Update fio io engine version
[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 unsigned long page_mask;
41 unsigned long page_size;
42 #define ALIGN(buf)      \
43         (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
44
45 int groupid = 0;
46 int thread_number = 0;
47 int nr_process = 0;
48 int nr_thread = 0;
49 int shm_id = 0;
50 int temp_stall_ts;
51
52 static struct fio_sem *startup_sem;
53 static volatile int fio_abort;
54 static int exit_value;
55
56 struct io_log *agg_io_log[2];
57
58 #define TERMINATE_ALL           (-1)
59 #define JOB_START_TIMEOUT       (5 * 1000)
60
61 static inline void td_set_runstate(struct thread_data *td, int runstate)
62 {
63         td->runstate = runstate;
64 }
65
66 static void terminate_threads(int group_id)
67 {
68         struct thread_data *td;
69         int i;
70
71         for_each_td(td, i) {
72                 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
73                         /*
74                          * if the thread is running, just let it exit
75                          */
76                         if (td->runstate < TD_RUNNING)
77                                 kill(td->pid, SIGQUIT);
78                         td->terminate = 1;
79                         td->o.start_delay = 0;
80                 }
81         }
82 }
83
84 static void sig_handler(int sig)
85 {
86         switch (sig) {
87                 case SIGALRM:
88                         update_io_ticks();
89                         disk_util_timer_arm();
90                         print_thread_status();
91                         break;
92                 default:
93                         printf("\nfio: terminating on signal %d\n", sig);
94                         fflush(stdout);
95                         terminate_threads(TERMINATE_ALL);
96                         break;
97         }
98 }
99
100 /*
101  * Check if we are above the minimum rate given.
102  */
103 static int check_min_rate(struct thread_data *td, struct timeval *now)
104 {
105         unsigned long long bytes = 0;
106         unsigned long iops = 0;
107         unsigned long spent;
108         unsigned long rate;
109
110         /*
111          * No minimum rate set, always ok
112          */
113         if (!td->o.ratemin && !td->o.rate_iops_min)
114                 return 0;
115
116         /*
117          * allow a 2 second settle period in the beginning
118          */
119         if (mtime_since(&td->start, now) < 2000)
120                 return 0;
121
122         if (td_read(td)) {
123                 iops += td->io_blocks[DDIR_READ];
124                 bytes += td->this_io_bytes[DDIR_READ];
125         }
126         if (td_write(td)) {
127                 iops += td->io_blocks[DDIR_WRITE];
128                 bytes += td->this_io_bytes[DDIR_WRITE];
129         }
130
131         /*
132          * if rate blocks is set, sample is running
133          */
134         if (td->rate_bytes || td->rate_blocks) {
135                 spent = mtime_since(&td->lastrate, now);
136                 if (spent < td->o.ratecycle)
137                         return 0;
138
139                 if (td->o.rate) {
140                         /*
141                          * check bandwidth specified rate
142                          */
143                         if (bytes < td->rate_bytes) {
144                                 log_err("%s: min rate %u not met\n", td->o.name, td->o.ratemin);
145                                 return 1;
146                         } else {
147                                 rate = (bytes - td->rate_bytes) / spent;
148                                 if (rate < td->o.ratemin || bytes < td->rate_bytes) {
149                                         log_err("%s: min rate %u not met, got %luKiB/sec\n", td->o.name, td->o.ratemin, rate);
150                                         return 1;
151                                 }
152                         }
153                 } else {
154                         /*
155                          * checks iops specified rate
156                          */
157                         if (iops < td->o.rate_iops) {
158                                 log_err("%s: min iops rate %u not met\n", td->o.name, td->o.rate_iops);
159                                 return 1;
160                         } else {
161                                 rate = (iops - td->rate_blocks) / spent;
162                                 if (rate < td->o.rate_iops_min || iops < td->rate_blocks) {
163                                         log_err("%s: min iops rate %u not met, got %lu\n", td->o.name, td->o.rate_iops_min, rate);
164                                 }
165                         }
166                 }
167         }
168
169         td->rate_bytes = bytes;
170         td->rate_blocks = iops;
171         memcpy(&td->lastrate, now, sizeof(*now));
172         return 0;
173 }
174
175 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
176 {
177         if (!td->o.timeout)
178                 return 0;
179         if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
180                 return 1;
181
182         return 0;
183 }
184
185 /*
186  * When job exits, we can cancel the in-flight IO if we are using async
187  * io. Attempt to do so.
188  */
189 static void cleanup_pending_aio(struct thread_data *td)
190 {
191         struct list_head *entry, *n;
192         struct io_u *io_u;
193         int r;
194
195         /*
196          * get immediately available events, if any
197          */
198         r = io_u_queued_complete(td, 0);
199         if (r < 0)
200                 return;
201
202         /*
203          * now cancel remaining active events
204          */
205         if (td->io_ops->cancel) {
206                 list_for_each_safe(entry, n, &td->io_u_busylist) {
207                         io_u = list_entry(entry, struct io_u, list);
208
209                         /*
210                          * if the io_u isn't in flight, then that generally
211                          * means someone leaked an io_u. complain but fix
212                          * it up, so we don't stall here.
213                          */
214                         if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
215                                 log_err("fio: non-busy IO on busy list\n");
216                                 put_io_u(td, io_u);
217                         } else {
218                                 r = td->io_ops->cancel(td, io_u);
219                                 if (!r)
220                                         put_io_u(td, io_u);
221                         }
222                 }
223         }
224
225         if (td->cur_depth)
226                 r = io_u_queued_complete(td, td->cur_depth);
227 }
228
229 /*
230  * Helper to handle the final sync of a file. Works just like the normal
231  * io path, just does everything sync.
232  */
233 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
234 {
235         struct io_u *io_u = __get_io_u(td);
236         int ret;
237
238         if (!io_u)
239                 return 1;
240
241         io_u->ddir = DDIR_SYNC;
242         io_u->file = f;
243
244         if (td_io_prep(td, io_u)) {
245                 put_io_u(td, io_u);
246                 return 1;
247         }
248
249 requeue:
250         ret = td_io_queue(td, io_u);
251         if (ret < 0) {
252                 td_verror(td, io_u->error, "td_io_queue");
253                 put_io_u(td, io_u);
254                 return 1;
255         } else if (ret == FIO_Q_QUEUED) {
256                 if (io_u_queued_complete(td, 1) < 0)
257                         return 1;
258         } else if (ret == FIO_Q_COMPLETED) {
259                 if (io_u->error) {
260                         td_verror(td, io_u->error, "td_io_queue");
261                         return 1;
262                 }
263
264                 if (io_u_sync_complete(td, io_u) < 0)
265                         return 1;
266         } else if (ret == FIO_Q_BUSY) {
267                 if (td_io_commit(td))
268                         return 1;
269                 goto requeue;
270         }
271
272         return 0;
273 }
274
275 /*
276  * The main verify engine. Runs over the writes we previously submitted,
277  * reads the blocks back in, and checks the crc/md5 of the data.
278  */
279 static void do_verify(struct thread_data *td)
280 {
281         struct fio_file *f;
282         struct io_u *io_u;
283         int ret, min_events;
284         unsigned int i;
285
286         /*
287          * sync io first and invalidate cache, to make sure we really
288          * read from disk.
289          */
290         for_each_file(td, f, i) {
291                 if (!(f->flags & FIO_FILE_OPEN))
292                         continue;
293                 if (fio_io_sync(td, f))
294                         break;
295                 if (file_invalidate_cache(td, f))
296                         break;
297         }
298
299         if (td->error)
300                 return;
301
302         td_set_runstate(td, TD_VERIFYING);
303
304         io_u = NULL;
305         while (!td->terminate) {
306                 int ret2;
307
308                 io_u = __get_io_u(td);
309                 if (!io_u)
310                         break;
311
312                 if (runtime_exceeded(td, &io_u->start_time)) {
313                         put_io_u(td, io_u);
314                         break;
315                 }
316
317                 if (get_next_verify(td, io_u)) {
318                         put_io_u(td, io_u);
319                         break;
320                 }
321
322                 if (td_io_prep(td, io_u)) {
323                         put_io_u(td, io_u);
324                         break;
325                 }
326
327                 io_u->end_io = verify_io_u;
328
329                 ret = td_io_queue(td, io_u);
330                 switch (ret) {
331                 case FIO_Q_COMPLETED:
332                         if (io_u->error)
333                                 ret = -io_u->error;
334                         else if (io_u->resid) {
335                                 int bytes = io_u->xfer_buflen - io_u->resid;
336
337                                 /*
338                                  * zero read, fail
339                                  */
340                                 if (!bytes) {
341                                         td_verror(td, ENODATA, "full resid");
342                                         put_io_u(td, io_u);
343                                         break;
344                                 }
345                                 io_u->xfer_buflen = io_u->resid;
346                                 io_u->xfer_buf += bytes;
347                                 requeue_io_u(td, &io_u);
348                         } else {
349                                 ret = io_u_sync_complete(td, io_u);
350                                 if (ret < 0)
351                                         break;
352                         }
353                         continue;
354                 case FIO_Q_QUEUED:
355                         break;
356                 case FIO_Q_BUSY:
357                         requeue_io_u(td, &io_u);
358                         ret2 = td_io_commit(td);
359                         if (ret2 < 0)
360                                 ret = ret2;
361                         break;
362                 default:
363                         assert(ret < 0);
364                         td_verror(td, -ret, "td_io_queue");
365                         break;
366                 }
367
368                 if (ret < 0 || td->error)
369                         break;
370
371                 /*
372                  * if we can queue more, do so. but check if there are
373                  * completed io_u's first.
374                  */
375                 min_events = 0;
376                 if (queue_full(td) || ret == FIO_Q_BUSY) {
377                         min_events = 1;
378
379                         if (td->cur_depth > td->o.iodepth_low)
380                                 min_events = td->cur_depth - td->o.iodepth_low;
381                 }
382
383                 /*
384                  * Reap required number of io units, if any, and do the
385                  * verification on them through the callback handler
386                  */
387                 if (io_u_queued_complete(td, min_events) < 0)
388                         break;
389         }
390
391         if (!td->error) {
392                 min_events = td->cur_depth;
393
394                 if (min_events)
395                         ret = io_u_queued_complete(td, min_events);
396         } else
397                 cleanup_pending_aio(td);
398
399         td_set_runstate(td, TD_RUNNING);
400 }
401
402 /*
403  * Main IO worker function. It retrieves io_u's to process and queues
404  * and reaps them, checking for rate and errors along the way.
405  */
406 static void do_io(struct thread_data *td)
407 {
408         struct timeval s;
409         unsigned long usec;
410         unsigned int i;
411         int ret = 0;
412
413         td_set_runstate(td, TD_RUNNING);
414
415         while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) {
416                 struct timeval comp_time;
417                 long bytes_done = 0;
418                 int min_evts = 0;
419                 struct io_u *io_u;
420                 int ret2;
421
422                 if (td->terminate)
423                         break;
424
425                 io_u = get_io_u(td);
426                 if (!io_u)
427                         break;
428
429                 memcpy(&s, &io_u->start_time, sizeof(s));
430
431                 if (runtime_exceeded(td, &s)) {
432                         put_io_u(td, io_u);
433                         break;
434                 }
435
436                 ret = td_io_queue(td, io_u);
437                 switch (ret) {
438                 case FIO_Q_COMPLETED:
439                         if (io_u->error)
440                                 ret = -io_u->error;
441                         else if (io_u->resid) {
442                                 int bytes = io_u->xfer_buflen - io_u->resid;
443
444                                 /*
445                                  * zero read, fail
446                                  */
447                                 if (!bytes) {
448                                         td_verror(td, ENODATA, "full resid");
449                                         put_io_u(td, io_u);
450                                         break;
451                                 }
452
453                                 io_u->xfer_buflen = io_u->resid;
454                                 io_u->xfer_buf += bytes;
455                                 requeue_io_u(td, &io_u);
456                         } else {
457                                 fio_gettime(&comp_time, NULL);
458                                 bytes_done = io_u_sync_complete(td, io_u);
459                                 if (bytes_done < 0)
460                                         ret = bytes_done;
461                         }
462                         break;
463                 case FIO_Q_QUEUED:
464                         /*
465                          * if the engine doesn't have a commit hook,
466                          * the io_u is really queued. if it does have such
467                          * a hook, it has to call io_u_queued() itself.
468                          */
469                         if (td->io_ops->commit == NULL)
470                                 io_u_queued(td, io_u);
471                         break;
472                 case FIO_Q_BUSY:
473                         requeue_io_u(td, &io_u);
474                         ret2 = td_io_commit(td);
475                         if (ret2 < 0)
476                                 ret = ret2;
477                         break;
478                 default:
479                         assert(ret < 0);
480                         put_io_u(td, io_u);
481                         break;
482                 }
483
484                 if (ret < 0 || td->error)
485                         break;
486
487                 /*
488                  * See if we need to complete some commands
489                  */
490                 if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
491                         min_evts = 0;
492                         if (queue_full(td) || ret == FIO_Q_BUSY) {
493                                 min_evts = 1;
494
495                                 if (td->cur_depth > td->o.iodepth_low)
496                                         min_evts = td->cur_depth - td->o.iodepth_low;
497                         }
498
499                         fio_gettime(&comp_time, NULL);
500                         bytes_done = io_u_queued_complete(td, min_evts);
501                         if (bytes_done < 0)
502                                 break;
503                 }
504
505                 if (!bytes_done)
506                         continue;
507
508                 /*
509                  * the rate is batched for now, it should work for batches
510                  * of completions except the very first one which may look
511                  * a little bursty
512                  */
513                 usec = utime_since(&s, &comp_time);
514
515                 rate_throttle(td, usec, bytes_done);
516
517                 if (check_min_rate(td, &comp_time)) {
518                         if (exitall_on_terminate)
519                                 terminate_threads(td->groupid);
520                         td_verror(td, ENODATA, "check_min_rate");
521                         break;
522                 }
523
524                 if (td->o.thinktime) {
525                         unsigned long long b;
526
527                         b = td->io_blocks[0] + td->io_blocks[1];
528                         if (!(b % td->o.thinktime_blocks)) {
529                                 int left;
530
531                                 if (td->o.thinktime_spin)
532                                         __usec_sleep(td->o.thinktime_spin);
533
534                                 left = td->o.thinktime - td->o.thinktime_spin;
535                                 if (left)
536                                         usec_sleep(td, left);
537                         }
538                 }
539         }
540
541         if (!td->error) {
542                 struct fio_file *f;
543
544                 i = td->cur_depth;
545                 if (i)
546                         ret = io_u_queued_complete(td, i);
547
548                 if (should_fsync(td) && td->o.end_fsync) {
549                         td_set_runstate(td, TD_FSYNCING);
550
551                         for_each_file(td, f, i) {
552                                 if (!(f->flags & FIO_FILE_OPEN))
553                                         continue;
554                                 fio_io_sync(td, f);
555                         }
556                 }
557         } else
558                 cleanup_pending_aio(td);
559 }
560
561 static void cleanup_io_u(struct thread_data *td)
562 {
563         struct list_head *entry, *n;
564         struct io_u *io_u;
565
566         list_for_each_safe(entry, n, &td->io_u_freelist) {
567                 io_u = list_entry(entry, struct io_u, list);
568
569                 list_del(&io_u->list);
570                 free(io_u);
571         }
572
573         free_io_mem(td);
574 }
575
576 /*
577  * "randomly" fill the buffer contents
578  */
579 static void fill_rand_buf(struct io_u *io_u, int max_bs)
580 {
581         int *ptr = io_u->buf;
582
583         while ((void *) ptr - io_u->buf < max_bs) {
584                 *ptr = rand() * 0x9e370001;
585                 ptr++;
586         }
587 }
588
589 static int init_io_u(struct thread_data *td)
590 {
591         unsigned long long buf_size;
592         struct io_u *io_u;
593         unsigned int max_bs;
594         int i, max_units;
595         char *p;
596
597         if (td->io_ops->flags & FIO_SYNCIO)
598                 max_units = 1;
599         else
600                 max_units = td->o.iodepth;
601
602         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
603         buf_size = (unsigned long long) max_bs * (unsigned long long) max_units;
604         buf_size += page_mask;
605         if (buf_size != (size_t) buf_size) {
606                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
607                 return 1;
608         }
609
610         td->orig_buffer_size = buf_size;
611
612         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE)
613                 td->orig_buffer_size = (td->orig_buffer_size + td->o.hugepage_size - 1) & ~(td->o.hugepage_size - 1);
614         else if (td->orig_buffer_size & page_mask)
615                 td->orig_buffer_size = (td->orig_buffer_size + page_mask) & ~page_mask;
616
617         if (allocate_io_mem(td))
618                 return 1;
619
620         p = ALIGN(td->orig_buffer);
621         for (i = 0; i < max_units; i++) {
622                 io_u = malloc(sizeof(*io_u));
623                 memset(io_u, 0, sizeof(*io_u));
624                 INIT_LIST_HEAD(&io_u->list);
625
626                 io_u->buf = p + max_bs * i;
627                 if (td_write(td) || td_rw(td))
628                         fill_rand_buf(io_u, max_bs);
629
630                 io_u->index = i;
631                 io_u->flags = IO_U_F_FREE;
632                 list_add(&io_u->list, &td->io_u_freelist);
633         }
634
635         io_u_init_timeout();
636
637         return 0;
638 }
639
640 static int switch_ioscheduler(struct thread_data *td)
641 {
642         char tmp[256], tmp2[128];
643         FILE *f;
644         int ret;
645
646         if (td->io_ops->flags & FIO_DISKLESSIO)
647                 return 0;
648
649         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
650
651         f = fopen(tmp, "r+");
652         if (!f) {
653                 td_verror(td, errno, "fopen");
654                 return 1;
655         }
656
657         /*
658          * Set io scheduler.
659          */
660         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
661         if (ferror(f) || ret != 1) {
662                 td_verror(td, errno, "fwrite");
663                 fclose(f);
664                 return 1;
665         }
666
667         rewind(f);
668
669         /*
670          * Read back and check that the selected scheduler is now the default.
671          */
672         ret = fread(tmp, 1, sizeof(tmp), f);
673         if (ferror(f) || ret < 0) {
674                 td_verror(td, errno, "fread");
675                 fclose(f);
676                 return 1;
677         }
678
679         sprintf(tmp2, "[%s]", td->o.ioscheduler);
680         if (!strstr(tmp, tmp2)) {
681                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
682                 td_verror(td, EINVAL, "iosched_switch");
683                 fclose(f);
684                 return 1;
685         }
686
687         fclose(f);
688         return 0;
689 }
690
691 static int clear_io_state(struct thread_data *td)
692 {
693         struct fio_file *f;
694         unsigned int i;
695         int ret;
696
697         td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
698         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
699         td->zone_bytes = 0;
700         td->rate_bytes = 0;
701         td->rate_blocks = 0;
702         td->rw_end_set[0] = td->rw_end_set[1] = 0;
703
704         td->last_was_sync = 0;
705
706         for_each_file(td, f, i)
707                 td_io_close_file(td, f);
708
709         ret = 0;
710         for_each_file(td, f, i) {
711                 ret = td_io_open_file(td, f);
712                 if (ret)
713                         break;
714         }
715
716         return ret;
717 }
718
719 /*
720  * Entry point for the thread based jobs. The process based jobs end up
721  * here as well, after a little setup.
722  */
723 static void *thread_main(void *data)
724 {
725         unsigned long long runtime[2];
726         struct thread_data *td = data;
727         unsigned long elapsed;
728         int clear_state;
729
730         if (!td->o.use_thread)
731                 setsid();
732
733         td->pid = getpid();
734
735         INIT_LIST_HEAD(&td->io_u_freelist);
736         INIT_LIST_HEAD(&td->io_u_busylist);
737         INIT_LIST_HEAD(&td->io_u_requeues);
738         INIT_LIST_HEAD(&td->io_log_list);
739         INIT_LIST_HEAD(&td->io_hist_list);
740         td->io_hist_tree = RB_ROOT;
741
742         if (init_io_u(td))
743                 goto err_sem;
744
745         if (fio_setaffinity(td) == -1) {
746                 td_verror(td, errno, "cpu_set_affinity");
747                 goto err_sem;
748         }
749
750         if (init_iolog(td))
751                 goto err_sem;
752
753         if (td->ioprio) {
754                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
755                         td_verror(td, errno, "ioprio_set");
756                         goto err_sem;
757                 }
758         }
759
760         if (nice(td->o.nice) == -1) {
761                 td_verror(td, errno, "nice");
762                 goto err_sem;
763         }
764
765         if (td->o.ioscheduler && switch_ioscheduler(td))
766                 goto err_sem;
767
768         td_set_runstate(td, TD_INITIALIZED);
769         fio_sem_up(startup_sem);
770         fio_sem_down(td->mutex);
771
772         /*
773          * the ->mutex semaphore is now no longer used, close it to avoid
774          * eating a file descriptor
775          */
776         fio_sem_remove(td->mutex);
777
778         if (!td->o.create_serialize && setup_files(td))
779                 goto err;
780
781         if (td_io_init(td))
782                 goto err;
783
784         if (open_files(td))
785                 goto err;
786
787         if (init_random_map(td))
788                 goto err;
789
790         if (td->o.exec_prerun) {
791                 if (system(td->o.exec_prerun) < 0)
792                         goto err;
793         }
794
795         fio_gettime(&td->epoch, NULL);
796         memcpy(&td->timeout_end, &td->epoch, sizeof(td->epoch));
797         getrusage(RUSAGE_SELF, &td->ts.ru_start);
798
799         runtime[0] = runtime[1] = 0;
800         clear_state = 0;
801         while (td->o.loops--) {
802                 fio_gettime(&td->start, NULL);
803                 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
804
805                 if (td->o.ratemin)
806                         memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
807
808                 if (clear_state && clear_io_state(td))
809                         break;
810
811                 prune_io_piece_log(td);
812
813                 do_io(td);
814
815                 clear_state = 1;
816
817                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
818                         if (td->rw_end_set[DDIR_READ])
819                                 elapsed = utime_since(&td->start, &td->rw_end[DDIR_READ]);
820                         else
821                                 elapsed = utime_since_now(&td->start);
822
823                         runtime[DDIR_READ] += elapsed;
824                 }
825                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
826                         if (td->rw_end_set[DDIR_WRITE])
827                                 elapsed = utime_since(&td->start, &td->rw_end[DDIR_WRITE]);
828                         else
829                                 elapsed = utime_since_now(&td->start);
830
831                         runtime[DDIR_WRITE] += elapsed;
832                 }
833                 
834                 if (td->error || td->terminate)
835                         break;
836
837                 if (td->o.verify == VERIFY_NONE)
838                         continue;
839
840                 if (clear_io_state(td))
841                         break;
842
843                 fio_gettime(&td->start, NULL);
844
845                 do_verify(td);
846
847                 runtime[DDIR_READ] += utime_since_now(&td->start);
848
849                 if (td->error || td->terminate)
850                         break;
851         }
852
853         update_rusage_stat(td);
854         td->ts.runtime[0] = runtime[0] / 1000;
855         td->ts.runtime[1] = runtime[1] / 1000;
856         td->ts.total_run_time = mtime_since_now(&td->epoch);
857         td->ts.io_bytes[0] = td->io_bytes[0];
858         td->ts.io_bytes[1] = td->io_bytes[1];
859
860         if (td->ts.bw_log)
861                 finish_log(td, td->ts.bw_log, "bw");
862         if (td->ts.slat_log)
863                 finish_log(td, td->ts.slat_log, "slat");
864         if (td->ts.clat_log)
865                 finish_log(td, td->ts.clat_log, "clat");
866         if (td->o.write_iolog_file)
867                 write_iolog_close(td);
868         if (td->o.exec_postrun) {
869                 if (system(td->o.exec_postrun) < 0)
870                         log_err("fio: postrun %s failed\n", td->o.exec_postrun);
871         }
872
873         if (exitall_on_terminate)
874                 terminate_threads(td->groupid);
875
876 err:
877         if (td->error)
878                 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
879         close_files(td);
880         close_ioengine(td);
881         cleanup_io_u(td);
882         options_mem_free(td);
883         td_set_runstate(td, TD_EXITED);
884         return (void *) (unsigned long) td->error;
885 err_sem:
886         fio_sem_up(startup_sem);
887         goto err;
888 }
889
890 /*
891  * We cannot pass the td data into a forked process, so attach the td and
892  * pass it to the thread worker.
893  */
894 static int fork_main(int shmid, int offset)
895 {
896         struct thread_data *td;
897         void *data, *ret;
898
899         data = shmat(shmid, NULL, 0);
900         if (data == (void *) -1) {
901                 int __err = errno;
902
903                 perror("shmat");
904                 return __err;
905         }
906
907         td = data + offset * sizeof(struct thread_data);
908         ret = thread_main(td);
909         shmdt(data);
910         return (int) (unsigned long) ret;
911 }
912
913 /*
914  * Run over the job map and reap the threads that have exited, if any.
915  */
916 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
917 {
918         struct thread_data *td;
919         int i, cputhreads, pending, status, ret;
920
921         /*
922          * reap exited threads (TD_EXITED -> TD_REAPED)
923          */
924         pending = cputhreads = 0;
925         for_each_td(td, i) {
926                 int flags = 0;
927
928                 /*
929                  * ->io_ops is NULL for a thread that has closed its
930                  * io engine
931                  */
932                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
933                         cputhreads++;
934
935                 if (!td->pid || td->runstate == TD_REAPED)
936                         continue;
937                 if (td->o.use_thread) {
938                         if (td->runstate == TD_EXITED) {
939                                 td_set_runstate(td, TD_REAPED);
940                                 goto reaped;
941                         }
942                         continue;
943                 }
944
945                 flags = WNOHANG;
946                 if (td->runstate == TD_EXITED)
947                         flags = 0;
948
949                 /*
950                  * check if someone quit or got killed in an unusual way
951                  */
952                 ret = waitpid(td->pid, &status, flags);
953                 if (ret < 0) {
954                         if (errno == ECHILD) {
955                                 log_err("fio: pid=%d disappeared %d\n", td->pid, td->runstate);
956                                 td_set_runstate(td, TD_REAPED);
957                                 goto reaped;
958                         }
959                         perror("waitpid");
960                 } else if (ret == td->pid) {
961                         if (WIFSIGNALED(status)) {
962                                 int sig = WTERMSIG(status);
963
964                                 if (sig != SIGQUIT)
965                                         log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
966                                 td_set_runstate(td, TD_REAPED);
967                                 goto reaped;
968                         }
969                         if (WIFEXITED(status)) {
970                                 if (WEXITSTATUS(status) && !td->error)
971                                         td->error = WEXITSTATUS(status);
972
973                                 td_set_runstate(td, TD_REAPED);
974                                 goto reaped;
975                         }
976                 }
977
978                 /*
979                  * thread is not dead, continue
980                  */
981                 continue;
982 reaped:
983                 if (td->o.use_thread) {
984                         long ret;
985
986                         if (pthread_join(td->thread, (void *) &ret))
987                                 perror("pthread_join");
988                 }
989
990                 (*nr_running)--;
991                 (*m_rate) -= td->o.ratemin;
992                 (*t_rate) -= td->o.rate;
993
994                 if (td->error)
995                         exit_value++;
996         }
997
998         if (*nr_running == cputhreads && !pending)
999                 terminate_threads(TERMINATE_ALL);
1000 }
1001
1002 /*
1003  * Main function for kicking off and reaping jobs, as needed.
1004  */
1005 static void run_threads(void)
1006 {
1007         struct thread_data *td;
1008         unsigned long spent;
1009         int i, todo, nr_running, m_rate, t_rate, nr_started;
1010
1011         if (fio_pin_memory())
1012                 return;
1013
1014         if (!terse_output) {
1015                 printf("Starting ");
1016                 if (nr_thread)
1017                         printf("%d thread%s", nr_thread, nr_thread > 1 ? "s" : "");
1018                 if (nr_process) {
1019                         if (nr_thread)
1020                                 printf(" and ");
1021                         printf("%d process%s", nr_process, nr_process > 1 ? "es" : "");
1022                 }
1023                 printf("\n");
1024                 fflush(stdout);
1025         }
1026
1027         signal(SIGINT, sig_handler);
1028         signal(SIGALRM, sig_handler);
1029
1030         todo = thread_number;
1031         nr_running = 0;
1032         nr_started = 0;
1033         m_rate = t_rate = 0;
1034
1035         for_each_td(td, i) {
1036                 print_status_init(td->thread_number - 1);
1037
1038                 if (!td->o.create_serialize) {
1039                         init_disk_util(td);
1040                         continue;
1041                 }
1042
1043                 /*
1044                  * do file setup here so it happens sequentially,
1045                  * we don't want X number of threads getting their
1046                  * client data interspersed on disk
1047                  */
1048                 if (setup_files(td)) {
1049                         exit_value++;
1050                         if (td->error)
1051                                 log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
1052                         td_set_runstate(td, TD_REAPED);
1053                         todo--;
1054                 }
1055
1056                 init_disk_util(td);
1057         }
1058
1059         set_genesis_time();
1060
1061         while (todo) {
1062                 struct thread_data *map[MAX_JOBS];
1063                 struct timeval this_start;
1064                 int this_jobs = 0, left;
1065
1066                 /*
1067                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1068                  */
1069                 for_each_td(td, i) {
1070                         if (td->runstate != TD_NOT_CREATED)
1071                                 continue;
1072
1073                         /*
1074                          * never got a chance to start, killed by other
1075                          * thread for some reason
1076                          */
1077                         if (td->terminate) {
1078                                 todo--;
1079                                 continue;
1080                         }
1081
1082                         if (td->o.start_delay) {
1083                                 spent = mtime_since_genesis();
1084
1085                                 if (td->o.start_delay * 1000 > spent)
1086                                         continue;
1087                         }
1088
1089                         if (td->o.stonewall && (nr_started || nr_running))
1090                                 break;
1091
1092                         /*
1093                          * Set state to created. Thread will transition
1094                          * to TD_INITIALIZED when it's done setting up.
1095                          */
1096                         td_set_runstate(td, TD_CREATED);
1097                         map[this_jobs++] = td;
1098                         nr_started++;
1099
1100                         if (td->o.use_thread) {
1101                                 if (pthread_create(&td->thread, NULL, thread_main, td)) {
1102                                         perror("thread_create");
1103                                         nr_started--;
1104                                         break;
1105                                 }
1106                         } else {
1107                                 if (!fork()) {
1108                                         int ret = fork_main(shm_id, i);
1109
1110                                         exit(ret);
1111                                 }
1112                         }
1113                         fio_sem_down(startup_sem);
1114                 }
1115
1116                 /*
1117                  * Wait for the started threads to transition to
1118                  * TD_INITIALIZED.
1119                  */
1120                 fio_gettime(&this_start, NULL);
1121                 left = this_jobs;
1122                 while (left && !fio_abort) {
1123                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1124                                 break;
1125
1126                         usleep(100000);
1127
1128                         for (i = 0; i < this_jobs; i++) {
1129                                 td = map[i];
1130                                 if (!td)
1131                                         continue;
1132                                 if (td->runstate == TD_INITIALIZED) {
1133                                         map[i] = NULL;
1134                                         left--;
1135                                 } else if (td->runstate >= TD_EXITED) {
1136                                         map[i] = NULL;
1137                                         left--;
1138                                         todo--;
1139                                         nr_running++; /* work-around... */
1140                                 }
1141                         }
1142                 }
1143
1144                 if (left) {
1145                         log_err("fio: %d jobs failed to start\n", left);
1146                         for (i = 0; i < this_jobs; i++) {
1147                                 td = map[i];
1148                                 if (!td)
1149                                         continue;
1150                                 kill(td->pid, SIGTERM);
1151                         }
1152                         break;
1153                 }
1154
1155                 /*
1156                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1157                  */
1158                 for_each_td(td, i) {
1159                         if (td->runstate != TD_INITIALIZED)
1160                                 continue;
1161
1162                         td_set_runstate(td, TD_RUNNING);
1163                         nr_running++;
1164                         nr_started--;
1165                         m_rate += td->o.ratemin;
1166                         t_rate += td->o.rate;
1167                         todo--;
1168                         fio_sem_up(td->mutex);
1169                 }
1170
1171                 reap_threads(&nr_running, &t_rate, &m_rate);
1172
1173                 if (todo)
1174                         usleep(100000);
1175         }
1176
1177         while (nr_running) {
1178                 reap_threads(&nr_running, &t_rate, &m_rate);
1179                 usleep(10000);
1180         }
1181
1182         update_io_ticks();
1183         fio_unpin_memory();
1184 }
1185
1186 int main(int argc, char *argv[])
1187 {
1188         long ps;
1189
1190         /*
1191          * We need locale for number printing, if it isn't set then just
1192          * go with the US format.
1193          */
1194         if (!getenv("LC_NUMERIC"))
1195                 setlocale(LC_NUMERIC, "en_US");
1196
1197         if (parse_options(argc, argv))
1198                 return 1;
1199
1200         if (!thread_number) {
1201                 log_err("Nothing to do\n");
1202                 return 1;
1203         }
1204
1205         ps = sysconf(_SC_PAGESIZE);
1206         if (ps < 0) {
1207                 log_err("Failed to get page size\n");
1208                 return 1;
1209         }
1210
1211         page_size = ps;
1212         page_mask = ps - 1;
1213
1214         if (write_bw_log) {
1215                 setup_log(&agg_io_log[DDIR_READ]);
1216                 setup_log(&agg_io_log[DDIR_WRITE]);
1217         }
1218
1219         startup_sem = fio_sem_init(0);
1220
1221         set_genesis_time();
1222
1223         disk_util_timer_arm();
1224
1225         run_threads();
1226
1227         if (!fio_abort) {
1228                 show_run_stats();
1229                 if (write_bw_log) {
1230                         __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1231                         __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");
1232                 }
1233         }
1234
1235         fio_sem_remove(startup_sem);
1236         return exit_value;
1237 }