Document file_service_type option
[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 <assert.h>
31 #include <sys/stat.h>
32 #include <sys/wait.h>
33 #include <sys/ipc.h>
34 #include <sys/shm.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 static int exit_value;
52
53 struct io_log *agg_io_log[2];
54
55 #define TERMINATE_ALL           (-1)
56 #define JOB_START_TIMEOUT       (5 * 1000)
57
58 static inline void td_set_runstate(struct thread_data *td, int runstate)
59 {
60         td->runstate = runstate;
61 }
62
63 static void terminate_threads(int group_id, int forced_kill)
64 {
65         struct thread_data *td;
66         int i;
67
68         for_each_td(td, i) {
69                 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
70                         td->terminate = 1;
71                         td->start_delay = 0;
72                         if (forced_kill)
73                                 td_set_runstate(td, TD_EXITED);
74                 }
75         }
76 }
77
78 static void sig_handler(int sig)
79 {
80         switch (sig) {
81                 case SIGALRM:
82                         update_io_ticks();
83                         disk_util_timer_arm();
84                         print_thread_status();
85                         break;
86                 default:
87                         printf("\nfio: terminating on signal %d\n", sig);
88                         fflush(stdout);
89                         terminate_threads(TERMINATE_ALL, 0);
90                         break;
91         }
92 }
93
94 /*
95  * Check if we are above the minimum rate given.
96  */
97 static int check_min_rate(struct thread_data *td, struct timeval *now)
98 {
99         unsigned long long bytes = 0;
100         unsigned long spent;
101         unsigned long rate;
102
103         /*
104          * No minimum rate set, always ok
105          */
106         if (!td->ratemin)
107                 return 0;
108
109         /*
110          * allow a 2 second settle period in the beginning
111          */
112         if (mtime_since(&td->start, now) < 2000)
113                 return 0;
114
115         if (td_read(td))
116                 bytes += td->this_io_bytes[DDIR_READ];
117         if (td_write(td))
118                 bytes += td->this_io_bytes[DDIR_WRITE];
119
120         /*
121          * if rate blocks is set, sample is running
122          */
123         if (td->rate_bytes) {
124                 spent = mtime_since(&td->lastrate, now);
125                 if (spent < td->ratecycle)
126                         return 0;
127
128                 if (bytes < td->rate_bytes) {
129                         fprintf(f_out, "%s: min rate %u not met\n", td->name, td->ratemin);
130                         return 1;
131                 } else {
132                         rate = (bytes - td->rate_bytes) / spent;
133                         if (rate < td->ratemin || bytes < td->rate_bytes) {
134                                 fprintf(f_out, "%s: min rate %u not met, got %luKiB/sec\n", td->name, td->ratemin, rate);
135                                 return 1;
136                         }
137                 }
138         }
139
140         td->rate_bytes = bytes;
141         memcpy(&td->lastrate, now, sizeof(*now));
142         return 0;
143 }
144
145 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
146 {
147         if (!td->timeout)
148                 return 0;
149         if (mtime_since(&td->epoch, t) >= td->timeout * 1000)
150                 return 1;
151
152         return 0;
153 }
154
155 /*
156  * When job exits, we can cancel the in-flight IO if we are using async
157  * io. Attempt to do so.
158  */
159 static void cleanup_pending_aio(struct thread_data *td)
160 {
161         struct list_head *entry, *n;
162         struct io_u *io_u;
163         int r;
164
165         /*
166          * get immediately available events, if any
167          */
168         r = io_u_queued_complete(td, 0);
169         if (r < 0)
170                 return;
171
172         /*
173          * now cancel remaining active events
174          */
175         if (td->io_ops->cancel) {
176                 list_for_each_safe(entry, n, &td->io_u_busylist) {
177                         io_u = list_entry(entry, struct io_u, list);
178
179                         /*
180                          * if the io_u isn't in flight, then that generally
181                          * means someone leaked an io_u. complain but fix
182                          * it up, so we don't stall here.
183                          */
184                         if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
185                                 log_err("fio: non-busy IO on busy list\n");
186                                 put_io_u(td, io_u);
187                         } else {
188                                 r = td->io_ops->cancel(td, io_u);
189                                 if (!r)
190                                         put_io_u(td, io_u);
191                         }
192                 }
193         }
194
195         if (td->cur_depth)
196                 r = io_u_queued_complete(td, td->cur_depth);
197 }
198
199 /*
200  * Helper to handle the final sync of a file. Works just like the normal
201  * io path, just does everything sync.
202  */
203 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
204 {
205         struct io_u *io_u = __get_io_u(td);
206         int ret;
207
208         if (!io_u)
209                 return 1;
210
211         io_u->ddir = DDIR_SYNC;
212         io_u->file = f;
213
214         if (td_io_prep(td, io_u)) {
215                 put_io_u(td, io_u);
216                 return 1;
217         }
218
219 requeue:
220         ret = td_io_queue(td, io_u);
221         if (ret < 0) {
222                 td_verror(td, io_u->error, "td_io_queue");
223                 put_io_u(td, io_u);
224                 return 1;
225         } else if (ret == FIO_Q_QUEUED) {
226                 if (io_u_queued_complete(td, 1) < 0)
227                         return 1;
228         } else if (ret == FIO_Q_COMPLETED) {
229                 if (io_u->error) {
230                         td_verror(td, io_u->error, "td_io_queue");
231                         return 1;
232                 }
233
234                 if (io_u_sync_complete(td, io_u) < 0)
235                         return 1;
236         } else if (ret == FIO_Q_BUSY) {
237                 if (td_io_commit(td))
238                         return 1;
239                 goto requeue;
240         }
241
242         return 0;
243 }
244
245 /*
246  * The main verify engine. Runs over the writes we previusly submitted,
247  * reads the blocks back in, and checks the crc/md5 of the data.
248  */
249 static void do_verify(struct thread_data *td)
250 {
251         struct fio_file *f;
252         struct io_u *io_u;
253         int ret, i, min_events;
254
255         /*
256          * sync io first and invalidate cache, to make sure we really
257          * read from disk.
258          */
259         for_each_file(td, f, i) {
260                 if (fio_io_sync(td, f))
261                         break;
262                 if (file_invalidate_cache(td, f))
263                         break;
264         }
265
266         if (td->error)
267                 return;
268
269         td_set_runstate(td, TD_VERIFYING);
270
271         io_u = NULL;
272         while (!td->terminate) {
273                 io_u = __get_io_u(td);
274                 if (!io_u)
275                         break;
276
277                 if (runtime_exceeded(td, &io_u->start_time)) {
278                         put_io_u(td, io_u);
279                         break;
280                 }
281
282                 if (get_next_verify(td, io_u)) {
283                         put_io_u(td, io_u);
284                         break;
285                 }
286
287                 if (td_io_prep(td, io_u)) {
288                         put_io_u(td, io_u);
289                         break;
290                 }
291
292                 io_u->end_io = verify_io_u;
293 requeue:
294                 ret = td_io_queue(td, io_u);
295
296                 switch (ret) {
297                 case FIO_Q_COMPLETED:
298                         if (io_u->error)
299                                 ret = -io_u->error;
300                         if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
301                                 int bytes = io_u->xfer_buflen - io_u->resid;
302
303                                 io_u->xfer_buflen = io_u->resid;
304                                 io_u->xfer_buf += bytes;
305                                 goto requeue;
306                         }
307                         ret = io_u_sync_complete(td, io_u);
308                         if (ret < 0)
309                                 break;
310                         continue;
311                 case FIO_Q_QUEUED:
312                         break;
313                 case FIO_Q_BUSY:
314                         requeue_io_u(td, &io_u);
315                         ret = td_io_commit(td);
316                         break;
317                 default:
318                         assert(ret < 0);
319                         td_verror(td, -ret, "td_io_queue");
320                         break;
321                 }
322
323                 if (ret < 0 || td->error)
324                         break;
325
326                 /*
327                  * if we can queue more, do so. but check if there are
328                  * completed io_u's first.
329                  */
330                 min_events = 0;
331                 if (queue_full(td) || ret == FIO_Q_BUSY) {
332                         min_events = 1;
333
334                         if (td->cur_depth > td->iodepth_low)
335                                 min_events = td->cur_depth - td->iodepth_low;
336                 }
337
338                 /*
339                  * Reap required number of io units, if any, and do the
340                  * verification on them through the callback handler
341                  */
342                 if (io_u_queued_complete(td, min_events) < 0)
343                         break;
344         }
345
346         if (!td->error) {
347                 min_events = td->cur_depth;
348
349                 if (min_events)
350                         ret = io_u_queued_complete(td, min_events);
351         } else
352                 cleanup_pending_aio(td);
353
354         td_set_runstate(td, TD_RUNNING);
355 }
356
357 /*
358  * Not really an io thread, all it does is burn CPU cycles in the specified
359  * manner.
360  */
361 static void do_cpuio(struct thread_data *td)
362 {
363         struct timeval e;
364         int split = 100 / td->cpuload;
365         int i = 0;
366
367         while (!td->terminate) {
368                 fio_gettime(&e, NULL);
369
370                 if (runtime_exceeded(td, &e))
371                         break;
372
373                 if (!(i % split))
374                         __usec_sleep(10000);
375                 else
376                         usec_sleep(td, 10000);
377
378                 i++;
379         }
380 }
381
382 /*
383  * Main IO worker function. It retrieves io_u's to process and queues
384  * and reaps them, checking for rate and errors along the way.
385  */
386 static void do_io(struct thread_data *td)
387 {
388         struct timeval s;
389         unsigned long usec;
390         int i, ret = 0;
391
392         td_set_runstate(td, TD_RUNNING);
393
394         while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->io_size) {
395                 struct timeval comp_time;
396                 long bytes_done = 0;
397                 int min_evts = 0;
398                 struct io_u *io_u;
399
400                 if (td->terminate)
401                         break;
402
403                 io_u = get_io_u(td);
404                 if (!io_u)
405                         break;
406
407                 memcpy(&s, &io_u->start_time, sizeof(s));
408
409                 if (runtime_exceeded(td, &s)) {
410                         put_io_u(td, io_u);
411                         break;
412                 }
413 requeue:
414                 ret = td_io_queue(td, io_u);
415
416                 switch (ret) {
417                 case FIO_Q_COMPLETED:
418                         if (io_u->error) {
419                                 ret = io_u->error;
420                                 break;
421                         }
422                         if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
423                                 int bytes = io_u->xfer_buflen - io_u->resid;
424
425                                 io_u->xfer_buflen = io_u->resid;
426                                 io_u->xfer_buf += bytes;
427                                 goto requeue;
428                         }
429                         fio_gettime(&comp_time, NULL);
430                         bytes_done = io_u_sync_complete(td, io_u);
431                         if (bytes_done < 0)
432                                 ret = bytes_done;
433                         break;
434                 case FIO_Q_QUEUED:
435                         /*
436                          * if the engine doesn't have a commit hook,
437                          * the io_u is really queued. if it does have such
438                          * a hook, it has to call io_u_queued() itself.
439                          */
440                         if (td->io_ops->commit == NULL)
441                                 io_u_queued(td, io_u);
442                         break;
443                 case FIO_Q_BUSY:
444                         requeue_io_u(td, &io_u);
445                         ret = td_io_commit(td);
446                         break;
447                 default:
448                         assert(ret < 0);
449                         put_io_u(td, io_u);
450                         break;
451                 }
452
453                 if (ret < 0 || td->error)
454                         break;
455
456                 /*
457                  * See if we need to complete some commands
458                  */
459                 if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
460                         min_evts = 0;
461                         if (queue_full(td) || ret == FIO_Q_BUSY) {
462                                 min_evts = 1;
463
464                                 if (td->cur_depth > td->iodepth_low)
465                                         min_evts = td->cur_depth - td->iodepth_low;
466                         }
467
468                         fio_gettime(&comp_time, NULL);
469                         bytes_done = io_u_queued_complete(td, min_evts);
470                         if (bytes_done < 0)
471                                 break;
472                 }
473
474                 if (!bytes_done)
475                         continue;
476
477                 /*
478                  * the rate is batched for now, it should work for batches
479                  * of completions except the very first one which may look
480                  * a little bursty
481                  */
482                 usec = utime_since(&s, &comp_time);
483
484                 rate_throttle(td, usec, bytes_done);
485
486                 if (check_min_rate(td, &comp_time)) {
487                         if (exitall_on_terminate)
488                                 terminate_threads(td->groupid, 0);
489                         td_verror(td, ENODATA, "check_min_rate");
490                         break;
491                 }
492
493                 if (td->thinktime) {
494                         unsigned long long b;
495
496                         b = td->io_blocks[0] + td->io_blocks[1];
497                         if (!(b % td->thinktime_blocks)) {
498                                 int left;
499
500                                 if (td->thinktime_spin)
501                                         __usec_sleep(td->thinktime_spin);
502
503                                 left = td->thinktime - td->thinktime_spin;
504                                 if (left)
505                                         usec_sleep(td, left);
506                         }
507                 }
508         }
509
510         if (!td->error) {
511                 struct fio_file *f;
512
513                 i = td->cur_depth;
514                 if (i)
515                         ret = io_u_queued_complete(td, i);
516
517                 if (should_fsync(td) && td->end_fsync) {
518                         td_set_runstate(td, TD_FSYNCING);
519                         for_each_file(td, f, i)
520                                 fio_io_sync(td, f);
521                 }
522         } else
523                 cleanup_pending_aio(td);
524 }
525
526 static void cleanup_io_u(struct thread_data *td)
527 {
528         struct list_head *entry, *n;
529         struct io_u *io_u;
530
531         list_for_each_safe(entry, n, &td->io_u_freelist) {
532                 io_u = list_entry(entry, struct io_u, list);
533
534                 list_del(&io_u->list);
535                 free(io_u);
536         }
537
538         free_io_mem(td);
539 }
540
541 /*
542  * "randomly" fill the buffer contents
543  */
544 static void fill_rand_buf(struct io_u *io_u, int max_bs)
545 {
546         int *ptr = io_u->buf;
547
548         while ((void *) ptr - io_u->buf < max_bs) {
549                 *ptr = rand() * 0x9e370001;
550                 ptr++;
551         }
552 }
553
554 static int init_io_u(struct thread_data *td)
555 {
556         struct io_u *io_u;
557         unsigned int max_bs;
558         int i, max_units;
559         char *p;
560
561         if (td->io_ops->flags & FIO_CPUIO)
562                 return 0;
563
564         if (td->io_ops->flags & FIO_SYNCIO)
565                 max_units = 1;
566         else
567                 max_units = td->iodepth;
568
569         max_bs = max(td->max_bs[DDIR_READ], td->max_bs[DDIR_WRITE]);
570         td->orig_buffer_size = max_bs * max_units;
571
572         if (td->mem_type == MEM_SHMHUGE || td->mem_type == MEM_MMAPHUGE)
573                 td->orig_buffer_size = (td->orig_buffer_size + td->hugepage_size - 1) & ~(td->hugepage_size - 1);
574         else
575                 td->orig_buffer_size += page_mask;
576
577         if (allocate_io_mem(td))
578                 return 1;
579
580         p = ALIGN(td->orig_buffer);
581         for (i = 0; i < max_units; i++) {
582                 io_u = malloc(sizeof(*io_u));
583                 memset(io_u, 0, sizeof(*io_u));
584                 INIT_LIST_HEAD(&io_u->list);
585
586                 io_u->buf = p + max_bs * i;
587                 if (td_write(td) || td_rw(td))
588                         fill_rand_buf(io_u, max_bs);
589
590                 io_u->index = i;
591                 io_u->flags = IO_U_F_FREE;
592                 list_add(&io_u->list, &td->io_u_freelist);
593         }
594
595         io_u_init_timeout();
596
597         return 0;
598 }
599
600 static int switch_ioscheduler(struct thread_data *td)
601 {
602         char tmp[256], tmp2[128];
603         FILE *f;
604         int ret;
605
606         if (td->io_ops->flags & FIO_CPUIO)
607                 return 0;
608
609         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
610
611         f = fopen(tmp, "r+");
612         if (!f) {
613                 td_verror(td, errno, "fopen");
614                 return 1;
615         }
616
617         /*
618          * Set io scheduler.
619          */
620         ret = fwrite(td->ioscheduler, strlen(td->ioscheduler), 1, f);
621         if (ferror(f) || ret != 1) {
622                 td_verror(td, errno, "fwrite");
623                 fclose(f);
624                 return 1;
625         }
626
627         rewind(f);
628
629         /*
630          * Read back and check that the selected scheduler is now the default.
631          */
632         ret = fread(tmp, 1, sizeof(tmp), f);
633         if (ferror(f) || ret < 0) {
634                 td_verror(td, errno, "fread");
635                 fclose(f);
636                 return 1;
637         }
638
639         sprintf(tmp2, "[%s]", td->ioscheduler);
640         if (!strstr(tmp, tmp2)) {
641                 log_err("fio: io scheduler %s not found\n", td->ioscheduler);
642                 td_verror(td, EINVAL, "iosched_switch");
643                 fclose(f);
644                 return 1;
645         }
646
647         fclose(f);
648         return 0;
649 }
650
651 static void clear_io_state(struct thread_data *td)
652 {
653         struct fio_file *f;
654         int i;
655
656         td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
657         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
658         td->zone_bytes = 0;
659
660         td->last_was_sync = 0;
661
662         for_each_file(td, f, i) {
663                 f->last_completed_pos = 0;
664
665                 f->last_pos = 0;
666                 if (td->io_ops->flags & FIO_SYNCIO)
667                         lseek(f->fd, SEEK_SET, 0);
668
669                 if (f->file_map)
670                         memset(f->file_map, 0, f->num_maps * sizeof(long));
671         }
672 }
673
674 /*
675  * Entry point for the thread based jobs. The process based jobs end up
676  * here as well, after a little setup.
677  */
678 static void *thread_main(void *data)
679 {
680         unsigned long long runtime[2];
681         struct thread_data *td = data;
682
683         if (!td->use_thread)
684                 setsid();
685
686         td->pid = getpid();
687
688         INIT_LIST_HEAD(&td->io_u_freelist);
689         INIT_LIST_HEAD(&td->io_u_busylist);
690         INIT_LIST_HEAD(&td->io_u_requeues);
691         INIT_LIST_HEAD(&td->io_hist_list);
692         INIT_LIST_HEAD(&td->io_log_list);
693
694         if (init_io_u(td))
695                 goto err;
696
697         if (fio_setaffinity(td) == -1) {
698                 td_verror(td, errno, "cpu_set_affinity");
699                 goto err;
700         }
701
702         if (init_iolog(td))
703                 goto err;
704
705         if (td->ioprio) {
706                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
707                         td_verror(td, errno, "ioprio_set");
708                         goto err;
709                 }
710         }
711
712         if (nice(td->nice) == -1) {
713                 td_verror(td, errno, "nice");
714                 goto err;
715         }
716
717         if (init_random_state(td))
718                 goto err;
719
720         if (td->ioscheduler && switch_ioscheduler(td))
721                 goto err;
722
723         td_set_runstate(td, TD_INITIALIZED);
724         fio_sem_up(&startup_sem);
725         fio_sem_down(&td->mutex);
726
727         if (!td->create_serialize && setup_files(td))
728                 goto err;
729         if (open_files(td))
730                 goto err;
731
732         /*
733          * Do this late, as some IO engines would like to have the
734          * files setup prior to initializing structures.
735          */
736         if (td_io_init(td))
737                 goto err;
738
739         if (td->exec_prerun) {
740                 if (system(td->exec_prerun) < 0)
741                         goto err;
742         }
743
744         fio_gettime(&td->epoch, NULL);
745         memcpy(&td->timeout_end, &td->epoch, sizeof(td->epoch));
746         getrusage(RUSAGE_SELF, &td->ts.ru_start);
747
748         runtime[0] = runtime[1] = 0;
749         while (td->loops--) {
750                 fio_gettime(&td->start, NULL);
751                 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
752
753                 if (td->ratemin)
754                         memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
755
756                 clear_io_state(td);
757                 prune_io_piece_log(td);
758
759                 if (td->io_ops->flags & FIO_CPUIO)
760                         do_cpuio(td);
761                 else
762                         do_io(td);
763
764                 if (td_read(td) && td->io_bytes[DDIR_READ])
765                         runtime[DDIR_READ] += utime_since_now(&td->start);
766                 if (td_write(td) && td->io_bytes[DDIR_WRITE])
767                         runtime[DDIR_WRITE] += utime_since_now(&td->start);
768                 
769                 if (td->error || td->terminate)
770                         break;
771
772                 if (td->verify == VERIFY_NONE)
773                         continue;
774
775                 clear_io_state(td);
776                 fio_gettime(&td->start, NULL);
777
778                 do_verify(td);
779
780                 runtime[DDIR_READ] += utime_since_now(&td->start);
781
782                 if (td->error || td->terminate)
783                         break;
784         }
785
786         update_rusage_stat(td);
787         fio_gettime(&td->end_time, NULL);
788         td->runtime[0] = runtime[0] / 1000;
789         td->runtime[1] = runtime[1] / 1000;
790
791         if (td->ts.bw_log)
792                 finish_log(td, td->ts.bw_log, "bw");
793         if (td->ts.slat_log)
794                 finish_log(td, td->ts.slat_log, "slat");
795         if (td->ts.clat_log)
796                 finish_log(td, td->ts.clat_log, "clat");
797         if (td->write_iolog_file)
798                 write_iolog_close(td);
799         if (td->exec_postrun) {
800                 if (system(td->exec_postrun) < 0)
801                         log_err("fio: postrun %s failed\n", td->exec_postrun);
802         }
803
804         if (exitall_on_terminate)
805                 terminate_threads(td->groupid, 0);
806
807 err:
808         if (td->error)
809                 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
810         close_files(td);
811         close_ioengine(td);
812         cleanup_io_u(td);
813         td_set_runstate(td, TD_EXITED);
814         return (void *) (unsigned long) td->error;
815 }
816
817 /*
818  * We cannot pass the td data into a forked process, so attach the td and
819  * pass it to the thread worker.
820  */
821 static int fork_main(int shmid, int offset)
822 {
823         struct thread_data *td;
824         void *data, *ret;
825
826         data = shmat(shmid, NULL, 0);
827         if (data == (void *) -1) {
828                 int __err = errno;
829
830                 perror("shmat");
831                 return __err;
832         }
833
834         td = data + offset * sizeof(struct thread_data);
835         ret = thread_main(td);
836         shmdt(data);
837         return (int) (unsigned long) ret;
838 }
839
840 /*
841  * Run over the job map and reap the threads that have exited, if any.
842  */
843 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
844 {
845         struct thread_data *td;
846         int i, cputhreads, pending, status, ret;
847
848         /*
849          * reap exited threads (TD_EXITED -> TD_REAPED)
850          */
851         pending = cputhreads = 0;
852         for_each_td(td, i) {
853                 int flags = 0;
854
855                 /*
856                  * ->io_ops is NULL for a thread that has closed its
857                  * io engine
858                  */
859                 if (td->io_ops && td->io_ops->flags & FIO_CPUIO)
860                         cputhreads++;
861
862                 if (!td->pid || td->runstate == TD_REAPED)
863                         continue;
864                 if (td->use_thread) {
865                         if (td->runstate == TD_EXITED) {
866                                 td_set_runstate(td, TD_REAPED);
867                                 goto reaped;
868                         }
869                         continue;
870                 }
871
872                 flags = WNOHANG;
873                 if (td->runstate == TD_EXITED)
874                         flags = 0;
875
876                 /*
877                  * check if someone quit or got killed in an unusual way
878                  */
879                 ret = waitpid(td->pid, &status, flags);
880                 if (ret < 0) {
881                         if (errno == ECHILD) {
882                                 log_err("fio: pid=%d disappeared %d\n", td->pid, td->runstate);
883                                 td_set_runstate(td, TD_REAPED);
884                                 goto reaped;
885                         }
886                         perror("waitpid");
887                 } else if (ret == td->pid) {
888                         if (WIFSIGNALED(status)) {
889                                 int sig = WTERMSIG(status);
890
891                                 log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
892                                 td_set_runstate(td, TD_REAPED);
893                                 goto reaped;
894                         }
895                         if (WIFEXITED(status)) {
896                                 if (WEXITSTATUS(status) && !td->error)
897                                         td->error = WEXITSTATUS(status);
898
899                                 td_set_runstate(td, TD_REAPED);
900                                 goto reaped;
901                         }
902                 }
903
904                 /*
905                  * thread is not dead, continue
906                  */
907                 continue;
908 reaped:
909                 if (td->use_thread) {
910                         long ret;
911
912                         if (pthread_join(td->thread, (void *) &ret))
913                                 perror("pthread_join");
914                 }
915
916                 (*nr_running)--;
917                 (*m_rate) -= td->ratemin;
918                 (*t_rate) -= td->rate;
919
920                 if (td->error)
921                         exit_value++;
922         }
923
924         if (*nr_running == cputhreads && !pending)
925                 terminate_threads(TERMINATE_ALL, 0);
926 }
927
928 /*
929  * Main function for kicking off and reaping jobs, as needed.
930  */
931 static void run_threads(void)
932 {
933         struct thread_data *td;
934         unsigned long spent;
935         int i, todo, nr_running, m_rate, t_rate, nr_started;
936
937         if (fio_pin_memory())
938                 return;
939
940         if (!terse_output) {
941                 printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
942                 fflush(stdout);
943         }
944
945         signal(SIGINT, sig_handler);
946         signal(SIGALRM, sig_handler);
947
948         todo = thread_number;
949         nr_running = 0;
950         nr_started = 0;
951         m_rate = t_rate = 0;
952
953         for_each_td(td, i) {
954                 print_status_init(td->thread_number - 1);
955
956                 if (!td->create_serialize) {
957                         init_disk_util(td);
958                         continue;
959                 }
960
961                 /*
962                  * do file setup here so it happens sequentially,
963                  * we don't want X number of threads getting their
964                  * client data interspersed on disk
965                  */
966                 if (setup_files(td)) {
967                         exit_value++;
968                         if (td->error)
969                                 log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
970                         td_set_runstate(td, TD_REAPED);
971                         todo--;
972                 }
973
974                 init_disk_util(td);
975         }
976
977         set_genesis_time();
978
979         while (todo) {
980                 struct thread_data *map[MAX_JOBS];
981                 struct timeval this_start;
982                 int this_jobs = 0, left;
983
984                 /*
985                  * create threads (TD_NOT_CREATED -> TD_CREATED)
986                  */
987                 for_each_td(td, i) {
988                         if (td->runstate != TD_NOT_CREATED)
989                                 continue;
990
991                         /*
992                          * never got a chance to start, killed by other
993                          * thread for some reason
994                          */
995                         if (td->terminate) {
996                                 todo--;
997                                 continue;
998                         }
999
1000                         if (td->start_delay) {
1001                                 spent = mtime_since_genesis();
1002
1003                                 if (td->start_delay * 1000 > spent)
1004                                         continue;
1005                         }
1006
1007                         if (td->stonewall && (nr_started || nr_running))
1008                                 break;
1009
1010                         /*
1011                          * Set state to created. Thread will transition
1012                          * to TD_INITIALIZED when it's done setting up.
1013                          */
1014                         td_set_runstate(td, TD_CREATED);
1015                         map[this_jobs++] = td;
1016                         fio_sem_init(&startup_sem, 1);
1017                         nr_started++;
1018
1019                         if (td->use_thread) {
1020                                 if (pthread_create(&td->thread, NULL, thread_main, td)) {
1021                                         perror("thread_create");
1022                                         nr_started--;
1023                                 }
1024                         } else {
1025                                 if (fork())
1026                                         fio_sem_down(&startup_sem);
1027                                 else {
1028                                         int ret = fork_main(shm_id, i);
1029
1030                                         exit(ret);
1031                                 }
1032                         }
1033                 }
1034
1035                 /*
1036                  * Wait for the started threads to transition to
1037                  * TD_INITIALIZED.
1038                  */
1039                 fio_gettime(&this_start, NULL);
1040                 left = this_jobs;
1041                 while (left && !fio_abort) {
1042                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1043                                 break;
1044
1045                         usleep(100000);
1046
1047                         for (i = 0; i < this_jobs; i++) {
1048                                 td = map[i];
1049                                 if (!td)
1050                                         continue;
1051                                 if (td->runstate == TD_INITIALIZED) {
1052                                         map[i] = NULL;
1053                                         left--;
1054                                 } else if (td->runstate >= TD_EXITED) {
1055                                         map[i] = NULL;
1056                                         left--;
1057                                         todo--;
1058                                         nr_running++; /* work-around... */
1059                                 }
1060                         }
1061                 }
1062
1063                 if (left) {
1064                         log_err("fio: %d jobs failed to start\n", left);
1065                         for (i = 0; i < this_jobs; i++) {
1066                                 td = map[i];
1067                                 if (!td)
1068                                         continue;
1069                                 kill(td->pid, SIGTERM);
1070                         }
1071                         break;
1072                 }
1073
1074                 /*
1075                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1076                  */
1077                 for_each_td(td, i) {
1078                         if (td->runstate != TD_INITIALIZED)
1079                                 continue;
1080
1081                         td_set_runstate(td, TD_RUNNING);
1082                         nr_running++;
1083                         nr_started--;
1084                         m_rate += td->ratemin;
1085                         t_rate += td->rate;
1086                         todo--;
1087                         fio_sem_up(&td->mutex);
1088                 }
1089
1090                 reap_threads(&nr_running, &t_rate, &m_rate);
1091
1092                 if (todo)
1093                         usleep(100000);
1094         }
1095
1096         while (nr_running) {
1097                 reap_threads(&nr_running, &t_rate, &m_rate);
1098                 usleep(10000);
1099         }
1100
1101         update_io_ticks();
1102         fio_unpin_memory();
1103 }
1104
1105 int main(int argc, char *argv[])
1106 {
1107         long ps;
1108
1109         /*
1110          * We need locale for number printing, if it isn't set then just
1111          * go with the US format.
1112          */
1113         if (!getenv("LC_NUMERIC"))
1114                 setlocale(LC_NUMERIC, "en_US");
1115
1116         if (parse_options(argc, argv))
1117                 return 1;
1118
1119         if (!thread_number) {
1120                 log_err("Nothing to do\n");
1121                 return 1;
1122         }
1123
1124         ps = sysconf(_SC_PAGESIZE);
1125         if (ps < 0) {
1126                 log_err("Failed to get page size\n");
1127                 return 1;
1128         }
1129
1130         page_mask = ps - 1;
1131
1132         if (write_bw_log) {
1133                 setup_log(&agg_io_log[DDIR_READ]);
1134                 setup_log(&agg_io_log[DDIR_WRITE]);
1135         }
1136
1137         set_genesis_time();
1138
1139         disk_util_timer_arm();
1140
1141         run_threads();
1142
1143         if (!fio_abort) {
1144                 show_run_stats();
1145                 if (write_bw_log) {
1146                         __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1147                         __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");
1148                 }
1149         }
1150
1151         return exit_value;
1152 }