Some more bits of blktrace support
[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         unsigned long long buf_size;
618         struct io_u *io_u;
619         unsigned int max_bs;
620         int i, max_units;
621         char *p;
622
623         if (td->io_ops->flags & FIO_SYNCIO)
624                 max_units = 1;
625         else
626                 max_units = td->o.iodepth;
627
628         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
629         buf_size = (unsigned long long) max_bs * (unsigned long long) max_units;
630         buf_size += page_mask;
631         if (buf_size != (size_t) buf_size) {
632                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
633                 return 1;
634         }
635
636         td->orig_buffer_size = buf_size;
637
638         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE)
639                 td->orig_buffer_size = (td->orig_buffer_size + td->o.hugepage_size - 1) & ~(td->o.hugepage_size - 1);
640         else if (td->orig_buffer_size & page_mask)
641                 td->orig_buffer_size = (td->orig_buffer_size + page_mask) & ~page_mask;
642
643         if (allocate_io_mem(td))
644                 return 1;
645
646         p = ALIGN(td->orig_buffer);
647         for (i = 0; i < max_units; i++) {
648                 if (td->terminate)
649                         return 1;
650                 io_u = malloc(sizeof(*io_u));
651                 memset(io_u, 0, sizeof(*io_u));
652                 INIT_LIST_HEAD(&io_u->list);
653
654                 io_u->buf = p + max_bs * i;
655
656                 if (td_write(td))
657                         fill_io_buf(td, io_u, max_bs);
658
659                 io_u->index = i;
660                 io_u->flags = IO_U_F_FREE;
661                 list_add(&io_u->list, &td->io_u_freelist);
662         }
663
664         io_u_init_timeout();
665
666         return 0;
667 }
668
669 static int switch_ioscheduler(struct thread_data *td)
670 {
671         char tmp[256], tmp2[128];
672         FILE *f;
673         int ret;
674
675         if (td->io_ops->flags & FIO_DISKLESSIO)
676                 return 0;
677
678         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
679
680         f = fopen(tmp, "r+");
681         if (!f) {
682                 if (errno == ENOENT) {
683                         log_err("fio: os or kernel doesn't support IO scheduler switching\n");
684                         return 0;
685                 }
686                 td_verror(td, errno, "fopen iosched");
687                 return 1;
688         }
689
690         /*
691          * Set io scheduler.
692          */
693         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
694         if (ferror(f) || ret != 1) {
695                 td_verror(td, errno, "fwrite");
696                 fclose(f);
697                 return 1;
698         }
699
700         rewind(f);
701
702         /*
703          * Read back and check that the selected scheduler is now the default.
704          */
705         ret = fread(tmp, 1, sizeof(tmp), f);
706         if (ferror(f) || ret < 0) {
707                 td_verror(td, errno, "fread");
708                 fclose(f);
709                 return 1;
710         }
711
712         sprintf(tmp2, "[%s]", td->o.ioscheduler);
713         if (!strstr(tmp, tmp2)) {
714                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
715                 td_verror(td, EINVAL, "iosched_switch");
716                 fclose(f);
717                 return 1;
718         }
719
720         fclose(f);
721         return 0;
722 }
723
724 static int clear_io_state(struct thread_data *td)
725 {
726         struct fio_file *f;
727         unsigned int i;
728         int ret;
729
730         td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
731         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
732         td->zone_bytes = 0;
733         td->rate_bytes = 0;
734         td->rate_blocks = 0;
735         td->rw_end_set[0] = td->rw_end_set[1] = 0;
736
737         td->last_was_sync = 0;
738
739         for_each_file(td, f, i)
740                 td_io_close_file(td, f);
741
742         ret = 0;
743         for_each_file(td, f, i) {
744                 ret = td_io_open_file(td, f);
745                 if (ret)
746                         break;
747         }
748
749         return ret;
750 }
751
752 /*
753  * Entry point for the thread based jobs. The process based jobs end up
754  * here as well, after a little setup.
755  */
756 static void *thread_main(void *data)
757 {
758         unsigned long long runtime[2];
759         struct thread_data *td = data;
760         unsigned long elapsed;
761         int clear_state;
762
763         if (!td->o.use_thread)
764                 setsid();
765
766         td->pid = getpid();
767
768         INIT_LIST_HEAD(&td->io_u_freelist);
769         INIT_LIST_HEAD(&td->io_u_busylist);
770         INIT_LIST_HEAD(&td->io_u_requeues);
771         INIT_LIST_HEAD(&td->io_log_list);
772         INIT_LIST_HEAD(&td->io_hist_list);
773         td->io_hist_tree = RB_ROOT;
774
775         td_set_runstate(td, TD_INITIALIZED);
776         fio_sem_up(startup_sem);
777         fio_sem_down(td->mutex);
778
779         /*
780          * the ->mutex semaphore is now no longer used, close it to avoid
781          * eating a file descriptor
782          */
783         fio_sem_remove(td->mutex);
784
785         if (init_io_u(td))
786                 goto err;
787
788         if (fio_setaffinity(td) == -1) {
789                 td_verror(td, errno, "cpu_set_affinity");
790                 goto err;
791         }
792
793         if (init_iolog(td))
794                 goto err;
795
796         if (td->ioprio) {
797                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
798                         td_verror(td, errno, "ioprio_set");
799                         goto err;
800                 }
801         }
802
803         if (nice(td->o.nice) == -1) {
804                 td_verror(td, errno, "nice");
805                 goto err;
806         }
807
808         if (td->o.ioscheduler && switch_ioscheduler(td))
809                 goto err;
810
811         if (!td->o.create_serialize && setup_files(td))
812                 goto err;
813
814         if (td_io_init(td))
815                 goto err;
816
817         if (open_files(td))
818                 goto err;
819
820         if (init_random_map(td))
821                 goto err;
822
823         if (td->o.exec_prerun) {
824                 if (system(td->o.exec_prerun) < 0)
825                         goto err;
826         }
827
828         fio_gettime(&td->epoch, NULL);
829         memcpy(&td->timeout_end, &td->epoch, sizeof(td->epoch));
830         getrusage(RUSAGE_SELF, &td->ts.ru_start);
831
832         runtime[0] = runtime[1] = 0;
833         clear_state = 0;
834         while (td->o.time_based || td->o.loops--) {
835                 fio_gettime(&td->start, NULL);
836                 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
837
838                 if (td->o.ratemin)
839                         memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
840
841                 if (clear_state && clear_io_state(td))
842                         break;
843
844                 prune_io_piece_log(td);
845
846                 do_io(td);
847
848                 clear_state = 1;
849
850                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
851                         if (td->rw_end_set[DDIR_READ])
852                                 elapsed = utime_since(&td->start, &td->rw_end[DDIR_READ]);
853                         else
854                                 elapsed = utime_since_now(&td->start);
855
856                         runtime[DDIR_READ] += elapsed;
857                 }
858                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
859                         if (td->rw_end_set[DDIR_WRITE])
860                                 elapsed = utime_since(&td->start, &td->rw_end[DDIR_WRITE]);
861                         else
862                                 elapsed = utime_since_now(&td->start);
863
864                         runtime[DDIR_WRITE] += elapsed;
865                 }
866                 
867                 if (td->error || td->terminate)
868                         break;
869
870                 if (td->o.verify == VERIFY_NONE)
871                         continue;
872
873                 if (clear_io_state(td))
874                         break;
875
876                 fio_gettime(&td->start, NULL);
877
878                 do_verify(td);
879
880                 runtime[DDIR_READ] += utime_since_now(&td->start);
881
882                 if (td->error || td->terminate)
883                         break;
884         }
885
886         update_rusage_stat(td);
887         td->ts.runtime[0] = runtime[0] / 1000;
888         td->ts.runtime[1] = runtime[1] / 1000;
889         td->ts.total_run_time = mtime_since_now(&td->epoch);
890         td->ts.io_bytes[0] = td->io_bytes[0];
891         td->ts.io_bytes[1] = td->io_bytes[1];
892
893         if (td->ts.bw_log)
894                 finish_log(td, td->ts.bw_log, "bw");
895         if (td->ts.slat_log)
896                 finish_log(td, td->ts.slat_log, "slat");
897         if (td->ts.clat_log)
898                 finish_log(td, td->ts.clat_log, "clat");
899         if (td->o.write_iolog_file)
900                 write_iolog_close(td);
901         if (td->o.exec_postrun) {
902                 if (system(td->o.exec_postrun) < 0)
903                         log_err("fio: postrun %s failed\n", td->o.exec_postrun);
904         }
905
906         if (exitall_on_terminate)
907                 terminate_threads(td->groupid);
908
909 err:
910         if (td->error)
911                 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
912         close_files(td);
913         close_ioengine(td);
914         cleanup_io_u(td);
915         options_mem_free(td);
916         td_set_runstate(td, TD_EXITED);
917         return (void *) (unsigned long) td->error;
918 }
919
920 /*
921  * We cannot pass the td data into a forked process, so attach the td and
922  * pass it to the thread worker.
923  */
924 static int fork_main(int shmid, int offset)
925 {
926         struct thread_data *td;
927         void *data, *ret;
928
929         data = shmat(shmid, NULL, 0);
930         if (data == (void *) -1) {
931                 int __err = errno;
932
933                 perror("shmat");
934                 return __err;
935         }
936
937         td = data + offset * sizeof(struct thread_data);
938         ret = thread_main(td);
939         shmdt(data);
940         return (int) (unsigned long) ret;
941 }
942
943 /*
944  * Run over the job map and reap the threads that have exited, if any.
945  */
946 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
947 {
948         struct thread_data *td;
949         int i, cputhreads, pending, status, ret;
950
951         /*
952          * reap exited threads (TD_EXITED -> TD_REAPED)
953          */
954         pending = cputhreads = 0;
955         for_each_td(td, i) {
956                 int flags = 0;
957
958                 /*
959                  * ->io_ops is NULL for a thread that has closed its
960                  * io engine
961                  */
962                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
963                         cputhreads++;
964
965                 if (!td->pid || td->runstate == TD_REAPED)
966                         continue;
967                 if (td->o.use_thread) {
968                         if (td->runstate == TD_EXITED) {
969                                 td_set_runstate(td, TD_REAPED);
970                                 goto reaped;
971                         }
972                         continue;
973                 }
974
975                 flags = WNOHANG;
976                 if (td->runstate == TD_EXITED)
977                         flags = 0;
978
979                 /*
980                  * check if someone quit or got killed in an unusual way
981                  */
982                 ret = waitpid(td->pid, &status, flags);
983                 if (ret < 0) {
984                         if (errno == ECHILD) {
985                                 log_err("fio: pid=%d disappeared %d\n", td->pid, td->runstate);
986                                 td_set_runstate(td, TD_REAPED);
987                                 goto reaped;
988                         }
989                         perror("waitpid");
990                 } else if (ret == td->pid) {
991                         if (WIFSIGNALED(status)) {
992                                 int sig = WTERMSIG(status);
993
994                                 if (sig != SIGQUIT)
995                                         log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
996                                 td_set_runstate(td, TD_REAPED);
997                                 goto reaped;
998                         }
999                         if (WIFEXITED(status)) {
1000                                 if (WEXITSTATUS(status) && !td->error)
1001                                         td->error = WEXITSTATUS(status);
1002
1003                                 td_set_runstate(td, TD_REAPED);
1004                                 goto reaped;
1005                         }
1006                 }
1007
1008                 /*
1009                  * thread is not dead, continue
1010                  */
1011                 pending++;
1012                 continue;
1013 reaped:
1014                 if (td->o.use_thread) {
1015                         long ret;
1016
1017                         if (pthread_join(td->thread, (void *) &ret))
1018                                 perror("pthread_join");
1019                 }
1020
1021                 (*nr_running)--;
1022                 (*m_rate) -= td->o.ratemin;
1023                 (*t_rate) -= td->o.rate;
1024                 pending--;
1025
1026                 if (td->error)
1027                         exit_value++;
1028         }
1029
1030         if (*nr_running == cputhreads && !pending)
1031                 terminate_threads(TERMINATE_ALL);
1032 }
1033
1034 /*
1035  * Main function for kicking off and reaping jobs, as needed.
1036  */
1037 static void run_threads(void)
1038 {
1039         struct thread_data *td;
1040         unsigned long spent;
1041         int i, todo, nr_running, m_rate, t_rate, nr_started;
1042
1043         if (fio_pin_memory())
1044                 return;
1045
1046         if (!terse_output) {
1047                 printf("Starting ");
1048                 if (nr_thread)
1049                         printf("%d thread%s", nr_thread, nr_thread > 1 ? "s" : "");
1050                 if (nr_process) {
1051                         if (nr_thread)
1052                                 printf(" and ");
1053                         printf("%d process%s", nr_process, nr_process > 1 ? "es" : "");
1054                 }
1055                 printf("\n");
1056                 fflush(stdout);
1057         }
1058
1059         signal(SIGINT, sig_handler);
1060         signal(SIGALRM, sig_handler);
1061
1062         todo = thread_number;
1063         nr_running = 0;
1064         nr_started = 0;
1065         m_rate = t_rate = 0;
1066
1067         for_each_td(td, i) {
1068                 print_status_init(td->thread_number - 1);
1069
1070                 if (!td->o.create_serialize) {
1071                         init_disk_util(td);
1072                         continue;
1073                 }
1074
1075                 /*
1076                  * do file setup here so it happens sequentially,
1077                  * we don't want X number of threads getting their
1078                  * client data interspersed on disk
1079                  */
1080                 if (setup_files(td)) {
1081                         exit_value++;
1082                         if (td->error)
1083                                 log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
1084                         td_set_runstate(td, TD_REAPED);
1085                         todo--;
1086                 }
1087
1088                 init_disk_util(td);
1089         }
1090
1091         set_genesis_time();
1092
1093         while (todo) {
1094                 struct thread_data *map[MAX_JOBS];
1095                 struct timeval this_start;
1096                 int this_jobs = 0, left;
1097
1098                 /*
1099                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1100                  */
1101                 for_each_td(td, i) {
1102                         if (td->runstate != TD_NOT_CREATED)
1103                                 continue;
1104
1105                         /*
1106                          * never got a chance to start, killed by other
1107                          * thread for some reason
1108                          */
1109                         if (td->terminate) {
1110                                 todo--;
1111                                 continue;
1112                         }
1113
1114                         if (td->o.start_delay) {
1115                                 spent = mtime_since_genesis();
1116
1117                                 if (td->o.start_delay * 1000 > spent)
1118                                         continue;
1119                         }
1120
1121                         if (td->o.stonewall && (nr_started || nr_running))
1122                                 break;
1123
1124                         /*
1125                          * Set state to created. Thread will transition
1126                          * to TD_INITIALIZED when it's done setting up.
1127                          */
1128                         td_set_runstate(td, TD_CREATED);
1129                         map[this_jobs++] = td;
1130                         nr_started++;
1131
1132                         if (td->o.use_thread) {
1133                                 if (pthread_create(&td->thread, NULL, thread_main, td)) {
1134                                         perror("thread_create");
1135                                         nr_started--;
1136                                         break;
1137                                 }
1138                         } else {
1139                                 if (!fork()) {
1140                                         int ret = fork_main(shm_id, i);
1141
1142                                         exit(ret);
1143                                 }
1144                         }
1145                         fio_sem_down(startup_sem);
1146                 }
1147
1148                 /*
1149                  * Wait for the started threads to transition to
1150                  * TD_INITIALIZED.
1151                  */
1152                 fio_gettime(&this_start, NULL);
1153                 left = this_jobs;
1154                 while (left && !fio_abort) {
1155                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1156                                 break;
1157
1158                         usleep(100000);
1159
1160                         for (i = 0; i < this_jobs; i++) {
1161                                 td = map[i];
1162                                 if (!td)
1163                                         continue;
1164                                 if (td->runstate == TD_INITIALIZED) {
1165                                         map[i] = NULL;
1166                                         left--;
1167                                 } else if (td->runstate >= TD_EXITED) {
1168                                         map[i] = NULL;
1169                                         left--;
1170                                         todo--;
1171                                         nr_running++; /* work-around... */
1172                                 }
1173                         }
1174                 }
1175
1176                 if (left) {
1177                         log_err("fio: %d jobs failed to start\n", left);
1178                         for (i = 0; i < this_jobs; i++) {
1179                                 td = map[i];
1180                                 if (!td)
1181                                         continue;
1182                                 kill(td->pid, SIGTERM);
1183                         }
1184                         break;
1185                 }
1186
1187                 /*
1188                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1189                  */
1190                 for_each_td(td, i) {
1191                         if (td->runstate != TD_INITIALIZED)
1192                                 continue;
1193
1194                         td_set_runstate(td, TD_RUNNING);
1195                         nr_running++;
1196                         nr_started--;
1197                         m_rate += td->o.ratemin;
1198                         t_rate += td->o.rate;
1199                         todo--;
1200                         fio_sem_up(td->mutex);
1201                 }
1202
1203                 reap_threads(&nr_running, &t_rate, &m_rate);
1204
1205                 if (todo)
1206                         usleep(100000);
1207         }
1208
1209         while (nr_running) {
1210                 reap_threads(&nr_running, &t_rate, &m_rate);
1211                 usleep(10000);
1212         }
1213
1214         update_io_ticks();
1215         fio_unpin_memory();
1216 }
1217
1218 int main(int argc, char *argv[])
1219 {
1220         long ps;
1221
1222         /*
1223          * We need locale for number printing, if it isn't set then just
1224          * go with the US format.
1225          */
1226         if (!getenv("LC_NUMERIC"))
1227                 setlocale(LC_NUMERIC, "en_US");
1228
1229         if (parse_options(argc, argv))
1230                 return 1;
1231
1232         if (!thread_number)
1233                 return 0;
1234
1235         ps = sysconf(_SC_PAGESIZE);
1236         if (ps < 0) {
1237                 log_err("Failed to get page size\n");
1238                 return 1;
1239         }
1240
1241         page_size = ps;
1242         page_mask = ps - 1;
1243
1244         if (write_bw_log) {
1245                 setup_log(&agg_io_log[DDIR_READ]);
1246                 setup_log(&agg_io_log[DDIR_WRITE]);
1247         }
1248
1249         startup_sem = fio_sem_init(0);
1250
1251         set_genesis_time();
1252
1253         disk_util_timer_arm();
1254
1255         run_threads();
1256
1257         if (!fio_abort) {
1258                 show_run_stats();
1259                 if (write_bw_log) {
1260                         __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1261                         __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");
1262                 }
1263         }
1264
1265         fio_sem_remove(startup_sem);
1266         return exit_value;
1267 }