td, 0) ||
678                                     __should_check_rate(td, 1))
679                                         fio_gettime(&comp_time, NULL);
680
681                                 ret = io_u_sync_complete(td, io_u, bytes_done);
682                                 if (ret < 0)
683                                         break;
684                         }
685                         break;
686                 case FIO_Q_QUEUED:
687                         /*
688                          * if the engine doesn't have a commit hook,
689                          * the io_u is really queued. if it does have such
690                          * a hook, it has to call io_u_queued() itself.
691                          */
692                         if (td->io_ops->commit == NULL)
693                                 io_u_queued(td, io_u);
694                         break;
695                 case FIO_Q_BUSY:
696                         requeue_io_u(td, &io_u);
697                         ret2 = td_io_commit(td);
698                         if (ret2 < 0)
699                                 ret = ret2;
700                         break;
701                 default:
702                         assert(ret < 0);
703                         put_io_u(td, io_u);
704                         break;
705                 }
706
707                 if (break_on_this_error(td, &ret))
708                         break;
709
710                 /*
711                  * See if we need to complete some commands
712                  */
713                 full = queue_full(td) || ret == FIO_Q_BUSY;
714                 if (full || !td->o.iodepth_batch_complete) {
715                         min_evts = min(td->o.iodepth_batch_complete,
716                                         td->cur_depth);
717                         if (full && !min_evts)
718                                 min_evts = 1;
719
720                         if (__should_check_rate(td, 0) ||
721                             __should_check_rate(td, 1))
722                                 fio_gettime(&comp_time, NULL);
723
724                         do {
725                                 ret = io_u_queued_complete(td, min_evts, bytes_done);
726                                 if (ret < 0)
727                                         break;
728
729                         } while (full && (td->cur_depth > td->o.iodepth_low));
730                 }
731
732                 if (ret < 0)
733                         break;
734                 if (!(bytes_done[0] + bytes_done[1]))
735                         continue;
736
737                 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
738                         if (check_min_rate(td, &comp_time, bytes_done)) {
739                                 if (exitall_on_terminate)
740                                         terminate_threads(td->groupid);
741                                 td_verror(td, EIO, "check_min_rate");
742                                 break;
743                         }
744                 }
745
746                 if (td->o.thinktime) {
747                         unsigned long long b;
748
749                         b = td->io_blocks[0] + td->io_blocks[1];
750                         if (!(b % td->o.thinktime_blocks)) {
751                                 int left;
752
753                                 if (td->o.thinktime_spin)
754                                         usec_spin(td->o.thinktime_spin);
755
756                                 left = td->o.thinktime - td->o.thinktime_spin;
757                                 if (left)
758                                         usec_sleep(td, left);
759                         }
760                 }
761         }
762
763         if (td->trim_entries)
764                 printf("trim entries %ld\n", td->trim_entries);
765
766         if (td->o.fill_device && td->error == ENOSPC) {
767                 td->error = 0;
768                 td->terminate = 1;
769         }
770         if (!td->error) {
771                 struct fio_file *f;
772
773                 i = td->cur_depth;
774                 if (i) {
775                         ret = io_u_queued_complete(td, i, NULL);
776                         if (td->o.fill_device && td->error == ENOSPC)
777                                 td->error = 0;
778                 }
779
780                 if (should_fsync(td) && td->o.end_fsync) {
781                         td_set_runstate(td, TD_FSYNCING);
782
783                         for_each_file(td, f, i) {
784                                 if (!fio_file_open(f))
785                                         continue;
786                                 fio_io_sync(td, f);
787                         }
788                 }
789         } else
790                 cleanup_pending_aio(td);
791
792         /*
793          * stop job if we failed doing any IO
794          */
795         if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
796                 td->done = 1;
797 }
798
799 static void cleanup_io_u(struct thread_data *td)
800 {
801         struct flist_head *entry, *n;
802         struct io_u *io_u;
803
804         flist_for_each_safe(entry, n, &td->io_u_freelist) {
805                 io_u = flist_entry(entry, struct io_u, list);
806
807                 flist_del(&io_u->list);
808                 fio_memfree(io_u, sizeof(*io_u));
809         }
810
811         free_io_mem(td);
812 }
813
814 static int init_io_u(struct thread_data *td)
815 {
816         struct io_u *io_u;
817         unsigned int max_bs;
818         int cl_align, i, max_units;
819         char *p;
820
821         max_units = td->o.iodepth;
822         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
823         td->orig_buffer_size = (unsigned long long) max_bs
824                                         * (unsigned long long) max_units;
825
826         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
827                 unsigned long bs;
828
829                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
830                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
831         }
832
833         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
834                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
835                 return 1;
836         }
837
838         if (allocate_io_mem(td))
839                 return 1;
840
841         if (td->o.odirect || td->o.mem_align ||
842             (td->io_ops->flags & FIO_RAWIO))
843                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
844         else
845                 p = td->orig_buffer;
846
847         cl_align = os_cache_line_size();
848
849         for (i = 0; i < max_units; i++) {
850                 void *ptr;
851
852                 if (td->terminate)
853                         return 1;
854
855                 ptr = fio_memalign(cl_align, sizeof(*io_u));
856                 if (!ptr) {
857                         log_err("fio: unable to allocate aligned memory\n");
858                         break;
859                 }
860
861                 io_u = ptr;
862                 memset(io_u, 0, sizeof(*io_u));
863                 INIT_FLIST_HEAD(&io_u->list);
864                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
865
866                 if (!(td->io_ops->flags & FIO_NOIO)) {
867                         io_u->buf = p + max_bs * i;
868                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
869
870                         if (td_write(td) && !td->o.refill_buffers)
871                                 io_u_fill_buffer(td, io_u, max_bs);
872                         else if (td_write(td) && td->o.verify_pattern_bytes) {
873                                 /*
874                                  * Fill the buffer with the pattern if we are
875                                  * going to be doing writes.
876                                  */
877                                 fill_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
878                         }
879                 }
880
881                 io_u->index = i;
882                 io_u->flags = IO_U_F_FREE;
883                 flist_add(&io_u->list, &td->io_u_freelist);
884         }
885
886         return 0;
887 }
888
889 static int switch_ioscheduler(struct thread_data *td)
890 {
891         char tmp[256], tmp2[128];
892         FILE *f;
893         int ret;
894
895         if (td->io_ops->flags & FIO_DISKLESSIO)
896                 return 0;
897
898         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
899
900         f = fopen(tmp, "r+");
901         if (!f) {
902                 if (errno == ENOENT) {
903                         log_err("fio: os or kernel doesn't support IO scheduler"
904                                 " switching\n");
905                         return 0;
906                 }
907                 td_verror(td, errno, "fopen iosched");
908                 return 1;
909         }
910
911         /*
912          * Set io scheduler.
913          */
914         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
915         if (ferror(f) || ret != 1) {
916                 td_verror(td, errno, "fwrite");
917                 fclose(f);
918                 return 1;
919         }
920
921         rewind(f);
922
923         /*
924          * Read back and check that the selected scheduler is now the default.
925          */
926         ret = fread(tmp, 1, sizeof(tmp), f);
927         if (ferror(f) || ret < 0) {
928                 td_verror(td, errno, "fread");
929                 fclose(f);
930                 return 1;
931         }
932
933         sprintf(tmp2, "[%s]", td->o.ioscheduler);
934         if (!strstr(tmp, tmp2)) {
935                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
936                 td_verror(td, EINVAL, "iosched_switch");
937                 fclose(f);
938                 return 1;
939         }
940
941         fclose(f);
942         return 0;
943 }
944
945 static int keep_running(struct thread_data *td)
946 {
947         unsigned long long io_done;
948
949         if (td->done)
950                 return 0;
951         if (td->o.time_based)
952                 return 1;
953         if (td->o.loops) {
954                 td->o.loops--;
955                 return 1;
956         }
957
958         io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
959                         + td->io_skip_bytes;
960         if (io_done < td->o.size)
961                 return 1;
962
963         return 0;
964 }
965
966 static void reset_io_counters(struct thread_data *td)
967 {
968         td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
969         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
970         td->zone_bytes = 0;
971         td->rate_bytes[0] = td->rate_bytes[1] = 0;
972         td->rate_blocks[0] = td->rate_blocks[1] = 0;
973
974         td->last_was_sync = 0;
975
976         /*
977          * reset file done count if we are to start over
978          */
979         if (td->o.time_based || td->o.loops)
980                 td->nr_done_files = 0;
981 }
982
983 void reset_all_stats(struct thread_data *td)
984 {
985         struct timeval tv;
986         int i;
987
988         reset_io_counters(td);
989
990         for (i = 0; i < 2; i++) {
991                 td->io_bytes[i] = 0;
992                 td->io_blocks[i] = 0;
993                 td->io_issues[i] = 0;
994                 td->ts.total_io_u[i] = 0;
995         }
996
997         fio_gettime(&tv, NULL);
998         td->ts.runtime[0] = 0;
999         td->ts.runtime[1] = 0;
1000         memcpy(&td->epoch, &tv, sizeof(tv));
1001         memcpy(&td->start, &tv, sizeof(tv));
1002 }
1003
1004 static void clear_io_state(struct thread_data *td)
1005 {
1006         struct fio_file *f;
1007         unsigned int i;
1008
1009         reset_io_counters(td);
1010
1011         close_files(td);
1012         for_each_file(td, f, i)
1013                 fio_file_clear_done(f);
1014
1015         /*
1016          * Set the same seed to get repeatable runs
1017          */
1018         td_fill_rand_seeds(td);
1019 }
1020
1021 static int exec_string(const char *string)
1022 {
1023         int ret, newlen = strlen(string) + 1 + 8;
1024         char *str;
1025
1026         str = malloc(newlen);
1027         sprintf(str, "sh -c %s", string);
1028
1029         ret = system(str);
1030         if (ret == -1)
1031                 log_err("fio: exec of cmd <%s> failed\n", str);
1032
1033         free(str);
1034         return ret;
1035 }
1036
1037 /*
1038  * Entry point for the thread based jobs. The process based jobs end up
1039  * here as well, after a little setup.
1040  */
1041 static void *thread_main(void *data)
1042 {
1043         unsigned long long elapsed;
1044         struct thread_data *td = data;
1045         pthread_condattr_t attr;
1046         int clear_state;
1047
1048         if (!td->o.use_thread) {
1049                 setsid();
1050                 td->pid = getpid();
1051         } else
1052                 td->pid = gettid();
1053
1054         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1055
1056         INIT_FLIST_HEAD(&td->io_u_freelist);
1057         INIT_FLIST_HEAD(&td->io_u_busylist);
1058         INIT_FLIST_HEAD(&td->io_u_requeues);
1059         INIT_FLIST_HEAD(&td->io_log_list);
1060         INIT_FLIST_HEAD(&td->io_hist_list);
1061         INIT_FLIST_HEAD(&td->verify_list);
1062         INIT_FLIST_HEAD(&td->trim_list);
1063         pthread_mutex_init(&td->io_u_lock, NULL);
1064         td->io_hist_tree = RB_ROOT;
1065
1066         pthread_condattr_init(&attr);
1067         pthread_cond_init(&td->verify_cond, &attr);
1068         pthread_cond_init(&td->free_cond, &attr);
1069
1070         td_set_runstate(td, TD_INITIALIZED);
1071         dprint(FD_MUTEX, "up startup_mutex\n");
1072         fio_mutex_up(startup_mutex);
1073         dprint(FD_MUTEX, "wait on td->mutex\n");
1074         fio_mutex_down(td->mutex);
1075         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1076
1077         /*
1078          * the ->mutex mutex is now no longer used, close it to avoid
1079          * eating a file descriptor
1080          */
1081         fio_mutex_remove(td->mutex);
1082
1083         /*
1084          * A new gid requires privilege, so we need to do this before setting
1085          * the uid.
1086          */
1087         if (td->o.gid != -1U && setgid(td->o.gid)) {
1088                 td_verror(td, errno, "setgid");
1089                 goto err;
1090         }
1091         if (td->o.uid != -1U && setuid(td->o.uid)) {
1092                 td_verror(td, errno, "setuid");
1093                 goto err;
1094         }
1095
1096         /*
1097          * If we have a gettimeofday() thread, make sure we exclude that
1098          * thread from this job
1099          */
1100         if (td->o.gtod_cpu)
1101                 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1102
1103         /*
1104          * Set affinity first, in case it has an impact on the memory
1105          * allocations.
1106          */
1107         if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1108                 td_verror(td, errno, "cpu_set_affinity");
1109                 goto err;
1110         }
1111
1112         /*
1113          * May alter parameters that init_io_u() will use, so we need to
1114          * do this first.
1115          */
1116         if (init_iolog(td))
1117                 goto err;
1118
1119         if (init_io_u(td))
1120                 goto err;
1121
1122         if (td->o.verify_async && verify_async_init(td))
1123                 goto err;
1124
1125         if (td->ioprio_set) {
1126                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1127                         td_verror(td, errno, "ioprio_set");
1128                         goto err;
1129                 }
1130         }
1131
1132         if (td->o.cgroup_weight && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1133                 goto err;
1134
1135         if (nice(td->o.nice) == -1) {
1136                 td_verror(td, errno, "nice");
1137                 goto err;
1138         }
1139
1140         if (td->o.ioscheduler && switch_ioscheduler(td))
1141                 goto err;
1142
1143         if (!td->o.create_serialize && setup_files(td))
1144                 goto err;
1145
1146         if (td_io_init(td))
1147                 goto err;
1148
1149         if (init_random_map(td))
1150                 goto err;
1151
1152         if (td->o.exec_prerun) {
1153                 if (exec_string(td->o.exec_prerun))
1154                         goto err;
1155         }
1156
1157         if (td->o.pre_read) {
1158                 if (pre_read_files(td) < 0)
1159                         goto err;
1160         }
1161
1162         fio_gettime(&td->epoch, NULL);
1163         getrusage(RUSAGE_SELF, &td->ts.ru_start);
1164
1165         clear_state = 0;
1166         while (keep_running(td)) {
1167                 fio_gettime(&td->start, NULL);
1168                 memcpy(&td->ts.stat_sample_time[0], &td->start,
1169                                 sizeof(td->start));
1170                 memcpy(&td->ts.stat_sample_time[1], &td->start,
1171                                 sizeof(td->start));
1172                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1173
1174                 if (td->o.ratemin[0] || td->o.ratemin[1])
1175                         memcpy(&td->lastrate, &td->ts.stat_sample_time,
1176                                                         sizeof(td->lastrate));
1177
1178                 if (clear_state)
1179                         clear_io_state(td);
1180
1181                 prune_io_piece_log(td);
1182
1183                 do_io(td);
1184
1185                 clear_state = 1;
1186
1187                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1188                         elapsed = utime_since_now(&td->start);
1189                         td->ts.runtime[DDIR_READ] += elapsed;
1190                 }
1191                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1192                         elapsed = utime_since_now(&td->start);
1193                         td->ts.runtime[DDIR_WRITE] += elapsed;
1194                 }
1195
1196                 if (td->error || td->terminate)
1197                         break;
1198
1199                 if (!td->o.do_verify ||
1200                     td->o.verify == VERIFY_NONE ||
1201                     (td->io_ops->flags & FIO_UNIDIR))
1202                         continue;
1203
1204                 clear_io_state(td);
1205
1206                 fio_gettime(&td->start, NULL);
1207
1208                 do_verify(td);
1209
1210                 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1211
1212                 if (td->error || td->terminate)
1213                         break;
1214         }
1215
1216         update_rusage_stat(td);
1217         td->ts.runtime[0] = (td->ts.runtime[0] + 999) / 1000;
1218         td->ts.runtime[1] = (td->ts.runtime[1] + 999) / 1000;
1219         td->ts.total_run_time = mtime_since_now(&td->epoch);
1220         td->ts.io_bytes[0] = td->io_bytes[0];
1221         td->ts.io_bytes[1] = td->io_bytes[1];
1222
1223         fio_mutex_down(writeout_mutex);
1224         if (td->ts.bw_log) {
1225                 if (td->o.bw_log_file) {
1226                         finish_log_named(td, td->ts.bw_log,
1227                                                 td->o.bw_log_file, "bw");
1228                 } else
1229                         finish_log(td, td->ts.bw_log, "bw");
1230         }
1231         if (td->ts.lat_log) {
1232                 if (td->o.lat_log_file) {
1233                         finish_log_named(td, td->ts.lat_log,
1234                                                 td->o.lat_log_file, "lat");
1235                 } else
1236                         finish_log(td, td->ts.lat_log, "lat");
1237         }
1238         if (td->ts.slat_log) {
1239                 if (td->o.lat_log_file) {
1240                         finish_log_named(td, td->ts.slat_log,
1241                                                 td->o.lat_log_file, "slat");
1242                 } else
1243                         finish_log(td, td->ts.slat_log, "slat");
1244         }
1245         if (td->ts.clat_log) {
1246                 if (td->o.lat_log_file) {
1247                         finish_log_named(td, td->ts.clat_log,
1248                                                 td->o.lat_log_file, "clat");
1249                 } else
1250                         finish_log(td, td->ts.clat_log, "clat");
1251         }
1252         fio_mutex_up(writeout_mutex);
1253         if (td->o.exec_postrun)
1254                 exec_string(td->o.exec_postrun);
1255
1256         if (exitall_on_terminate)
1257                 terminate_threads(td->groupid);
1258
1259 err:
1260         if (td->error)
1261                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1262                                                         td->verror);
1263
1264         if (td->o.verify_async)
1265                 verify_async_exit(td);
1266
1267         close_and_free_files(td);
1268         close_ioengine(td);
1269         cleanup_io_u(td);
1270         cgroup_shutdown(td, &cgroup_mnt);
1271
1272         if (td->o.cpumask_set) {
1273                 int ret = fio_cpuset_exit(&td->o.cpumask);
1274
1275                 td_verror(td, ret, "fio_cpuset_exit");
1276         }
1277
1278         /*
1279          * do this very late, it will log file closing as well
1280          */
1281         if (td->o.write_iolog_file)
1282                 write_iolog_close(td);
1283
1284         options_mem_free(td);
1285         td_set_runstate(td, TD_EXITED);
1286         return (void *) (unsigned long) td->error;
1287 }
1288
1289 /*
1290  * We cannot pass the td data into a forked process, so attach the td and
1291  * pass it to the thread worker.
1292  */
1293 static int fork_main(int shmid, int offset)
1294 {
1295         struct thread_data *td;
1296         void *data, *ret;
1297
1298 #ifndef __hpux
1299         data = shmat(shmid, NULL, 0);
1300         if (data == (void *) -1) {
1301                 int __err = errno;
1302
1303                 perror("shmat");
1304                 return __err;
1305         }
1306 #else
1307         /*
1308          * HP-UX inherits shm mappings?
1309          */
1310         data = threads;
1311 #endif
1312
1313         td = data + offset * sizeof(struct thread_data);
1314         ret = thread_main(td);
1315         shmdt(data);
1316         return (int) (unsigned long) ret;
1317 }
1318
1319 /*
1320  * Run over the job map and reap the threads that have exited, if any.
1321  */
1322 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1323 {
1324         struct thread_data *td;
1325         int i, cputhreads, realthreads, pending, status, ret;
1326
1327         /*
1328          * reap exited threads (TD_EXITED -> TD_REAPED)
1329          */
1330         realthreads = pending = cputhreads = 0;
1331         for_each_td(td, i) {
1332                 int flags = 0;
1333
1334                 /*
1335                  * ->io_ops is NULL for a thread that has closed its
1336                  * io engine
1337                  */
1338                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1339                         cputhreads++;
1340                 else
1341                         realthreads++;
1342
1343                 if (!td->pid) {
1344                         pending++;
1345                         continue;
1346                 }
1347                 if (td->runstate == TD_REAPED)
1348                         continue;
1349                 if (td->o.use_thread) {
1350                         if (td->runstate == TD_EXITED) {
1351                                 td_set_runstate(td, TD_REAPED);
1352                                 goto reaped;
1353                         }
1354                         continue;
1355                 }
1356
1357                 flags = WNOHANG;
1358                 if (td->runstate == TD_EXITED)
1359                         flags = 0;
1360
1361                 /*
1362                  * check if someone quit or got killed in an unusual way
1363                  */
1364                 ret = waitpid(td->pid, &status, flags);
1365                 if (ret < 0) {
1366                         if (errno == ECHILD) {
1367                                 log_err("fio: pid=%d disappeared %d\n",
1368                                                 (int) td->pid, td->runstate);
1369                                 td_set_runstate(td, TD_REAPED);
1370                                 goto reaped;
1371                         }
1372                         perror("waitpid");
1373                 } else if (ret == td->pid) {
1374                         if (WIFSIGNALED(status)) {
1375                                 int sig = WTERMSIG(status);
1376
1377                                 if (sig != SIGTERM)
1378                                         log_err("fio: pid=%d, got signal=%d\n",
1379                                                         (int) td->pid, sig);
1380                                 td_set_runstate(td, TD_REAPED);
1381                                 goto reaped;
1382                         }
1383                         if (WIFEXITED(status)) {
1384                                 if (WEXITSTATUS(status) && !td->error)
1385                                         td->error = WEXITSTATUS(status);
1386
1387                                 td_set_runstate(td, TD_REAPED);
1388                                 goto reaped;
1389                         }
1390                 }
1391
1392                 /*
1393                  * thread is not dead, continue
1394                  */
1395                 pending++;
1396                 continue;
1397 reaped:
1398                 (*nr_running)--;
1399                 (*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
1400                 (*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
1401                 if (!td->pid)
1402                         pending--;
1403
1404                 if (td->error)
1405                         exit_value++;
1406
1407                 done_secs += mtime_since_now(&td->epoch) / 1000;
1408         }
1409
1410         if (*nr_running == cputhreads && !pending && realthreads)
1411                 terminate_threads(TERMINATE_ALL);
1412 }
1413
1414 static void *gtod_thread_main(void *data)
1415 {
1416         fio_mutex_up(startup_mutex);
1417
1418         /*
1419          * As long as we have jobs around, update the clock. It would be nice
1420          * to have some way of NOT hammering that CPU with gettimeofday(),
1421          * but I'm not sure what to use outside of a simple CPU nop to relax
1422          * it - we don't want to lose precision.
1423          */
1424         while (threads) {
1425                 fio_gtod_update();
1426                 nop;
1427         }
1428
1429         return NULL;
1430 }
1431
1432 static int fio_start_gtod_thread(void)
1433 {
1434         pthread_attr_t attr;
1435         int ret;
1436
1437         pthread_attr_init(&attr);
1438         pthread_attr_setstacksize(&attr, PTHREAD_STACK_MIN);
1439         ret = pthread_create(&gtod_thread, &attr, gtod_thread_main, NULL);
1440         pthread_attr_destroy(&attr);
1441         if (ret) {
1442                 log_err("Can't create gtod thread: %s\n", strerror(ret));
1443                 return 1;
1444         }
1445
1446         ret = pthread_detach(gtod_thread);
1447         if (ret) {
1448                 log_err("Can't detatch gtod thread: %s\n", strerror(ret));
1449                 return 1;
1450         }
1451
1452         dprint(FD_MUTEX, "wait on startup_mutex\n");
1453         fio_mutex_down(startup_mutex);
1454         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1455         return 0;
1456 }
1457
1458 /*
1459  * Main function for kicking off and reaping jobs, as needed.
1460  */
1461 static void run_threads(void)
1462 {
1463         struct thread_data *td;
1464         unsigned long spent;
1465         int i, todo, nr_running, m_rate, t_rate, nr_started;
1466
1467         if (fio_pin_memory())
1468                 return;
1469
1470         if (fio_gtod_offload && fio_start_gtod_thread())
1471                 return;
1472
1473         if (!terse_output) {
1474                 log_info("Starting ");
1475                 if (nr_thread)
1476                         log_info("%d thread%s", nr_thread,
1477                                                 nr_thread > 1 ? "s" : "");
1478                 if (nr_process) {
1479                         if (nr_thread)
1480                                 printf(" and ");
1481                         log_info("%d process%s", nr_process,
1482                                                 nr_process > 1 ? "es" : "");
1483                 }
1484                 log_info("\n");
1485                 fflush(stdout);
1486         }
1487
1488         set_sig_handlers();
1489
1490         todo = thread_number;
1491         nr_running = 0;
1492         nr_started = 0;
1493         m_rate = t_rate = 0;
1494
1495         for_each_td(td, i) {
1496                 print_status_init(td->thread_number - 1);
1497
1498                 if (!td->o.create_serialize) {
1499                         init_disk_util(td);
1500                         continue;
1501                 }
1502
1503                 /*
1504                  * do file setup here so it happens sequentially,
1505                  * we don't want X number of threads getting their
1506                  * client data interspersed on disk
1507                  */
1508                 if (setup_files(td)) {
1509                         exit_value++;
1510                         if (td->error)
1511                                 log_err("fio: pid=%d, err=%d/%s\n",
1512                                         (int) td->pid, td->error, td->verror);
1513                         td_set_runstate(td, TD_REAPED);
1514                         todo--;
1515                 } else {
1516                         struct fio_file *f;
1517                         unsigned int j;
1518
1519                         /*
1520                          * for sharing to work, each job must always open
1521                          * its own files. so close them, if we opened them
1522                          * for creation
1523                          */
1524                         for_each_file(td, f, j) {
1525                                 if (fio_file_open(f))
1526                                         td_io_close_file(td, f);
1527                         }
1528                 }
1529
1530                 init_disk_util(td);
1531         }
1532
1533         set_genesis_time();
1534
1535         while (todo) {
1536                 struct thread_data *map[REAL_MAX_JOBS];
1537                 struct timeval this_start;
1538                 int this_jobs = 0, left;
1539
1540                 /*
1541                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1542                  */
1543                 for_each_td(td, i) {
1544                         if (td->runstate != TD_NOT_CREATED)
1545                                 continue;
1546
1547                         /*
1548                          * never got a chance to start, killed by other
1549                          * thread for some reason
1550                          */
1551                         if (td->terminate) {
1552                                 todo--;
1553                                 continue;
1554                         }
1555
1556                         if (td->o.start_delay) {
1557                                 spent = mtime_since_genesis();
1558
1559                                 if (td->o.start_delay * 1000 > spent)
1560                                         continue;
1561                         }
1562
1563                         if (td->o.stonewall && (nr_started || nr_running)) {
1564                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1565                                                         td->o.name);
1566                                 break;
1567                         }
1568
1569                         /*
1570                          * Set state to created. Thread will transition
1571                          * to TD_INITIALIZED when it's done setting up.
1572                          */
1573                         td_set_runstate(td, TD_CREATED);
1574                         map[this_jobs++] = td;
1575                         nr_started++;
1576
1577                         if (td->o.use_thread) {
1578                                 int ret;
1579
1580                                 dprint(FD_PROCESS, "will pthread_create\n");
1581                                 ret = pthread_create(&td->thread, NULL,
1582                                                         thread_main, td);
1583                                 if (ret) {
1584                                         log_err("pthread_create: %s\n",
1585                                                         strerror(ret));
1586                                         nr_started--;
1587                                         break;
1588                                 }
1589                                 ret = pthread_detach(td->thread);
1590                                 if (ret)
1591                                         log_err("pthread_detach: %s",
1592                                                         strerror(ret));
1593                         } else {
1594                                 pid_t pid;
1595                                 dprint(FD_PROCESS, "will fork\n");
1596                                 pid = fork();
1597                                 if (!pid) {
1598                                         int ret = fork_main(shm_id, i);
1599
1600                                         _exit(ret);
1601                                 } else if (i == fio_debug_jobno)
1602                                         *fio_debug_jobp = pid;
1603                         }
1604                         dprint(FD_MUTEX, "wait on startup_mutex\n");
1605                         if (fio_mutex_down_timeout(startup_mutex, 10)) {
1606                                 log_err("fio: job startup hung? exiting.\n");
1607                                 terminate_threads(TERMINATE_ALL);
1608                                 fio_abort = 1;
1609                                 nr_started--;
1610                                 break;
1611                         }
1612                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1613                 }
1614
1615                 /*
1616                  * Wait for the started threads to transition to
1617                  * TD_INITIALIZED.
1618                  */
1619                 fio_gettime(&this_start, NULL);
1620                 left = this_jobs;
1621                 while (left && !fio_abort) {
1622                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1623                                 break;
1624
1625                         usleep(100000);
1626
1627                         for (i = 0; i < this_jobs; i++) {
1628                                 td = map[i];
1629                                 if (!td)
1630                                         continue;
1631                                 if (td->runstate == TD_INITIALIZED) {
1632                                         map[i] = NULL;
1633                                         left--;
1634                                 } else if (td->runstate >= TD_EXITED) {
1635                                         map[i] = NULL;
1636                                         left--;
1637                                         todo--;
1638                                         nr_running++; /* work-around... */
1639                                 }
1640                         }
1641                 }
1642
1643                 if (left) {
1644                         log_err("fio: %d jobs failed to start\n", left);
1645                         for (i = 0; i < this_jobs; i++) {
1646                                 td = map[i];
1647                                 if (!td)
1648                                         continue;
1649                                 kill(td->pid, SIGTERM);
1650                         }
1651                         break;
1652                 }
1653
1654                 /*
1655                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1656                  */
1657                 for_each_td(td, i) {
1658                         if (td->runstate != TD_INITIALIZED)
1659                                 continue;
1660
1661                         if (in_ramp_time(td))
1662                                 td_set_runstate(td, TD_RAMP);
1663                         else
1664                                 td_set_runstate(td, TD_RUNNING);
1665                         nr_running++;
1666                         nr_started--;
1667                         m_rate += td->o.ratemin[0] + td->o.ratemin[1];
1668                         t_rate += td->o.rate[0] + td->o.rate[1];
1669                         todo--;
1670                         fio_mutex_up(td->mutex);
1671                 }
1672
1673                 reap_threads(&nr_running, &t_rate, &m_rate);
1674
1675                 if (todo)
1676                         usleep(100000);
1677         }
1678
1679         while (nr_running) {
1680                 reap_threads(&nr_running, &t_rate, &m_rate);
1681                 usleep(10000);
1682         }
1683
1684         update_io_ticks();
1685         fio_unpin_memory();
1686 }
1687
1688 int main(int argc, char *argv[])
1689 {
1690         long ps;
1691
1692         sinit();
1693         init_rand(&__fio_rand_state);
1694
1695         /*
1696          * We need locale for number printing, if it isn't set then just
1697          * go with the US format.
1698          */
1699         if (!getenv("LC_NUMERIC"))
1700                 setlocale(LC_NUMERIC, "en_US");
1701
1702         ps = sysconf(_SC_PAGESIZE);
1703         if (ps < 0) {
1704                 log_err("Failed to get page size\n");
1705                 return 1;
1706         }
1707
1708         page_size = ps;
1709         page_mask = ps - 1;
1710
1711         fio_keywords_init();
1712
1713         if (parse_options(argc, argv))
1714                 return 1;
1715
1716         if (exec_profile && load_profile(exec_profile))
1717                 return 1;
1718
1719         if (!thread_number)
1720                 return 0;
1721
1722         if (write_bw_log) {
1723                 setup_log(&agg_io_log[DDIR_READ]);
1724                 setup_log(&agg_io_log[DDIR_WRITE]);
1725         }
1726
1727         startup_mutex = fio_mutex_init(0);
1728         if (startup_mutex == NULL)
1729                 return 1;
1730         writeout_mutex = fio_mutex_init(1);
1731         if (writeout_mutex == NULL)
1732                 return 1;
1733
1734         set_genesis_time();
1735         create_disk_util_thread();
1736
1737         cgroup_list = smalloc(sizeof(*cgroup_list));
1738         INIT_FLIST_HEAD(cgroup_list);
1739
1740         run_threads();
1741
1742         if (!fio_abort) {
1743                 show_run_stats();
1744                 if (write_bw_log) {
1745                         __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1746                         __finish_log(agg_io_log[DDIR_WRITE],
1747                                         "agg-write_bw.log");
1748                 }
1749         }
1750
1751         cgroup_kill(cgroup_list);
1752         sfree(cgroup_list);
1753         sfree(cgroup_mnt);
1754
1755         fio_mutex_remove(startup_mutex);
1756         fio_mutex_remove(writeout_mutex);
1757         return exit_value;
1758 }