[PATCH] Make the libaio fsync fallback really work
[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 <assert.h>
30 #include <sys/stat.h>
31 #include <sys/wait.h>
32 #include <sys/ipc.h>
33 #include <sys/shm.h>
34 #include <sys/ioctl.h>
35 #include <sys/mman.h>
36
37 #include "fio.h"
38 #include "os.h"
39
40 static unsigned long page_mask;
41 #define ALIGN(buf)      \
42         (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
43
44 int groupid = 0;
45 int thread_number = 0;
46 int shm_id = 0;
47 int temp_stall_ts;
48
49 static volatile int startup_sem;
50 static volatile int fio_abort;
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[td->ddir] < 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                                 td_verror(td, io_u->error);
417                                 put_io_u(td, io_u);
418                                 break;
419                         }
420                 }
421
422                 add_slat_sample(td, io_u->ddir, mtime_since(&io_u->start_time, &io_u->issue_time));
423
424                 if (td->cur_depth < td->iodepth) {
425                         struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
426
427                         timeout = &ts;
428                         min_evts = 0;
429                 } else {
430                         timeout = NULL;
431                         min_evts = 1;
432                 }
433
434                 ret = td_io_getevents(td, min_evts, td->cur_depth, timeout);
435                 if (ret < 0) {
436                         td_verror(td, ret);
437                         break;
438                 } else if (!ret)
439                         continue;
440
441                 icd.nr = ret;
442                 ios_completed(td, &icd);
443                 if (icd.error) {
444                         td_verror(td, icd.error);
445                         break;
446                 }
447
448                 /*
449                  * the rate is batched for now, it should work for batches
450                  * of completions except the very first one which may look
451                  * a little bursty
452                  */
453                 usec = utime_since(&s, &icd.time);
454
455                 rate_throttle(td, usec, icd.bytes_done[td->ddir], td->ddir);
456
457                 if (check_min_rate(td, &icd.time)) {
458                         if (exitall_on_terminate)
459                                 terminate_threads(td->groupid, 0);
460                         td_verror(td, ENODATA);
461                         break;
462                 }
463
464                 if (runtime_exceeded(td, &icd.time))
465                         break;
466
467                 if (td->thinktime) {
468                         unsigned long long b;
469
470                         b = td->io_blocks[0] + td->io_blocks[1];
471                         if (!(b % td->thinktime_blocks))
472                                 usec_sleep(td, td->thinktime);
473                 }
474         }
475
476         if (!td->error) {
477                 if (td->cur_depth)
478                         cleanup_pending_aio(td);
479
480                 if (should_fsync(td) && td->end_fsync) {
481                         td_set_runstate(td, TD_FSYNCING);
482                         for_each_file(td, f, i)
483                                 fio_io_sync(td, f);
484                 }
485         }
486 }
487
488 static void cleanup_io_u(struct thread_data *td)
489 {
490         struct list_head *entry, *n;
491         struct io_u *io_u;
492
493         list_for_each_safe(entry, n, &td->io_u_freelist) {
494                 io_u = list_entry(entry, struct io_u, list);
495
496                 list_del(&io_u->list);
497                 free(io_u);
498         }
499
500         free_io_mem(td);
501 }
502
503 /*
504  * "randomly" fill the buffer contents
505  */
506 static void fill_rand_buf(struct io_u *io_u, int max_bs)
507 {
508         int *ptr = io_u->buf;
509
510         while ((void *) ptr - io_u->buf < max_bs) {
511                 *ptr = rand() * 0x9e370001;
512                 ptr++;
513         }
514 }
515
516 static int init_io_u(struct thread_data *td)
517 {
518         struct io_u *io_u;
519         unsigned int max_bs;
520         int i, max_units;
521         char *p;
522
523         if (td->io_ops->flags & FIO_CPUIO)
524                 return 0;
525
526         if (td->io_ops->flags & FIO_SYNCIO)
527                 max_units = 1;
528         else
529                 max_units = td->iodepth;
530
531         max_bs = max(td->max_bs[DDIR_READ], td->max_bs[DDIR_WRITE]);
532         td->orig_buffer_size = max_bs * max_units;
533
534         if (td->mem_type == MEM_SHMHUGE || td->mem_type == MEM_MMAPHUGE)
535                 td->orig_buffer_size = (td->orig_buffer_size + td->hugepage_size - 1) & ~(td->hugepage_size - 1);
536         else
537                 td->orig_buffer_size += page_mask;
538
539         if (allocate_io_mem(td))
540                 return 1;
541
542         p = ALIGN(td->orig_buffer);
543         for (i = 0; i < max_units; i++) {
544                 io_u = malloc(sizeof(*io_u));
545                 memset(io_u, 0, sizeof(*io_u));
546                 INIT_LIST_HEAD(&io_u->list);
547
548                 io_u->buf = p + max_bs * i;
549                 if (td_write(td) || td_rw(td))
550                         fill_rand_buf(io_u, max_bs);
551
552                 io_u->index = i;
553                 list_add(&io_u->list, &td->io_u_freelist);
554         }
555
556         return 0;
557 }
558
559 static int switch_ioscheduler(struct thread_data *td)
560 {
561         char tmp[256], tmp2[128];
562         FILE *f;
563         int ret;
564
565         if (td->io_ops->flags & FIO_CPUIO)
566                 return 0;
567
568         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
569
570         f = fopen(tmp, "r+");
571         if (!f) {
572                 td_verror(td, errno);
573                 return 1;
574         }
575
576         /*
577          * Set io scheduler.
578          */
579         ret = fwrite(td->ioscheduler, strlen(td->ioscheduler), 1, f);
580         if (ferror(f) || ret != 1) {
581                 td_verror(td, errno);
582                 fclose(f);
583                 return 1;
584         }
585
586         rewind(f);
587
588         /*
589          * Read back and check that the selected scheduler is now the default.
590          */
591         ret = fread(tmp, 1, sizeof(tmp), f);
592         if (ferror(f) || ret < 0) {
593                 td_verror(td, errno);
594                 fclose(f);
595                 return 1;
596         }
597
598         sprintf(tmp2, "[%s]", td->ioscheduler);
599         if (!strstr(tmp, tmp2)) {
600                 log_err("fio: io scheduler %s not found\n", td->ioscheduler);
601                 td_verror(td, EINVAL);
602                 fclose(f);
603                 return 1;
604         }
605
606         fclose(f);
607         return 0;
608 }
609
610 static void clear_io_state(struct thread_data *td)
611 {
612         struct fio_file *f;
613         int i;
614
615         td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0;
616         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
617         td->zone_bytes = 0;
618
619         for_each_file(td, f, i) {
620                 f->last_pos = 0;
621                 if (td->io_ops->flags & FIO_SYNCIO)
622                         lseek(f->fd, SEEK_SET, 0);
623
624                 if (f->file_map)
625                         memset(f->file_map, 0, f->num_maps * sizeof(long));
626         }
627 }
628
629 /*
630  * Entry point for the thread based jobs. The process based jobs end up
631  * here as well, after a little setup.
632  */
633 static void *thread_main(void *data)
634 {
635         unsigned long long runtime[2];
636         struct thread_data *td = data;
637
638         if (!td->use_thread)
639                 setsid();
640
641         td->pid = getpid();
642
643         INIT_LIST_HEAD(&td->io_u_freelist);
644         INIT_LIST_HEAD(&td->io_u_busylist);
645         INIT_LIST_HEAD(&td->io_hist_list);
646         INIT_LIST_HEAD(&td->io_log_list);
647
648         if (init_io_u(td))
649                 goto err;
650
651         if (fio_setaffinity(td) == -1) {
652                 td_verror(td, errno);
653                 goto err;
654         }
655
656         if (init_iolog(td))
657                 goto err;
658
659         if (td->ioprio) {
660                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
661                         td_verror(td, errno);
662                         goto err;
663                 }
664         }
665
666         if (nice(td->nice) == -1) {
667                 td_verror(td, errno);
668                 goto err;
669         }
670
671         if (init_random_state(td))
672                 goto err;
673
674         if (td->ioscheduler && switch_ioscheduler(td))
675                 goto err;
676
677         td_set_runstate(td, TD_INITIALIZED);
678         fio_sem_up(&startup_sem);
679         fio_sem_down(&td->mutex);
680
681         if (!td->create_serialize && setup_files(td))
682                 goto err;
683         if (open_files(td))
684                 goto err;
685
686         /*
687          * Do this late, as some IO engines would like to have the
688          * files setup prior to initializing structures.
689          */
690         if (td_io_init(td))
691                 goto err;
692
693         if (td->exec_prerun) {
694                 if (system(td->exec_prerun) < 0)
695                         goto err;
696         }
697
698         fio_gettime(&td->epoch, NULL);
699         getrusage(RUSAGE_SELF, &td->ru_start);
700
701         runtime[0] = runtime[1] = 0;
702         while (td->loops--) {
703                 fio_gettime(&td->start, NULL);
704                 memcpy(&td->stat_sample_time, &td->start, sizeof(td->start));
705
706                 if (td->ratemin)
707                         memcpy(&td->lastrate, &td->stat_sample_time, sizeof(td->lastrate));
708
709                 clear_io_state(td);
710                 prune_io_piece_log(td);
711
712                 if (td->io_ops->flags & FIO_CPUIO)
713                         do_cpuio(td);
714                 else
715                         do_io(td);
716
717                 runtime[td->ddir] += utime_since_now(&td->start);
718                 if (td_rw(td) && td->io_bytes[td->ddir ^ 1])
719                         runtime[td->ddir ^ 1] = runtime[td->ddir];
720
721                 if (td->error || td->terminate)
722                         break;
723
724                 if (td->verify == VERIFY_NONE)
725                         continue;
726
727                 clear_io_state(td);
728                 fio_gettime(&td->start, NULL);
729
730                 do_verify(td);
731
732                 runtime[DDIR_READ] += utime_since_now(&td->start);
733
734                 if (td->error || td->terminate)
735                         break;
736         }
737
738         update_rusage_stat(td);
739         fio_gettime(&td->end_time, NULL);
740         td->runtime[0] = runtime[0] / 1000;
741         td->runtime[1] = runtime[1] / 1000;
742
743         if (td->bw_log)
744                 finish_log(td, td->bw_log, "bw");
745         if (td->slat_log)
746                 finish_log(td, td->slat_log, "slat");
747         if (td->clat_log)
748                 finish_log(td, td->clat_log, "clat");
749         if (td->write_iolog_file)
750                 write_iolog_close(td);
751         if (td->exec_postrun) {
752                 if (system(td->exec_postrun) < 0)
753                         log_err("fio: postrun %s failed\n", td->exec_postrun);
754         }
755
756         if (exitall_on_terminate)
757                 terminate_threads(td->groupid, 0);
758
759 err:
760         close_files(td);
761         close_ioengine(td);
762         cleanup_io_u(td);
763         td_set_runstate(td, TD_EXITED);
764         return NULL;
765
766 }
767
768 /*
769  * We cannot pass the td data into a forked process, so attach the td and
770  * pass it to the thread worker.
771  */
772 static void *fork_main(int shmid, int offset)
773 {
774         struct thread_data *td;
775         void *data;
776
777         data = shmat(shmid, NULL, 0);
778         if (data == (void *) -1) {
779                 perror("shmat");
780                 return NULL;
781         }
782
783         td = data + offset * sizeof(struct thread_data);
784         thread_main(td);
785         shmdt(data);
786         return NULL;
787 }
788
789 /*
790  * Run over the job map and reap the threads that have exited, if any.
791  */
792 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
793 {
794         struct thread_data *td;
795         int i, cputhreads, pending;
796
797         /*
798          * reap exited threads (TD_EXITED -> TD_REAPED)
799          */
800         pending = cputhreads = 0;
801         for_each_td(td, i) {
802                 /*
803                  * ->io_ops is NULL for a thread that has closed its
804                  * io engine
805                  */
806                 if (td->io_ops && td->io_ops->flags & FIO_CPUIO)
807                         cputhreads++;
808
809                 if (td->runstate != TD_EXITED) {
810                         if (td->runstate < TD_RUNNING)
811                                 pending++;
812
813                         continue;
814                 }
815
816                 td_set_runstate(td, TD_REAPED);
817
818                 if (td->use_thread) {
819                         long ret;
820
821                         if (pthread_join(td->thread, (void *) &ret))
822                                 perror("thread_join");
823                 } else
824                         waitpid(td->pid, NULL, 0);
825
826                 (*nr_running)--;
827                 (*m_rate) -= td->ratemin;
828                 (*t_rate) -= td->rate;
829         }
830
831         if (*nr_running == cputhreads && !pending)
832                 terminate_threads(TERMINATE_ALL, 0);
833 }
834
835 /*
836  * Main function for kicking off and reaping jobs, as needed.
837  */
838 static void run_threads(void)
839 {
840         struct thread_data *td;
841         unsigned long spent;
842         int i, todo, nr_running, m_rate, t_rate, nr_started;
843
844         if (fio_pin_memory())
845                 return;
846
847         if (!terse_output) {
848                 printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
849                 fflush(stdout);
850         }
851
852         signal(SIGINT, sig_handler);
853         signal(SIGALRM, sig_handler);
854
855         todo = thread_number;
856         nr_running = 0;
857         nr_started = 0;
858         m_rate = t_rate = 0;
859
860         for_each_td(td, i) {
861                 print_status_init(td->thread_number - 1);
862
863                 if (!td->create_serialize) {
864                         init_disk_util(td);
865                         continue;
866                 }
867
868                 /*
869                  * do file setup here so it happens sequentially,
870                  * we don't want X number of threads getting their
871                  * client data interspersed on disk
872                  */
873                 if (setup_files(td)) {
874                         td_set_runstate(td, TD_REAPED);
875                         todo--;
876                 }
877
878                 init_disk_util(td);
879         }
880
881         while (todo) {
882                 struct thread_data *map[MAX_JOBS];
883                 struct timeval this_start;
884                 int this_jobs = 0, left;
885
886                 /*
887                  * create threads (TD_NOT_CREATED -> TD_CREATED)
888                  */
889                 for_each_td(td, i) {
890                         if (td->runstate != TD_NOT_CREATED)
891                                 continue;
892
893                         /*
894                          * never got a chance to start, killed by other
895                          * thread for some reason
896                          */
897                         if (td->terminate) {
898                                 todo--;
899                                 continue;
900                         }
901
902                         if (td->start_delay) {
903                                 spent = mtime_since_genesis();
904
905                                 if (td->start_delay * 1000 > spent)
906                                         continue;
907                         }
908
909                         if (td->stonewall && (nr_started || nr_running))
910                                 break;
911
912                         /*
913                          * Set state to created. Thread will transition
914                          * to TD_INITIALIZED when it's done setting up.
915                          */
916                         td_set_runstate(td, TD_CREATED);
917                         map[this_jobs++] = td;
918                         fio_sem_init(&startup_sem, 1);
919                         nr_started++;
920
921                         if (td->use_thread) {
922                                 if (pthread_create(&td->thread, NULL, thread_main, td)) {
923                                         perror("thread_create");
924                                         nr_started--;
925                                 }
926                         } else {
927                                 if (fork())
928                                         fio_sem_down(&startup_sem);
929                                 else {
930                                         fork_main(shm_id, i);
931                                         exit(0);
932                                 }
933                         }
934                 }
935
936                 /*
937                  * Wait for the started threads to transition to
938                  * TD_INITIALIZED.
939                  */
940                 fio_gettime(&this_start, NULL);
941                 left = this_jobs;
942                 while (left && !fio_abort) {
943                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
944                                 break;
945
946                         usleep(100000);
947
948                         for (i = 0; i < this_jobs; i++) {
949                                 td = map[i];
950                                 if (!td)
951                                         continue;
952                                 if (td->runstate == TD_INITIALIZED) {
953                                         map[i] = NULL;
954                                         left--;
955                                 } else if (td->runstate >= TD_EXITED) {
956                                         map[i] = NULL;
957                                         left--;
958                                         todo--;
959                                         nr_running++; /* work-around... */
960                                 }
961                         }
962                 }
963
964                 if (left) {
965                         log_err("fio: %d jobs failed to start\n", left);
966                         for (i = 0; i < this_jobs; i++) {
967                                 td = map[i];
968                                 if (!td)
969                                         continue;
970                                 kill(td->pid, SIGTERM);
971                         }
972                         break;
973                 }
974
975                 /*
976                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
977                  */
978                 for_each_td(td, i) {
979                         if (td->runstate != TD_INITIALIZED)
980                                 continue;
981
982                         td_set_runstate(td, TD_RUNNING);
983                         nr_running++;
984                         nr_started--;
985                         m_rate += td->ratemin;
986                         t_rate += td->rate;
987                         todo--;
988                         fio_sem_up(&td->mutex);
989                 }
990
991                 reap_threads(&nr_running, &t_rate, &m_rate);
992
993                 if (todo)
994                         usleep(100000);
995         }
996
997         while (nr_running) {
998                 reap_threads(&nr_running, &t_rate, &m_rate);
999                 usleep(10000);
1000         }
1001
1002         update_io_ticks();
1003         fio_unpin_memory();
1004 }
1005
1006 int main(int argc, char *argv[])
1007 {
1008         long ps;
1009
1010         if (parse_options(argc, argv))
1011                 return 1;
1012
1013         if (!thread_number) {
1014                 log_err("Nothing to do\n");
1015                 return 1;
1016         }
1017
1018         ps = sysconf(_SC_PAGESIZE);
1019         if (ps < 0) {
1020                 log_err("Failed to get page size\n");
1021                 return 1;
1022         }
1023
1024         page_mask = ps - 1;
1025
1026         if (write_bw_log) {
1027                 setup_log(&agg_io_log[DDIR_READ]);
1028                 setup_log(&agg_io_log[DDIR_WRITE]);
1029         }
1030
1031         disk_util_timer_arm();
1032
1033         run_threads();
1034
1035         if (!fio_abort) {
1036                 show_run_stats();
1037                 if (write_bw_log) {
1038                         __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1039                         __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");
1040                 }
1041         }
1042
1043         return 0;
1044 }