ead118ec35a0d02572a5366f4fe065a904c0cf05
[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 "hash.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                         td->terminate = 1;
315                         break;
316                 }
317
318                 if (get_next_verify(td, io_u)) {
319                         put_io_u(td, io_u);
320                         break;
321                 }
322
323                 if (td_io_prep(td, io_u)) {
324                         put_io_u(td, io_u);
325                         break;
326                 }
327
328                 io_u->end_io = verify_io_u;
329
330                 ret = td_io_queue(td, io_u);
331                 switch (ret) {
332                 case FIO_Q_COMPLETED:
333                         if (io_u->error)
334                                 ret = -io_u->error;
335                         else if (io_u->resid) {
336                                 int bytes = io_u->xfer_buflen - io_u->resid;
337                                 struct fio_file *f = io_u->file;
338
339                                 /*
340                                  * zero read, fail
341                                  */
342                                 if (!bytes) {
343                                         td_verror(td, ENODATA, "full resid");
344                                         put_io_u(td, io_u);
345                                         break;
346                                 }
347
348                                 io_u->xfer_buflen = io_u->resid;
349                                 io_u->xfer_buf += bytes;
350                                 io_u->offset += bytes;
351                                 f->last_completed_pos = io_u->offset;
352
353                                 td->ts.short_io_u[io_u->ddir]++;
354
355                                 if (io_u->offset == f->real_file_size)
356                                         goto sync_done;
357
358                                 requeue_io_u(td, &io_u);
359                         } else {
360 sync_done:
361                                 ret = io_u_sync_complete(td, io_u);
362                                 if (ret < 0)
363                                         break;
364                         }
365                         continue;
366                 case FIO_Q_QUEUED:
367                         break;
368                 case FIO_Q_BUSY:
369                         requeue_io_u(td, &io_u);
370                         ret2 = td_io_commit(td);
371                         if (ret2 < 0)
372                                 ret = ret2;
373                         break;
374                 default:
375                         assert(ret < 0);
376                         td_verror(td, -ret, "td_io_queue");
377                         break;
378                 }
379
380                 if (ret < 0 || td->error)
381                         break;
382
383                 /*
384                  * if we can queue more, do so. but check if there are
385                  * completed io_u's first.
386                  */
387                 min_events = 0;
388                 if (queue_full(td) || ret == FIO_Q_BUSY) {
389                         min_events = 1;
390
391                         if (td->cur_depth > td->o.iodepth_low)
392                                 min_events = td->cur_depth - td->o.iodepth_low;
393                 }
394
395                 /*
396                  * Reap required number of io units, if any, and do the
397                  * verification on them through the callback handler
398                  */
399                 if (io_u_queued_complete(td, min_events) < 0)
400                         break;
401         }
402
403         if (!td->error) {
404                 min_events = td->cur_depth;
405
406                 if (min_events)
407                         ret = io_u_queued_complete(td, min_events);
408         } else
409                 cleanup_pending_aio(td);
410
411         td_set_runstate(td, TD_RUNNING);
412 }
413
414 /*
415  * Main IO worker function. It retrieves io_u's to process and queues
416  * and reaps them, checking for rate and errors along the way.
417  */
418 static void do_io(struct thread_data *td)
419 {
420         struct timeval s;
421         unsigned long usec;
422         unsigned int i;
423         int ret = 0;
424
425         td_set_runstate(td, TD_RUNNING);
426
427         while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) {
428                 struct timeval comp_time;
429                 long bytes_done = 0;
430                 int min_evts = 0;
431                 struct io_u *io_u;
432                 int ret2;
433
434                 if (td->terminate)
435                         break;
436
437                 io_u = get_io_u(td);
438                 if (!io_u)
439                         break;
440
441                 memcpy(&s, &io_u->start_time, sizeof(s));
442
443                 if (runtime_exceeded(td, &s)) {
444                         put_io_u(td, io_u);
445                         td->terminate = 1;
446                         break;
447                 }
448
449                 ret = td_io_queue(td, io_u);
450                 switch (ret) {
451                 case FIO_Q_COMPLETED:
452                         if (io_u->error)
453                                 ret = -io_u->error;
454                         else if (io_u->resid) {
455                                 int bytes = io_u->xfer_buflen - io_u->resid;
456                                 struct fio_file *f = io_u->file;
457
458                                 /*
459                                  * zero read, fail
460                                  */
461                                 if (!bytes) {
462                                         td_verror(td, ENODATA, "full resid");
463                                         put_io_u(td, io_u);
464                                         break;
465                                 }
466
467                                 io_u->xfer_buflen = io_u->resid;
468                                 io_u->xfer_buf += bytes;
469                                 io_u->offset += bytes;
470                                 f->last_completed_pos = io_u->offset;
471
472                                 td->ts.short_io_u[io_u->ddir]++;
473
474                                 if (io_u->offset == f->real_file_size)
475                                         goto sync_done;
476
477                                 requeue_io_u(td, &io_u);
478                         } else {
479 sync_done:
480                                 fio_gettime(&comp_time, NULL);
481                                 bytes_done = io_u_sync_complete(td, io_u);
482                                 if (bytes_done < 0)
483                                         ret = bytes_done;
484                         }
485                         break;
486                 case FIO_Q_QUEUED:
487                         /*
488                          * if the engine doesn't have a commit hook,
489                          * the io_u is really queued. if it does have such
490                          * a hook, it has to call io_u_queued() itself.
491                          */
492                         if (td->io_ops->commit == NULL)
493                                 io_u_queued(td, io_u);
494                         break;
495                 case FIO_Q_BUSY:
496                         requeue_io_u(td, &io_u);
497                         ret2 = td_io_commit(td);
498                         if (ret2 < 0)
499                                 ret = ret2;
500                         break;
501                 default:
502                         assert(ret < 0);
503                         put_io_u(td, io_u);
504                         break;
505                 }
506
507                 if (ret < 0 || td->error)
508                         break;
509
510                 /*
511                  * See if we need to complete some commands
512                  */
513                 if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
514                         min_evts = 0;
515                         if (queue_full(td) || ret == FIO_Q_BUSY) {
516                                 min_evts = 1;
517
518                                 if (td->cur_depth > td->o.iodepth_low)
519                                         min_evts = td->cur_depth - td->o.iodepth_low;
520                         }
521
522                         fio_gettime(&comp_time, NULL);
523                         bytes_done = io_u_queued_complete(td, min_evts);
524                         if (bytes_done < 0)
525                                 break;
526                 }
527
528                 if (!bytes_done)
529                         continue;
530
531                 /*
532                  * the rate is batched for now, it should work for batches
533                  * of completions except the very first one which may look
534                  * a little bursty
535                  */
536                 usec = utime_since(&s, &comp_time);
537
538                 rate_throttle(td, usec, bytes_done);
539
540                 if (check_min_rate(td, &comp_time)) {
541                         if (exitall_on_terminate)
542                                 terminate_threads(td->groupid);
543                         td_verror(td, ENODATA, "check_min_rate");
544                         break;
545                 }
546
547                 if (td->o.thinktime) {
548                         unsigned long long b;
549
550                         b = td->io_blocks[0] + td->io_blocks[1];
551                         if (!(b % td->o.thinktime_blocks)) {
552                                 int left;
553
554                                 if (td->o.thinktime_spin)
555                                         __usec_sleep(td->o.thinktime_spin);
556
557                                 left = td->o.thinktime - td->o.thinktime_spin;
558                                 if (left)
559                                         usec_sleep(td, left);
560                         }
561                 }
562         }
563
564         if (!td->error) {
565                 struct fio_file *f;
566
567                 i = td->cur_depth;
568                 if (i)
569                         ret = io_u_queued_complete(td, i);
570
571                 if (should_fsync(td) && td->o.end_fsync) {
572                         td_set_runstate(td, TD_FSYNCING);
573
574                         for_each_file(td, f, i) {
575                                 if (!(f->flags & FIO_FILE_OPEN))
576                                         continue;
577                                 fio_io_sync(td, f);
578                         }
579                 }
580         } else
581                 cleanup_pending_aio(td);
582 }
583
584 static void cleanup_io_u(struct thread_data *td)
585 {
586         struct list_head *entry, *n;
587         struct io_u *io_u;
588
589         list_for_each_safe(entry, n, &td->io_u_freelist) {
590                 io_u = list_entry(entry, struct io_u, list);
591
592                 list_del(&io_u->list);
593                 free(io_u);
594         }
595
596         free_io_mem(td);
597 }
598
599 /*
600  * "randomly" fill the buffer contents
601  */
602 static void fill_io_buf(struct thread_data *td, struct io_u *io_u, int max_bs)
603 {
604         long *ptr = io_u->buf;
605
606         if (!td->o.zero_buffers) {
607                 while ((void *) ptr - io_u->buf < max_bs) {
608                         *ptr = rand() * GOLDEN_RATIO_PRIME;
609                         ptr++;
610                 }
611         } else
612                 memset(ptr, 0, max_bs);
613 }
614
615 static int init_io_u(struct thread_data *td)
616 {
617         struct io_u *io_u;
618         unsigned int max_bs;
619         int i, max_units;
620         char *p;
621
622         if (td->io_ops->flags & FIO_SYNCIO)
623                 max_units = 1;
624         else
625                 max_units = td->o.iodepth;
626
627         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
628         td->orig_buffer_size = (unsigned long long) max_bs * (unsigned long long) max_units;
629
630         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE)
631                 td->orig_buffer_size = (td->orig_buffer_size + td->o.hugepage_size - 1) & ~(td->o.hugepage_size - 1);
632
633         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
634                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
635                 return 1;
636         }
637
638         if (allocate_io_mem(td))
639                 return 1;
640
641         p = ALIGN(td->orig_buffer);
642         for (i = 0; i < max_units; i++) {
643                 if (td->terminate)
644                         return 1;
645                 io_u = malloc(sizeof(*io_u));
646                 memset(io_u, 0, sizeof(*io_u));
647                 INIT_LIST_HEAD(&io_u->list);
648
649                 io_u->buf = p + max_bs * i;
650
651                 if (td_write(td))
652                         fill_io_buf(td, io_u, max_bs);
653
654                 io_u->index = i;
655                 io_u->flags = IO_U_F_FREE;
656                 list_add(&io_u->list, &td->io_u_freelist);
657         }
658
659         io_u_init_timeout();
660
661         return 0;
662 }
663
664 static int switch_ioscheduler(struct thread_data *td)
665 {
666         char tmp[256], tmp2[128];
667         FILE *f;
668         int ret;
669
670         if (td->io_ops->flags & FIO_DISKLESSIO)
671                 return 0;
672
673         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
674
675         f = fopen(tmp, "r+");
676         if (!f) {
677                 if (errno == ENOENT) {
678                         log_err("fio: os or kernel doesn't support IO scheduler switching\n");
679                         return 0;
680                 }
681                 td_verror(td, errno, "fopen iosched");
682                 return 1;
683         }
684
685         /*
686          * Set io scheduler.
687          */
688         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
689         if (ferror(f) || ret != 1) {
690                 td_verror(td, errno, "fwrite");
691                 fclose(f);
692                 return 1;
693         }
694
695         rewind(f);
696
697         /*
698          * Read back and check that the selected scheduler is now the default.
699          */
700         ret = fread(tmp, 1, sizeof(tmp), f);
701         if (ferror(f) || ret < 0) {
702                 td_verror(td, errno, "fread");
703                 fclose(f);
704                 return 1;
705         }
706
707         sprintf(tmp2, "[%s]", td->o.ioscheduler);
708         if (!strstr(tmp, tmp2)) {
709                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
710                 td_verror(td, EINVAL, "iosched_switch");
711                 fclose(f);
712                 return 1;
713         }
714
715         fclose(f);
716         return 0;
717 }
718
719 static int keep_running(struct thread_data *td)
720 {
721         unsigned long long io_done;
722
723         if (td->o.time_based)
724                 return 1;
725         if (td->o.loops) {
726                 td->o.loops--;
727                 return 1;
728         }
729
730         io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE];
731         if (io_done < td->o.size)
732                 return 1;
733
734         return 0;
735 }
736
737 static int clear_io_state(struct thread_data *td)
738 {
739         struct fio_file *f;
740         unsigned int i;
741         int ret;
742
743         td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
744         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
745         td->zone_bytes = 0;
746         td->rate_bytes = 0;
747         td->rate_blocks = 0;
748         td->rw_end_set[0] = td->rw_end_set[1] = 0;
749
750         td->last_was_sync = 0;
751
752         td->nr_done_files = 0;
753
754         for_each_file(td, f, i)
755                 td_io_close_file(td, f);
756
757         ret = 0;
758         for_each_file(td, f, i) {
759                 f->flags &= ~FIO_FILE_DONE;
760                 ret = td_io_open_file(td, f);
761                 if (ret)
762                         break;
763         }
764
765         return ret;
766 }
767
768 /*
769  * Entry point for the thread based jobs. The process based jobs end up
770  * here as well, after a little setup.
771  */
772 static void *thread_main(void *data)
773 {
774         unsigned long long runtime[2];
775         struct thread_data *td = data;
776         unsigned long elapsed;
777         int clear_state;
778
779         if (!td->o.use_thread)
780                 setsid();
781
782         td->pid = getpid();
783
784         INIT_LIST_HEAD(&td->io_u_freelist);
785         INIT_LIST_HEAD(&td->io_u_busylist);
786         INIT_LIST_HEAD(&td->io_u_requeues);
787         INIT_LIST_HEAD(&td->io_log_list);
788         INIT_LIST_HEAD(&td->io_hist_list);
789         td->io_hist_tree = RB_ROOT;
790
791         td_set_runstate(td, TD_INITIALIZED);
792         fio_sem_up(startup_sem);
793         fio_sem_down(td->mutex);
794
795         /*
796          * the ->mutex semaphore is now no longer used, close it to avoid
797          * eating a file descriptor
798          */
799         fio_sem_remove(td->mutex);
800
801         /*
802          * May alter parameters that init_io_u() will use, so we need to
803          * do this first.
804          */
805         if (init_iolog(td))
806                 goto err;
807
808         if (init_io_u(td))
809                 goto err;
810
811         if (td->o.cpumask_set && fio_setaffinity(td) == -1) {
812                 td_verror(td, errno, "cpu_set_affinity");
813                 goto err;
814         }
815
816         if (td->ioprio) {
817                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
818                         td_verror(td, errno, "ioprio_set");
819                         goto err;
820                 }
821         }
822
823         if (nice(td->o.nice) == -1) {
824                 td_verror(td, errno, "nice");
825                 goto err;
826         }
827
828         if (td->o.ioscheduler && switch_ioscheduler(td))
829                 goto err;
830
831         if (!td->o.create_serialize && setup_files(td))
832                 goto err;
833
834         if (td_io_init(td))
835                 goto err;
836
837         if (open_files(td))
838                 goto err;
839
840         if (init_random_map(td))
841                 goto err;
842
843         if (td->o.exec_prerun) {
844                 if (system(td->o.exec_prerun) < 0)
845                         goto err;
846         }
847
848         fio_gettime(&td->epoch, NULL);
849         memcpy(&td->timeout_end, &td->epoch, sizeof(td->epoch));
850         getrusage(RUSAGE_SELF, &td->ts.ru_start);
851
852         runtime[0] = runtime[1] = 0;
853         clear_state = 0;
854         while (keep_running(td)) {
855                 fio_gettime(&td->start, NULL);
856                 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
857
858                 if (td->o.ratemin)
859                         memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
860
861                 if (clear_state && clear_io_state(td))
862                         break;
863
864                 prune_io_piece_log(td);
865
866                 do_io(td);
867
868                 clear_state = 1;
869
870                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
871                         if (td->rw_end_set[DDIR_READ])
872                                 elapsed = utime_since(&td->start, &td->rw_end[DDIR_READ]);
873                         else
874                                 elapsed = utime_since_now(&td->start);
875
876                         runtime[DDIR_READ] += elapsed;
877                 }
878                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
879                         if (td->rw_end_set[DDIR_WRITE])
880                                 elapsed = utime_since(&td->start, &td->rw_end[DDIR_WRITE]);
881                         else
882                                 elapsed = utime_since_now(&td->start);
883
884                         runtime[DDIR_WRITE] += elapsed;
885                 }
886                 
887                 if (td->error || td->terminate)
888                         break;
889
890                 if (td->o.verify == VERIFY_NONE)
891                         continue;
892
893                 if (clear_io_state(td))
894                         break;
895
896                 fio_gettime(&td->start, NULL);
897
898                 do_verify(td);
899
900                 runtime[DDIR_READ] += utime_since_now(&td->start);
901
902                 if (td->error || td->terminate)
903                         break;
904         }
905
906         update_rusage_stat(td);
907         td->ts.runtime[0] = runtime[0] / 1000;
908         td->ts.runtime[1] = runtime[1] / 1000;
909         td->ts.total_run_time = mtime_since_now(&td->epoch);
910         td->ts.io_bytes[0] = td->io_bytes[0];
911         td->ts.io_bytes[1] = td->io_bytes[1];
912
913         if (td->ts.bw_log)
914                 finish_log(td, td->ts.bw_log, "bw");
915         if (td->ts.slat_log)
916                 finish_log(td, td->ts.slat_log, "slat");
917         if (td->ts.clat_log)
918                 finish_log(td, td->ts.clat_log, "clat");
919         if (td->o.write_iolog_file)
920                 write_iolog_close(td);
921         if (td->o.exec_postrun) {
922                 if (system(td->o.exec_postrun) < 0)
923                         log_err("fio: postrun %s failed\n", td->o.exec_postrun);
924         }
925
926         if (exitall_on_terminate)
927                 terminate_threads(td->groupid);
928
929 err:
930         if (td->error)
931                 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
932         close_files(td);
933         close_ioengine(td);
934         cleanup_io_u(td);
935         options_mem_free(td);
936         td_set_runstate(td, TD_EXITED);
937         return (void *) (unsigned long) td->error;
938 }
939
940 /*
941  * We cannot pass the td data into a forked process, so attach the td and
942  * pass it to the thread worker.
943  */
944 static int fork_main(int shmid, int offset)
945 {
946         struct thread_data *td;
947         void *data, *ret;
948
949         data = shmat(shmid, NULL, 0);
950         if (data == (void *) -1) {
951                 int __err = errno;
952
953                 perror("shmat");
954                 return __err;
955         }
956
957         td = data + offset * sizeof(struct thread_data);
958         ret = thread_main(td);
959         shmdt(data);
960         return (int) (unsigned long) ret;
961 }
962
963 /*
964  * Run over the job map and reap the threads that have exited, if any.
965  */
966 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
967 {
968         struct thread_data *td;
969         int i, cputhreads, pending, status, ret;
970
971         /*
972          * reap exited threads (TD_EXITED -> TD_REAPED)
973          */
974         pending = cputhreads = 0;
975         for_each_td(td, i) {
976                 int flags = 0;
977
978                 /*
979                  * ->io_ops is NULL for a thread that has closed its
980                  * io engine
981                  */
982                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
983                         cputhreads++;
984
985                 if (!td->pid || td->runstate == TD_REAPED)
986                         continue;
987                 if (td->o.use_thread) {
988                         if (td->runstate == TD_EXITED) {
989                                 td_set_runstate(td, TD_REAPED);
990                                 goto reaped;
991                         }
992                         continue;
993                 }
994
995                 flags = WNOHANG;
996                 if (td->runstate == TD_EXITED)
997                         flags = 0;
998
999                 /*
1000                  * check if someone quit or got killed in an unusual way
1001                  */
1002                 ret = waitpid(td->pid, &status, flags);
1003                 if (ret < 0) {
1004                         if (errno == ECHILD) {
1005                                 log_err("fio: pid=%d disappeared %d\n", td->pid, td->runstate);
1006                                 td_set_runstate(td, TD_REAPED);
1007                                 goto reaped;
1008                         }
1009                         perror("waitpid");
1010                 } else if (ret == td->pid) {
1011                         if (WIFSIGNALED(status)) {
1012                                 int sig = WTERMSIG(status);
1013
1014                                 if (sig != SIGQUIT)
1015                                         log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
1016                                 td_set_runstate(td, TD_REAPED);
1017                                 goto reaped;
1018                         }
1019                         if (WIFEXITED(status)) {
1020                                 if (WEXITSTATUS(status) && !td->error)
1021                                         td->error = WEXITSTATUS(status);
1022
1023                                 td_set_runstate(td, TD_REAPED);
1024                                 goto reaped;
1025                         }
1026                 }
1027
1028                 /*
1029                  * thread is not dead, continue
1030                  */
1031                 pending++;
1032                 continue;
1033 reaped:
1034                 if (td->o.use_thread) {
1035                         long ret;
1036
1037                         if (pthread_join(td->thread, (void *) &ret))
1038                                 perror("pthread_join");
1039                 }
1040
1041                 (*nr_running)--;
1042                 (*m_rate) -= td->o.ratemin;
1043                 (*t_rate) -= td->o.rate;
1044                 pending--;
1045
1046                 if (td->error)
1047                         exit_value++;
1048         }
1049
1050         if (*nr_running == cputhreads && !pending)
1051                 terminate_threads(TERMINATE_ALL);
1052 }
1053
1054 /*
1055  * Main function for kicking off and reaping jobs, as needed.
1056  */
1057 static void run_threads(void)
1058 {
1059         struct thread_data *td;
1060         unsigned long spent;
1061         int i, todo, nr_running, m_rate, t_rate, nr_started;
1062
1063         if (fio_pin_memory())
1064                 return;
1065
1066         if (!terse_output) {
1067                 printf("Starting ");
1068                 if (nr_thread)
1069                         printf("%d thread%s", nr_thread, nr_thread > 1 ? "s" : "");
1070                 if (nr_process) {
1071                         if (nr_thread)
1072                                 printf(" and ");
1073                         printf("%d process%s", nr_process, nr_process > 1 ? "es" : "");
1074                 }
1075                 printf("\n");
1076                 fflush(stdout);
1077         }
1078
1079         signal(SIGINT, sig_handler);
1080         signal(SIGALRM, sig_handler);
1081
1082         todo = thread_number;
1083         nr_running = 0;
1084         nr_started = 0;
1085         m_rate = t_rate = 0;
1086
1087         for_each_td(td, i) {
1088                 print_status_init(td->thread_number - 1);
1089
1090                 if (!td->o.create_serialize) {
1091                         init_disk_util(td);
1092                         continue;
1093                 }
1094
1095                 /*
1096                  * do file setup here so it happens sequentially,
1097                  * we don't want X number of threads getting their
1098                  * client data interspersed on disk
1099                  */
1100                 if (setup_files(td)) {
1101                         exit_value++;
1102                         if (td->error)
1103                                 log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
1104                         td_set_runstate(td, TD_REAPED);
1105                         todo--;
1106                 }
1107
1108                 init_disk_util(td);
1109         }
1110
1111         set_genesis_time();
1112
1113         while (todo) {
1114                 struct thread_data *map[MAX_JOBS];
1115                 struct timeval this_start;
1116                 int this_jobs = 0, left;
1117
1118                 /*
1119                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1120                  */
1121                 for_each_td(td, i) {
1122                         if (td->runstate != TD_NOT_CREATED)
1123                                 continue;
1124
1125                         /*
1126                          * never got a chance to start, killed by other
1127                          * thread for some reason
1128                          */
1129                         if (td->terminate) {
1130                                 todo--;
1131                                 continue;
1132                         }
1133
1134                         if (td->o.start_delay) {
1135                                 spent = mtime_since_genesis();
1136
1137                                 if (td->o.start_delay * 1000 > spent)
1138                                         continue;
1139                         }
1140
1141                         if (td->o.stonewall && (nr_started || nr_running))
1142                                 break;
1143
1144                         /*
1145                          * Set state to created. Thread will transition
1146                          * to TD_INITIALIZED when it's done setting up.
1147                          */
1148                         td_set_runstate(td, TD_CREATED);
1149                         map[this_jobs++] = td;
1150                         nr_started++;
1151
1152                         if (td->o.use_thread) {
1153                                 if (pthread_create(&td->thread, NULL, thread_main, td)) {
1154                                         perror("thread_create");
1155                                         nr_started--;
1156                                         break;
1157                                 }
1158                         } else {
1159                                 if (!fork()) {
1160                                         int ret = fork_main(shm_id, i);
1161
1162                                         exit(ret);
1163                                 }
1164                         }
1165                         fio_sem_down(startup_sem);
1166                 }
1167
1168                 /*
1169                  * Wait for the started threads to transition to
1170                  * TD_INITIALIZED.
1171                  */
1172                 fio_gettime(&this_start, NULL);
1173                 left = this_jobs;
1174                 while (left && !fio_abort) {
1175                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1176                                 break;
1177
1178                         usleep(100000);
1179
1180                         for (i = 0; i < this_jobs; i++) {
1181                                 td = map[i];
1182                                 if (!td)
1183                                         continue;
1184                                 if (td->runstate == TD_INITIALIZED) {
1185                                         map[i] = NULL;
1186                                         left--;
1187                                 } else if (td->runstate >= TD_EXITED) {
1188                                         map[i] = NULL;
1189                                         left--;
1190                                         todo--;
1191                                         nr_running++; /* work-around... */
1192                                 }
1193                         }
1194                 }
1195
1196                 if (left) {
1197                         log_err("fio: %d jobs failed to start\n", left);
1198                         for (i = 0; i < this_jobs; i++) {
1199                                 td = map[i];
1200                                 if (!td)
1201                                         continue;
1202                                 kill(td->pid, SIGTERM);
1203                         }
1204                         break;
1205                 }
1206
1207                 /*
1208                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1209                  */
1210                 for_each_td(td, i) {
1211                         if (td->runstate != TD_INITIALIZED)
1212                                 continue;
1213
1214                         td_set_runstate(td, TD_RUNNING);
1215                         nr_running++;
1216                         nr_started--;
1217                         m_rate += td->o.ratemin;
1218                         t_rate += td->o.rate;
1219                         todo--;
1220                         fio_sem_up(td->mutex);
1221                 }
1222
1223                 reap_threads(&nr_running, &t_rate, &m_rate);
1224
1225                 if (todo)
1226                         usleep(100000);
1227         }
1228
1229         while (nr_running) {
1230                 reap_threads(&nr_running, &t_rate, &m_rate);
1231                 usleep(10000);
1232         }
1233
1234         update_io_ticks();
1235         fio_unpin_memory();
1236 }
1237
1238 int main(int argc, char *argv[])
1239 {
1240         long ps;
1241
1242         /*
1243          * We need locale for number printing, if it isn't set then just
1244          * go with the US format.
1245          */
1246         if (!getenv("LC_NUMERIC"))
1247                 setlocale(LC_NUMERIC, "en_US");
1248
1249         if (parse_options(argc, argv))
1250                 return 1;
1251
1252         if (!thread_number)
1253                 return 0;
1254
1255         ps = sysconf(_SC_PAGESIZE);
1256         if (ps < 0) {
1257                 log_err("Failed to get page size\n");
1258                 return 1;
1259         }
1260
1261         page_size = ps;
1262         page_mask = ps - 1;
1263
1264         if (write_bw_log) {
1265                 setup_log(&agg_io_log[DDIR_READ]);
1266                 setup_log(&agg_io_log[DDIR_WRITE]);
1267         }
1268
1269         startup_sem = fio_sem_init(0);
1270
1271         set_genesis_time();
1272
1273         disk_util_timer_arm();
1274
1275         run_threads();
1276
1277         if (!fio_abort) {
1278                 show_run_stats();
1279                 if (write_bw_log) {
1280                         __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1281                         __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");
1282                 }
1283         }
1284
1285         fio_sem_remove(startup_sem);
1286         return exit_value;
1287 }