[PATCH] syslet: fix leak of ring and ahu
[fio.git] / fio.c
1 /*
2  * fio - the flexible io tester
3  *
4  * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
5  * Copyright (C) 2006 Jens Axboe <axboe@kernel.dk>
6  *
7  * The license below covers all files distributed with fio unless otherwise
8  * noted in the file itself.
9  *
10  *  This program is free software; you can redistribute it and/or modify
11  *  it under the terms of the GNU General Public License version 2 as
12  *  published by the Free Software Foundation.
13  *
14  *  This program is distributed in the hope that it will be useful,
15  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
16  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
17  *  GNU General Public License for more details.
18  *
19  *  You should have received a copy of the GNU General Public License
20  *  along with this program; if not, write to the Free Software
21  *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
22  *
23  */
24 #include <unistd.h>
25 #include <fcntl.h>
26 #include <string.h>
27 #include <signal.h>
28 #include <time.h>
29 #include <locale.h>
30 #include <sys/stat.h>
31 #include <sys/wait.h>
32 #include <sys/ipc.h>
33 #include <sys/shm.h>
34 #include <sys/mman.h>
35
36 #include "fio.h"
37 #include "os.h"
38
39 static unsigned long page_mask;
40 #define ALIGN(buf)      \
41         (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
42
43 int groupid = 0;
44 int thread_number = 0;
45 int shm_id = 0;
46 int temp_stall_ts;
47
48 static volatile int startup_sem;
49 static volatile int fio_abort;
50 static int exit_value;
51
52 struct io_log *agg_io_log[2];
53
54 #define TERMINATE_ALL           (-1)
55 #define JOB_START_TIMEOUT       (5 * 1000)
56
57 static inline void td_set_runstate(struct thread_data *td, int runstate)
58 {
59         td->runstate = runstate;
60 }
61
62 static void terminate_threads(int group_id, int forced_kill)
63 {
64         struct thread_data *td;
65         int i;
66
67         for_each_td(td, i) {
68                 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
69                         td->terminate = 1;
70                         td->start_delay = 0;
71                         if (forced_kill)
72                                 td_set_runstate(td, TD_EXITED);
73                 }
74         }
75 }
76
77 static void sig_handler(int sig)
78 {
79         switch (sig) {
80                 case SIGALRM:
81                         update_io_ticks();
82                         disk_util_timer_arm();
83                         print_thread_status();
84                         break;
85                 default:
86                         printf("\nfio: terminating on signal %d\n", sig);
87                         fflush(stdout);
88                         terminate_threads(TERMINATE_ALL, 0);
89                         break;
90         }
91 }
92
93 /*
94  * Check if we are above the minimum rate given.
95  */
96 static int check_min_rate(struct thread_data *td, struct timeval *now)
97 {
98         unsigned long spent;
99         unsigned long rate;
100         int ddir = td->ddir;
101
102         /*
103          * allow a 2 second settle period in the beginning
104          */
105         if (mtime_since(&td->start, now) < 2000)
106                 return 0;
107
108         /*
109          * if rate blocks is set, sample is running
110          */
111         if (td->rate_bytes) {
112                 spent = mtime_since(&td->lastrate, now);
113                 if (spent < td->ratecycle)
114                         return 0;
115
116                 rate = (td->this_io_bytes[ddir] - td->rate_bytes) / spent;
117                 if (rate < td->ratemin) {
118                         fprintf(f_out, "%s: min rate %u not met, got %luKiB/sec\n", td->name, td->ratemin, rate);
119                         return 1;
120                 }
121         }
122
123         td->rate_bytes = td->this_io_bytes[ddir];
124         memcpy(&td->lastrate, now, sizeof(*now));
125         return 0;
126 }
127
128 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
129 {
130         if (!td->timeout)
131                 return 0;
132         if (mtime_since(&td->epoch, t) >= td->timeout * 1000)
133                 return 1;
134
135         return 0;
136 }
137
138 static struct fio_file *get_next_file(struct thread_data *td)
139 {
140         unsigned int old_next_file = td->next_file;
141         struct fio_file *f;
142
143         do {
144                 f = &td->files[td->next_file];
145
146                 td->next_file++;
147                 if (td->next_file >= td->nr_files)
148                         td->next_file = 0;
149
150                 if (f->fd != -1)
151                         break;
152
153                 f = NULL;
154         } while (td->next_file != old_next_file);
155
156         return f;
157 }
158
159 /*
160  * When job exits, we can cancel the in-flight IO if we are using async
161  * io. Attempt to do so.
162  */
163 static void cleanup_pending_aio(struct thread_data *td)
164 {
165         struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
166         struct list_head *entry, *n;
167         struct io_completion_data icd;
168         struct io_u *io_u;
169         int r;
170
171         /*
172          * get immediately available events, if any
173          */
174         r = td_io_getevents(td, 0, td->cur_depth, &ts);
175         if (r > 0) {
176                 icd.nr = r;
177                 ios_completed(td, &icd);
178         }
179
180         /*
181          * now cancel remaining active events
182          */
183         if (td->io_ops->cancel) {
184                 list_for_each_safe(entry, n, &td->io_u_busylist) {
185                         io_u = list_entry(entry, struct io_u, list);
186
187                         r = td->io_ops->cancel(td, io_u);
188                         if (!r)
189                                 put_io_u(td, io_u);
190                 }
191         }
192
193         if (td->cur_depth) {
194                 r = td_io_getevents(td, td->cur_depth, td->cur_depth, NULL);
195                 if (r > 0) {
196                         icd.nr = r;
197                         ios_completed(td, &icd);
198                 }
199         }
200 }
201
202 /*
203  * Helper to handle the final sync of a file. Works just like the normal
204  * io path, just does everything sync.
205  */
206 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
207 {
208         struct io_u *io_u = __get_io_u(td);
209         struct io_completion_data icd;
210         int ret;
211
212         if (!io_u)
213                 return 1;
214
215         io_u->ddir = DDIR_SYNC;
216         io_u->file = f;
217
218         if (td_io_prep(td, io_u)) {
219                 put_io_u(td, io_u);
220                 return 1;
221         }
222
223         ret = td_io_queue(td, io_u);
224         if (ret) {
225                 td_verror(td, io_u->error);
226                 put_io_u(td, io_u);
227                 return 1;
228         }
229
230         ret = td_io_getevents(td, 1, td->cur_depth, NULL);
231         if (ret < 0) {
232                 td_verror(td, ret);
233                 return 1;
234         }
235
236         icd.nr = ret;
237         ios_completed(td, &icd);
238         if (icd.error) {
239                 td_verror(td, icd.error);
240                 return 1;
241         }
242
243         return 0;
244 }
245
246 /*
247  * The main verify engine. Runs over the writes we previusly submitted,
248  * reads the blocks back in, and checks the crc/md5 of the data.
249  */
250 static void do_verify(struct thread_data *td)
251 {
252         struct io_u *io_u, *v_io_u = NULL;
253         struct io_completion_data icd;
254         struct fio_file *f;
255         int ret, i;
256
257         /*
258          * sync io first and invalidate cache, to make sure we really
259          * read from disk.
260          */
261         for_each_file(td, f, i) {
262                 fio_io_sync(td, f);
263                 file_invalidate_cache(td, f);
264         }
265
266         td_set_runstate(td, TD_VERIFYING);
267
268         do {
269                 if (td->terminate)
270                         break;
271
272                 io_u = __get_io_u(td);
273                 if (!io_u)
274                         break;
275
276                 if (runtime_exceeded(td, &io_u->start_time)) {
277                         put_io_u(td, io_u);
278                         break;
279                 }
280
281                 if (get_next_verify(td, io_u)) {
282                         put_io_u(td, io_u);
283                         break;
284                 }
285
286                 f = get_next_file(td);
287                 if (!f)
288                         break;
289
290                 io_u->file = f;
291
292                 if (td_io_prep(td, io_u)) {
293                         put_io_u(td, io_u);
294                         break;
295                 }
296
297                 ret = td_io_queue(td, io_u);
298                 if (ret) {
299                         td_verror(td, io_u->error);
300                         put_io_u(td, io_u);
301                         break;
302                 }
303
304                 /*
305                  * we have one pending to verify, do that while
306                  * we are doing io on the next one
307                  */
308                 if (do_io_u_verify(td, &v_io_u))
309                         break;
310
311                 ret = td_io_getevents(td, 1, 1, NULL);
312                 if (ret != 1) {
313                         if (ret < 0)
314                                 td_verror(td, ret);
315                         break;
316                 }
317
318                 v_io_u = td->io_ops->event(td, 0);
319                 icd.nr = 1;
320                 icd.error = 0;
321                 fio_gettime(&icd.time, NULL);
322                 io_completed(td, v_io_u, &icd);
323
324                 if (icd.error) {
325                         td_verror(td, icd.error);
326                         put_io_u(td, v_io_u);
327                         v_io_u = NULL;
328                         break;
329                 }
330
331                 /*
332                  * if we can't submit more io, we need to verify now
333                  */
334                 if (queue_full(td) && do_io_u_verify(td, &v_io_u))
335                         break;
336
337         } while (1);
338
339         do_io_u_verify(td, &v_io_u);
340
341         if (td->cur_depth)
342                 cleanup_pending_aio(td);
343
344         td_set_runstate(td, TD_RUNNING);
345 }
346
347 /*
348  * Not really an io thread, all it does is burn CPU cycles in the specified
349  * manner.
350  */
351 static void do_cpuio(struct thread_data *td)
352 {
353         struct timeval e;
354         int split = 100 / td->cpuload;
355         int i = 0;
356
357         while (!td->terminate) {
358                 fio_gettime(&e, NULL);
359
360                 if (runtime_exceeded(td, &e))
361                         break;
362
363                 if (!(i % split))
364                         __usec_sleep(10000);
365                 else
366                         usec_sleep(td, 10000);
367
368                 i++;
369         }
370 }
371
372 /*
373  * Main IO worker function. It retrieves io_u's to process and queues
374  * and reaps them, checking for rate and errors along the way.
375  */
376 static void do_io(struct thread_data *td)
377 {
378         struct io_completion_data icd;
379         struct timeval s;
380         unsigned long usec;
381         struct fio_file *f;
382         int i, ret = 0;
383
384         td_set_runstate(td, TD_RUNNING);
385
386         while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->io_size) {
387                 struct timespec *timeout;
388                 int min_evts = 0;
389                 struct io_u *io_u;
390
391                 if (td->terminate)
392                         break;
393
394                 f = get_next_file(td);
395                 if (!f)
396                         break;
397
398                 io_u = get_io_u(td, f);
399                 if (!io_u)
400                         break;
401
402                 memcpy(&s, &io_u->start_time, sizeof(s));
403
404 requeue:
405                 ret = td_io_queue(td, io_u);
406                 if (ret) {
407                         if (ret > 0 && (io_u->xfer_buflen != io_u->resid) &&
408                             io_u->resid) {
409                                 /*
410                                  * short read/write. requeue.
411                                  */
412                                 io_u->xfer_buflen = io_u->resid;
413                                 io_u->xfer_buf += ret;
414                                 goto requeue;
415                         } else {
416                                 put_io_u(td, io_u);
417                                 break;
418                         }
419                 }
420
421                 add_slat_sample(td, io_u->ddir, mtime_since(&io_u->start_time, &io_u->issue_time));
422
423                 if (td->cur_depth < td->iodepth) {
424                         struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
425
426                         timeout = &ts;
427                         min_evts = 0;
428                 } else {
429                         timeout = NULL;
430                         min_evts = 1;
431                 }
432
433                 ret = td_io_getevents(td, min_evts, td->cur_depth, timeout);
434                 if (ret < 0) {
435                         td_verror(td, ret);
436                         break;
437                 } else if (!ret)
438                         continue;
439
440                 icd.nr = ret;
441                 ios_completed(td, &icd);
442                 if (icd.error) {
443                         td_verror(td, icd.error);
444                         break;
445                 }
446
447                 /*
448                  * the rate is batched for now, it should work for batches
449                  * of completions except the very first one which may look
450                  * a little bursty
451                  */
452                 usec = utime_since(&s, &icd.time);
453
454                 rate_throttle(td, usec, icd.bytes_done[td->ddir], td->ddir);
455
456                 if (check_min_rate(td, &icd.time)) {
457                         if (exitall_on_terminate)
458                                 terminate_threads(td->groupid, 0);
459                         td_verror(td, ENODATA);
460                         break;
461                 }
462
463                 if (runtime_exceeded(td, &icd.time))
464                         break;
465
466                 if (td->thinktime) {
467                         unsigned long long b;
468
469                         b = td->io_blocks[0] + td->io_blocks[1];
470                         if (!(b % td->thinktime_blocks))
471                                 usec_sleep(td, td->thinktime);
472                 }
473         }
474
475         if (!td->error) {
476                 if (td->cur_depth)
477                         cleanup_pending_aio(td);
478
479                 if (should_fsync(td) && td->end_fsync) {
480                         td_set_runstate(td, TD_FSYNCING);
481                         for_each_file(td, f, i)
482                                 fio_io_sync(td, f);
483                 }
484         }
485 }
486
487 static void cleanup_io_u(struct thread_data *td)
488 {
489         struct list_head *entry, *n;
490         struct io_u *io_u;
491
492         list_for_each_safe(entry, n, &td->io_u_freelist) {
493                 io_u = list_entry(entry, struct io_u, list);
494
495                 list_del(&io_u->list);
496                 free(io_u);
497         }
498
499         free_io_mem(td);
500 }
501
502 /*
503  * "randomly" fill the buffer contents
504  */
505 static void fill_rand_buf(struct io_u *io_u, int max_bs)
506 {
507         int *ptr = io_u->buf;
508
509         while ((void *) ptr - io_u->buf < max_bs) {
510                 *ptr = rand() * 0x9e370001;
511                 ptr++;
512         }
513 }
514
515 static int init_io_u(struct thread_data *td)
516 {
517         struct io_u *io_u;
518         unsigned int max_bs;
519         int i, max_units;
520         char *p;
521
522         if (td->io_ops->flags & FIO_CPUIO)
523                 return 0;
524
525         if (td->io_ops->flags & FIO_SYNCIO)
526                 max_units = 1;
527         else
528                 max_units = td->iodepth;
529
530         max_bs = max(td->max_bs[DDIR_READ], td->max_bs[DDIR_WRITE]);
531         td->orig_buffer_size = max_bs * max_units;
532
533         if (td->mem_type == MEM_SHMHUGE || td->mem_type == MEM_MMAPHUGE)
534                 td->orig_buffer_size = (td->orig_buffer_size + td->hugepage_size - 1) & ~(td->hugepage_size - 1);
535         else
536                 td->orig_buffer_size += page_mask;
537
538         if (allocate_io_mem(td))
539                 return 1;
540
541         p = ALIGN(td->orig_buffer);
542         for (i = 0; i < max_units; i++) {
543                 io_u = malloc(sizeof(*io_u));
544                 memset(io_u, 0, sizeof(*io_u));
545                 INIT_LIST_HEAD(&io_u->list);
546
547                 io_u->buf = p + max_bs * i;
548                 if (td_write(td) || td_rw(td))
549                         fill_rand_buf(io_u, max_bs);
550
551                 io_u->index = i;
552                 list_add(&io_u->list, &td->io_u_freelist);
553         }
554
555         return 0;
556 }
557
558 static int switch_ioscheduler(struct thread_data *td)
559 {
560         char tmp[256], tmp2[128];
561         FILE *f;
562         int ret;
563
564         if (td->io_ops->flags & FIO_CPUIO)
565                 return 0;
566
567         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
568
569         f = fopen(tmp, "r+");
570         if (!f) {
571                 td_verror(td, errno);
572                 return 1;
573         }
574
575         /*
576          * Set io scheduler.
577          */
578         ret = fwrite(td->ioscheduler, strlen(td->ioscheduler), 1, f);
579         if (ferror(f) || ret != 1) {
580                 td_verror(td, errno);
581                 fclose(f);
582                 return 1;
583         }
584
585         rewind(f);
586
587         /*
588          * Read back and check that the selected scheduler is now the default.
589          */
590         ret = fread(tmp, 1, sizeof(tmp), f);
591         if (ferror(f) || ret < 0) {
592                 td_verror(td, errno);
593                 fclose(f);
594                 return 1;
595         }
596
597         sprintf(tmp2, "[%s]", td->ioscheduler);
598         if (!strstr(tmp, tmp2)) {
599                 log_err("fio: io scheduler %s not found\n", td->ioscheduler);
600                 td_verror(td, EINVAL);
601                 fclose(f);
602                 return 1;
603         }
604
605         fclose(f);
606         return 0;
607 }
608
609 static void clear_io_state(struct thread_data *td)
610 {
611         struct fio_file *f;
612         int i;
613
614         td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0;
615         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
616         td->zone_bytes = 0;
617
618         for_each_file(td, f, i) {
619                 f->last_pos = 0;
620                 if (td->io_ops->flags & FIO_SYNCIO)
621                         lseek(f->fd, SEEK_SET, 0);
622
623                 if (f->file_map)
624                         memset(f->file_map, 0, f->num_maps * sizeof(long));
625         }
626 }
627
628 /*
629  * Entry point for the thread based jobs. The process based jobs end up
630  * here as well, after a little setup.
631  */
632 static void *thread_main(void *data)
633 {
634         unsigned long long runtime[2];
635         struct thread_data *td = data;
636
637         if (!td->use_thread)
638                 setsid();
639
640         td->pid = getpid();
641
642         INIT_LIST_HEAD(&td->io_u_freelist);
643         INIT_LIST_HEAD(&td->io_u_busylist);
644         INIT_LIST_HEAD(&td->io_hist_list);
645         INIT_LIST_HEAD(&td->io_log_list);
646
647         if (init_io_u(td))
648                 goto err;
649
650         if (fio_setaffinity(td) == -1) {
651                 td_verror(td, errno);
652                 goto err;
653         }
654
655         if (init_iolog(td))
656                 goto err;
657
658         if (td->ioprio) {
659                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
660                         td_verror(td, errno);
661                         goto err;
662                 }
663         }
664
665         if (nice(td->nice) == -1) {
666                 td_verror(td, errno);
667                 goto err;
668         }
669
670         if (init_random_state(td))
671                 goto err;
672
673         if (td->ioscheduler && switch_ioscheduler(td))
674                 goto err;
675
676         td_set_runstate(td, TD_INITIALIZED);
677         fio_sem_up(&startup_sem);
678         fio_sem_down(&td->mutex);
679
680         if (!td->create_serialize && setup_files(td))
681                 goto err;
682         if (open_files(td))
683                 goto err;
684
685         /*
686          * Do this late, as some IO engines would like to have the
687          * files setup prior to initializing structures.
688          */
689         if (td_io_init(td))
690                 goto err;
691
692         if (td->exec_prerun) {
693                 if (system(td->exec_prerun) < 0)
694                         goto err;
695         }
696
697         fio_gettime(&td->epoch, NULL);
698         getrusage(RUSAGE_SELF, &td->ru_start);
699
700         runtime[0] = runtime[1] = 0;
701         while (td->loops--) {
702                 fio_gettime(&td->start, NULL);
703                 memcpy(&td->stat_sample_time, &td->start, sizeof(td->start));
704
705                 if (td->ratemin)
706                         memcpy(&td->lastrate, &td->stat_sample_time, sizeof(td->lastrate));
707
708                 clear_io_state(td);
709                 prune_io_piece_log(td);
710
711                 if (td->io_ops->flags & FIO_CPUIO)
712                         do_cpuio(td);
713                 else
714                         do_io(td);
715
716                 runtime[td->ddir] += utime_since_now(&td->start);
717                 if (td_rw(td) && td->io_bytes[td->ddir ^ 1])
718                         runtime[td->ddir ^ 1] = runtime[td->ddir];
719
720                 if (td->error || td->terminate)
721                         break;
722
723                 if (td->verify == VERIFY_NONE)
724                         continue;
725
726                 clear_io_state(td);
727                 fio_gettime(&td->start, NULL);
728
729                 do_verify(td);
730
731                 runtime[DDIR_READ] += utime_since_now(&td->start);
732
733                 if (td->error || td->terminate)
734                         break;
735         }
736
737         update_rusage_stat(td);
738         fio_gettime(&td->end_time, NULL);
739         td->runtime[0] = runtime[0] / 1000;
740         td->runtime[1] = runtime[1] / 1000;
741
742         if (td->bw_log)
743                 finish_log(td, td->bw_log, "bw");
744         if (td->slat_log)
745                 finish_log(td, td->slat_log, "slat");
746         if (td->clat_log)
747                 finish_log(td, td->clat_log, "clat");
748         if (td->write_iolog_file)
749                 write_iolog_close(td);
750         if (td->exec_postrun) {
751                 if (system(td->exec_postrun) < 0)
752                         log_err("fio: postrun %s failed\n", td->exec_postrun);
753         }
754
755         if (exitall_on_terminate)
756                 terminate_threads(td->groupid, 0);
757
758 err:
759         close_files(td);
760         close_ioengine(td);
761         cleanup_io_u(td);
762         td_set_runstate(td, TD_EXITED);
763         return NULL;
764
765 }
766
767 /*
768  * We cannot pass the td data into a forked process, so attach the td and
769  * pass it to the thread worker.
770  */
771 static void *fork_main(int shmid, int offset)
772 {
773         struct thread_data *td;
774         void *data;
775
776         data = shmat(shmid, NULL, 0);
777         if (data == (void *) -1) {
778                 perror("shmat");
779                 return NULL;
780         }
781
782         td = data + offset * sizeof(struct thread_data);
783         thread_main(td);
784         shmdt(data);
785         return NULL;
786 }
787
788 /*
789  * Run over the job map and reap the threads that have exited, if any.
790  */
791 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
792 {
793         struct thread_data *td;
794         int i, cputhreads, pending;
795
796         /*
797          * reap exited threads (TD_EXITED -> TD_REAPED)
798          */
799         pending = cputhreads = 0;
800         for_each_td(td, i) {
801                 /*
802                  * ->io_ops is NULL for a thread that has closed its
803                  * io engine
804                  */
805                 if (td->io_ops && td->io_ops->flags & FIO_CPUIO)
806                         cputhreads++;
807
808                 if (td->runstate != TD_EXITED) {
809                         if (td->runstate < TD_RUNNING)
810                                 pending++;
811
812                         continue;
813                 }
814
815                 if (td->error)
816                         exit_value++;
817
818                 td_set_runstate(td, TD_REAPED);
819
820                 if (td->use_thread) {
821                         long ret;
822
823                         if (pthread_join(td->thread, (void *) &ret))
824                                 perror("thread_join");
825                 } else
826                         waitpid(td->pid, NULL, 0);
827
828                 (*nr_running)--;
829                 (*m_rate) -= td->ratemin;
830                 (*t_rate) -= td->rate;
831         }
832
833         if (*nr_running == cputhreads && !pending)
834                 terminate_threads(TERMINATE_ALL, 0);
835 }
836
837 /*
838  * Main function for kicking off and reaping jobs, as needed.
839  */
840 static void run_threads(void)
841 {
842         struct thread_data *td;
843         unsigned long spent;
844         int i, todo, nr_running, m_rate, t_rate, nr_started;
845
846         if (fio_pin_memory())
847                 return;
848
849         if (!terse_output) {
850                 printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
851                 fflush(stdout);
852         }
853
854         signal(SIGINT, sig_handler);
855         signal(SIGALRM, sig_handler);
856
857         todo = thread_number;
858         nr_running = 0;
859         nr_started = 0;
860         m_rate = t_rate = 0;
861
862         for_each_td(td, i) {
863                 print_status_init(td->thread_number - 1);
864
865                 if (!td->create_serialize) {
866                         init_disk_util(td);
867                         continue;
868                 }
869
870                 /*
871                  * do file setup here so it happens sequentially,
872                  * we don't want X number of threads getting their
873                  * client data interspersed on disk
874                  */
875                 if (setup_files(td)) {
876                         td_set_runstate(td, TD_REAPED);
877                         todo--;
878                 }
879
880                 init_disk_util(td);
881         }
882
883         while (todo) {
884                 struct thread_data *map[MAX_JOBS];
885                 struct timeval this_start;
886                 int this_jobs = 0, left;
887
888                 /*
889                  * create threads (TD_NOT_CREATED -> TD_CREATED)
890                  */
891                 for_each_td(td, i) {
892                         if (td->runstate != TD_NOT_CREATED)
893                                 continue;
894
895                         /*
896                          * never got a chance to start, killed by other
897                          * thread for some reason
898                          */
899                         if (td->terminate) {
900                                 todo--;
901                                 continue;
902                         }
903
904                         if (td->start_delay) {
905                                 spent = mtime_since_genesis();
906
907                                 if (td->start_delay * 1000 > spent)
908                                         continue;
909                         }
910
911                         if (td->stonewall && (nr_started || nr_running))
912                                 break;
913
914                         /*
915                          * Set state to created. Thread will transition
916                          * to TD_INITIALIZED when it's done setting up.
917                          */
918                         td_set_runstate(td, TD_CREATED);
919                         map[this_jobs++] = td;
920                         fio_sem_init(&startup_sem, 1);
921                         nr_started++;
922
923                         if (td->use_thread) {
924                                 if (pthread_create(&td->thread, NULL, thread_main, td)) {
925                                         perror("thread_create");
926                                         nr_started--;
927                                 }
928                         } else {
929                                 if (fork())
930                                         fio_sem_down(&startup_sem);
931                                 else {
932                                         fork_main(shm_id, i);
933                                         exit(0);
934                                 }
935                         }
936                 }
937
938                 /*
939                  * Wait for the started threads to transition to
940                  * TD_INITIALIZED.
941                  */
942                 fio_gettime(&this_start, NULL);
943                 left = this_jobs;
944                 while (left && !fio_abort) {
945                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
946                                 break;
947
948                         usleep(100000);
949
950                         for (i = 0; i < this_jobs; i++) {
951                                 td = map[i];
952                                 if (!td)
953                                         continue;
954                                 if (td->runstate == TD_INITIALIZED) {
955                                         map[i] = NULL;
956                                         left--;
957                                 } else if (td->runstate >= TD_EXITED) {
958                                         map[i] = NULL;
959                                         left--;
960                                         todo--;
961                                         nr_running++; /* work-around... */
962                                 }
963                         }
964                 }
965
966                 if (left) {
967                         log_err("fio: %d jobs failed to start\n", left);
968                         for (i = 0; i < this_jobs; i++) {
969                                 td = map[i];
970                                 if (!td)
971                                         continue;
972                                 kill(td->pid, SIGTERM);
973                         }
974                         break;
975                 }
976
977                 /*
978                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
979                  */
980                 for_each_td(td, i) {
981                         if (td->runstate != TD_INITIALIZED)
982                                 continue;
983
984                         td_set_runstate(td, TD_RUNNING);
985                         nr_running++;
986                         nr_started--;
987                         m_rate += td->ratemin;
988                         t_rate += td->rate;
989                         todo--;
990                         fio_sem_up(&td->mutex);
991                 }
992
993                 reap_threads(&nr_running, &t_rate, &m_rate);
994
995                 if (todo)
996                         usleep(100000);
997         }
998
999         while (nr_running) {
1000                 reap_threads(&nr_running, &t_rate, &m_rate);
1001                 usleep(10000);
1002         }
1003
1004         update_io_ticks();
1005         fio_unpin_memory();
1006 }
1007
1008 int main(int argc, char *argv[])
1009 {
1010         long ps;
1011
1012         /*
1013          * We need locale for number printing, if it isn't set then just
1014          * go with the US format.
1015          */
1016         if (!getenv("LC_NUMERIC"))
1017                 setlocale(LC_NUMERIC, "en_US");
1018
1019         if (parse_options(argc, argv))
1020                 return 1;
1021
1022         if (!thread_number) {
1023                 log_err("Nothing to do\n");
1024                 return 1;
1025         }
1026
1027         ps = sysconf(_SC_PAGESIZE);
1028         if (ps < 0) {
1029                 log_err("Failed to get page size\n");
1030                 return 1;
1031         }
1032
1033         page_mask = ps - 1;
1034
1035         if (write_bw_log) {
1036                 setup_log(&agg_io_log[DDIR_READ]);
1037                 setup_log(&agg_io_log[DDIR_WRITE]);
1038         }
1039
1040         disk_util_timer_arm();
1041
1042         run_threads();
1043
1044         if (!fio_abort) {
1045                 show_run_stats();
1046                 if (write_bw_log) {
1047                         __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1048                         __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");
1049                 }
1050         }
1051
1052         return exit_value;
1053 }