Dump io_u on timeout
[fio.git] / fio.c
1 /*
2  * fio - the flexible io tester
3  *
4  * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
5  * Copyright (C) 2006 Jens Axboe <axboe@kernel.dk>
6  *
7  * The license below covers all files distributed with fio unless otherwise
8  * noted in the file itself.
9  *
10  *  This program is free software; you can redistribute it and/or modify
11  *  it under the terms of the GNU General Public License version 2 as
12  *  published by the Free Software Foundation.
13  *
14  *  This program is distributed in the hope that it will be useful,
15  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
16  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
17  *  GNU General Public License for more details.
18  *
19  *  You should have received a copy of the GNU General Public License
20  *  along with this program; if not, write to the Free Software
21  *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
22  *
23  */
24 #include <unistd.h>
25 #include <fcntl.h>
26 #include <string.h>
27 #include <signal.h>
28 #include <time.h>
29 #include <locale.h>
30 #include <assert.h>
31 #include <sys/stat.h>
32 #include <sys/wait.h>
33 #include <sys/ipc.h>
34 #include <sys/shm.h>
35 #include <sys/mman.h>
36
37 #include "fio.h"
38 #include "os.h"
39
40 static unsigned long page_mask;
41 #define ALIGN(buf)      \
42         (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
43
44 int groupid = 0;
45 int thread_number = 0;
46 int shm_id = 0;
47 int temp_stall_ts;
48
49 static volatile int startup_sem;
50 static volatile int fio_abort;
51 static int exit_value;
52
53 struct io_log *agg_io_log[2];
54
55 #define TERMINATE_ALL           (-1)
56 #define JOB_START_TIMEOUT       (5 * 1000)
57
58 static inline void td_set_runstate(struct thread_data *td, int runstate)
59 {
60         td->runstate = runstate;
61 }
62
63 static void terminate_threads(int group_id, int forced_kill)
64 {
65         struct thread_data *td;
66         int i;
67
68         for_each_td(td, i) {
69                 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
70                         td->terminate = 1;
71                         td->start_delay = 0;
72                         if (forced_kill)
73                                 td_set_runstate(td, TD_EXITED);
74                 }
75         }
76 }
77
78 static void sig_handler(int sig)
79 {
80         switch (sig) {
81                 case SIGALRM:
82                         update_io_ticks();
83                         disk_util_timer_arm();
84                         print_thread_status();
85                         break;
86                 default:
87                         printf("\nfio: terminating on signal %d\n", sig);
88                         fflush(stdout);
89                         terminate_threads(TERMINATE_ALL, 0);
90                         break;
91         }
92 }
93
94 /*
95  * Check if we are above the minimum rate given.
96  */
97 static int check_min_rate(struct thread_data *td, struct timeval *now)
98 {
99         unsigned long spent;
100         unsigned long rate;
101         int ddir = td->ddir;
102
103         /*
104          * allow a 2 second settle period in the beginning
105          */
106         if (mtime_since(&td->start, now) < 2000)
107                 return 0;
108
109         /*
110          * if rate blocks is set, sample is running
111          */
112         if (td->rate_bytes) {
113                 spent = mtime_since(&td->lastrate, now);
114                 if (spent < td->ratecycle)
115                         return 0;
116
117                 rate = (td->this_io_bytes[ddir] - td->rate_bytes) / spent;
118                 if (rate < td->ratemin) {
119                         fprintf(f_out, "%s: min rate %u not met, got %luKiB/sec\n", td->name, td->ratemin, rate);
120                         return 1;
121                 }
122         }
123
124         td->rate_bytes = td->this_io_bytes[ddir];
125         memcpy(&td->lastrate, now, sizeof(*now));
126         return 0;
127 }
128
129 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
130 {
131         if (!td->timeout)
132                 return 0;
133         if (mtime_since(&td->epoch, t) >= td->timeout * 1000)
134                 return 1;
135
136         return 0;
137 }
138
139 /*
140  * When job exits, we can cancel the in-flight IO if we are using async
141  * io. Attempt to do so.
142  */
143 static void cleanup_pending_aio(struct thread_data *td)
144 {
145         struct list_head *entry, *n;
146         struct io_u *io_u;
147         int r;
148
149         /*
150          * get immediately available events, if any
151          */
152         r = io_u_queued_complete(td, 0, NULL);
153         if (r < 0)
154                 return;
155
156         /*
157          * now cancel remaining active events
158          */
159         if (td->io_ops->cancel) {
160                 list_for_each_safe(entry, n, &td->io_u_busylist) {
161                         io_u = list_entry(entry, struct io_u, list);
162
163                         /*
164                          * if the io_u isn't in flight, then that generally
165                          * means someone leaked an io_u. complain but fix
166                          * it up, so we don't stall here.
167                          */
168                         if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
169                                 log_err("fio: non-busy IO on busy list\n");
170                                 put_io_u(td, io_u);
171                         } else {
172                                 r = td->io_ops->cancel(td, io_u);
173                                 if (!r)
174                                         put_io_u(td, io_u);
175                         }
176                 }
177         }
178
179         if (td->cur_depth)
180                 r = io_u_queued_complete(td, td->cur_depth, NULL);
181 }
182
183 /*
184  * Helper to handle the final sync of a file. Works just like the normal
185  * io path, just does everything sync.
186  */
187 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
188 {
189         struct io_u *io_u = __get_io_u(td);
190         int ret;
191
192         if (!io_u)
193                 return 1;
194
195         io_u->ddir = DDIR_SYNC;
196         io_u->file = f;
197
198         if (td_io_prep(td, io_u)) {
199                 put_io_u(td, io_u);
200                 return 1;
201         }
202
203 requeue:
204         ret = td_io_queue(td, io_u);
205         if (ret < 0) {
206                 td_verror(td, io_u->error, "td_io_queue");
207                 put_io_u(td, io_u);
208                 return 1;
209         } else if (ret == FIO_Q_QUEUED) {
210                 if (io_u_queued_complete(td, 1, NULL) < 0)
211                         return 1;
212         } else if (ret == FIO_Q_COMPLETED) {
213                 if (io_u->error) {
214                         td_verror(td, io_u->error, "td_io_queue");
215                         return 1;
216                 }
217
218                 if (io_u_sync_complete(td, io_u, NULL) < 0)
219                         return 1;
220         } else if (ret == FIO_Q_BUSY) {
221                 if (td_io_commit(td))
222                         return 1;
223                 goto requeue;
224         }
225
226         return 0;
227 }
228
229 /*
230  * The main verify engine. Runs over the writes we previusly submitted,
231  * reads the blocks back in, and checks the crc/md5 of the data.
232  */
233 static void do_verify(struct thread_data *td)
234 {
235         struct fio_file *f;
236         struct io_u *io_u;
237         int ret, i, min_events;
238
239         /*
240          * sync io first and invalidate cache, to make sure we really
241          * read from disk.
242          */
243         for_each_file(td, f, i) {
244                 if (fio_io_sync(td, f))
245                         break;
246                 if (file_invalidate_cache(td, f))
247                         break;
248         }
249
250         if (td->error)
251                 return;
252
253         td_set_runstate(td, TD_VERIFYING);
254
255         io_u = NULL;
256         while (!td->terminate) {
257                 io_u = __get_io_u(td);
258                 if (!io_u)
259                         break;
260
261                 if (runtime_exceeded(td, &io_u->start_time)) {
262                         put_io_u(td, io_u);
263                         break;
264                 }
265
266                 if (get_next_verify(td, io_u)) {
267                         put_io_u(td, io_u);
268                         break;
269                 }
270
271                 if (td_io_prep(td, io_u)) {
272                         put_io_u(td, io_u);
273                         break;
274                 }
275 requeue:
276                 ret = td_io_queue(td, io_u);
277
278                 switch (ret) {
279                 case FIO_Q_COMPLETED:
280                         if (io_u->error)
281                                 ret = -io_u->error;
282                         if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
283                                 int bytes = io_u->xfer_buflen - io_u->resid;
284
285                                 io_u->xfer_buflen = io_u->resid;
286                                 io_u->xfer_buf += bytes;
287                                 goto requeue;
288                         }
289                         ret = io_u_sync_complete(td, io_u, verify_io_u);
290                         if (ret < 0)
291                                 break;
292                         continue;
293                 case FIO_Q_QUEUED:
294                         break;
295                 case FIO_Q_BUSY:
296                         requeue_io_u(td, &io_u);
297                         ret = td_io_commit(td);
298                         break;
299                 default:
300                         assert(ret < 0);
301                         td_verror(td, -ret, "td_io_queue");
302                         break;
303                 }
304
305                 if (ret < 0 || td->error)
306                         break;
307
308                 /*
309                  * if we can queue more, do so. but check if there are
310                  * completed io_u's first.
311                  */
312                 min_events = 0;
313                 if (queue_full(td) || ret == FIO_Q_BUSY) {
314                         min_events = 1;
315
316                         if (td->cur_depth > td->iodepth_low)
317                                 min_events = td->cur_depth - td->iodepth_low;
318                 }
319
320                 /*
321                  * Reap required number of io units, if any, and do the
322                  * verification on them through the callback handler
323                  */
324                 if (io_u_queued_complete(td, min_events, verify_io_u) < 0)
325                         break;
326         }
327
328         if (td->cur_depth)
329                 cleanup_pending_aio(td);
330
331         td_set_runstate(td, TD_RUNNING);
332 }
333
334 /*
335  * Not really an io thread, all it does is burn CPU cycles in the specified
336  * manner.
337  */
338 static void do_cpuio(struct thread_data *td)
339 {
340         struct timeval e;
341         int split = 100 / td->cpuload;
342         int i = 0;
343
344         while (!td->terminate) {
345                 fio_gettime(&e, NULL);
346
347                 if (runtime_exceeded(td, &e))
348                         break;
349
350                 if (!(i % split))
351                         __usec_sleep(10000);
352                 else
353                         usec_sleep(td, 10000);
354
355                 i++;
356         }
357 }
358
359 /*
360  * Main IO worker function. It retrieves io_u's to process and queues
361  * and reaps them, checking for rate and errors along the way.
362  */
363 static void do_io(struct thread_data *td)
364 {
365         struct timeval s;
366         unsigned long usec;
367         int i, ret = 0;
368
369         td_set_runstate(td, TD_RUNNING);
370
371         while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->io_size) {
372                 struct timeval comp_time;
373                 long bytes_done = 0;
374                 int min_evts = 0;
375                 struct io_u *io_u;
376
377                 if (td->terminate)
378                         break;
379
380                 io_u = get_io_u(td);
381                 if (!io_u)
382                         break;
383
384                 memcpy(&s, &io_u->start_time, sizeof(s));
385
386                 if (runtime_exceeded(td, &s)) {
387                         put_io_u(td, io_u);
388                         break;
389                 }
390 requeue:
391                 ret = td_io_queue(td, io_u);
392
393                 switch (ret) {
394                 case FIO_Q_COMPLETED:
395                         if (io_u->error) {
396                                 ret = io_u->error;
397                                 break;
398                         }
399                         if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
400                                 int bytes = io_u->xfer_buflen - io_u->resid;
401
402                                 io_u->xfer_buflen = io_u->resid;
403                                 io_u->xfer_buf += bytes;
404                                 goto requeue;
405                         }
406                         fio_gettime(&comp_time, NULL);
407                         bytes_done = io_u_sync_complete(td, io_u, NULL);
408                         if (bytes_done < 0)
409                                 ret = bytes_done;
410                         break;
411                 case FIO_Q_QUEUED:
412                         /*
413                          * if the engine doesn't have a commit hook,
414                          * the io_u is really queued. if it does have such
415                          * a hook, it has to call io_u_queued() itself.
416                          */
417                         if (td->io_ops->commit == NULL)
418                                 io_u_queued(td, io_u);
419                         break;
420                 case FIO_Q_BUSY:
421                         requeue_io_u(td, &io_u);
422                         ret = td_io_commit(td);
423                         break;
424                 default:
425                         assert(ret < 0);
426                         put_io_u(td, io_u);
427                         break;
428                 }
429
430                 if (ret < 0 || td->error)
431                         break;
432
433                 /*
434                  * See if we need to complete some commands
435                  */
436                 if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
437                         min_evts = 0;
438                         if (queue_full(td) || ret == FIO_Q_BUSY) {
439                                 min_evts = 1;
440
441                                 if (td->cur_depth > td->iodepth_low)
442                                         min_evts = td->cur_depth - td->iodepth_low;
443                         }
444
445                         fio_gettime(&comp_time, NULL);
446                         bytes_done = io_u_queued_complete(td, min_evts, NULL);
447                         if (bytes_done < 0)
448                                 break;
449                 }
450
451                 if (!bytes_done)
452                         continue;
453
454                 /*
455                  * the rate is batched for now, it should work for batches
456                  * of completions except the very first one which may look
457                  * a little bursty
458                  */
459                 usec = utime_since(&s, &comp_time);
460
461                 rate_throttle(td, usec, bytes_done, td->ddir);
462
463                 if (check_min_rate(td, &comp_time)) {
464                         if (exitall_on_terminate)
465                                 terminate_threads(td->groupid, 0);
466                         td_verror(td, ENODATA, "check_min_rate");
467                         break;
468                 }
469
470                 if (td->thinktime) {
471                         unsigned long long b;
472
473                         b = td->io_blocks[0] + td->io_blocks[1];
474                         if (!(b % td->thinktime_blocks)) {
475                                 int left;
476
477                                 if (td->thinktime_spin)
478                                         __usec_sleep(td->thinktime_spin);
479
480                                 left = td->thinktime - td->thinktime_spin;
481                                 if (left)
482                                         usec_sleep(td, left);
483                         }
484                 }
485         }
486
487         if (!td->error) {
488                 struct fio_file *f;
489
490                 if (td->cur_depth)
491                         cleanup_pending_aio(td);
492
493                 if (should_fsync(td) && td->end_fsync) {
494                         td_set_runstate(td, TD_FSYNCING);
495                         for_each_file(td, f, i)
496                                 fio_io_sync(td, f);
497                 }
498         }
499 }
500
501 static void cleanup_io_u(struct thread_data *td)
502 {
503         struct list_head *entry, *n;
504         struct io_u *io_u;
505
506         list_for_each_safe(entry, n, &td->io_u_freelist) {
507                 io_u = list_entry(entry, struct io_u, list);
508
509                 list_del(&io_u->list);
510                 free(io_u);
511         }
512
513         free_io_mem(td);
514 }
515
516 /*
517  * "randomly" fill the buffer contents
518  */
519 static void fill_rand_buf(struct io_u *io_u, int max_bs)
520 {
521         int *ptr = io_u->buf;
522
523         while ((void *) ptr - io_u->buf < max_bs) {
524                 *ptr = rand() * 0x9e370001;
525                 ptr++;
526         }
527 }
528
529 static int init_io_u(struct thread_data *td)
530 {
531         struct io_u *io_u;
532         unsigned int max_bs;
533         int i, max_units;
534         char *p;
535
536         if (td->io_ops->flags & FIO_CPUIO)
537                 return 0;
538
539         if (td->io_ops->flags & FIO_SYNCIO)
540                 max_units = 1;
541         else
542                 max_units = td->iodepth;
543
544         max_bs = max(td->max_bs[DDIR_READ], td->max_bs[DDIR_WRITE]);
545         td->orig_buffer_size = max_bs * max_units;
546
547         if (td->mem_type == MEM_SHMHUGE || td->mem_type == MEM_MMAPHUGE)
548                 td->orig_buffer_size = (td->orig_buffer_size + td->hugepage_size - 1) & ~(td->hugepage_size - 1);
549         else
550                 td->orig_buffer_size += page_mask;
551
552         if (allocate_io_mem(td))
553                 return 1;
554
555         p = ALIGN(td->orig_buffer);
556         for (i = 0; i < max_units; i++) {
557                 io_u = malloc(sizeof(*io_u));
558                 memset(io_u, 0, sizeof(*io_u));
559                 INIT_LIST_HEAD(&io_u->list);
560
561                 io_u->buf = p + max_bs * i;
562                 if (td_write(td) || td_rw(td))
563                         fill_rand_buf(io_u, max_bs);
564
565                 io_u->index = i;
566                 io_u->flags = IO_U_F_FREE;
567                 list_add(&io_u->list, &td->io_u_freelist);
568         }
569
570         io_u_init_timeout();
571
572         return 0;
573 }
574
575 static int switch_ioscheduler(struct thread_data *td)
576 {
577         char tmp[256], tmp2[128];
578         FILE *f;
579         int ret;
580
581         if (td->io_ops->flags & FIO_CPUIO)
582                 return 0;
583
584         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
585
586         f = fopen(tmp, "r+");
587         if (!f) {
588                 td_verror(td, errno, "fopen");
589                 return 1;
590         }
591
592         /*
593          * Set io scheduler.
594          */
595         ret = fwrite(td->ioscheduler, strlen(td->ioscheduler), 1, f);
596         if (ferror(f) || ret != 1) {
597                 td_verror(td, errno, "fwrite");
598                 fclose(f);
599                 return 1;
600         }
601
602         rewind(f);
603
604         /*
605          * Read back and check that the selected scheduler is now the default.
606          */
607         ret = fread(tmp, 1, sizeof(tmp), f);
608         if (ferror(f) || ret < 0) {
609                 td_verror(td, errno, "fread");
610                 fclose(f);
611                 return 1;
612         }
613
614         sprintf(tmp2, "[%s]", td->ioscheduler);
615         if (!strstr(tmp, tmp2)) {
616                 log_err("fio: io scheduler %s not found\n", td->ioscheduler);
617                 td_verror(td, EINVAL, "iosched_switch");
618                 fclose(f);
619                 return 1;
620         }
621
622         fclose(f);
623         return 0;
624 }
625
626 static void clear_io_state(struct thread_data *td)
627 {
628         struct fio_file *f;
629         int i;
630
631         td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
632         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
633         td->zone_bytes = 0;
634
635         td->last_was_sync = 0;
636
637         for_each_file(td, f, i) {
638                 f->last_completed_pos = 0;
639
640                 f->last_pos = 0;
641                 if (td->io_ops->flags & FIO_SYNCIO)
642                         lseek(f->fd, SEEK_SET, 0);
643
644                 if (f->file_map)
645                         memset(f->file_map, 0, f->num_maps * sizeof(long));
646         }
647 }
648
649 /*
650  * Entry point for the thread based jobs. The process based jobs end up
651  * here as well, after a little setup.
652  */
653 static void *thread_main(void *data)
654 {
655         unsigned long long runtime[2];
656         struct thread_data *td = data;
657
658         if (!td->use_thread)
659                 setsid();
660
661         td->pid = getpid();
662
663         INIT_LIST_HEAD(&td->io_u_freelist);
664         INIT_LIST_HEAD(&td->io_u_busylist);
665         INIT_LIST_HEAD(&td->io_u_requeues);
666         INIT_LIST_HEAD(&td->io_hist_list);
667         INIT_LIST_HEAD(&td->io_log_list);
668
669         if (init_io_u(td))
670                 goto err;
671
672         if (fio_setaffinity(td) == -1) {
673                 td_verror(td, errno, "cpu_set_affinity");
674                 goto err;
675         }
676
677         if (init_iolog(td))
678                 goto err;
679
680         if (td->ioprio) {
681                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
682                         td_verror(td, errno, "ioprio_set");
683                         goto err;
684                 }
685         }
686
687         if (nice(td->nice) == -1) {
688                 td_verror(td, errno, "nice");
689                 goto err;
690         }
691
692         if (init_random_state(td))
693                 goto err;
694
695         if (td->ioscheduler && switch_ioscheduler(td))
696                 goto err;
697
698         td_set_runstate(td, TD_INITIALIZED);
699         fio_sem_up(&startup_sem);
700         fio_sem_down(&td->mutex);
701
702         if (!td->create_serialize && setup_files(td))
703                 goto err;
704         if (open_files(td))
705                 goto err;
706
707         /*
708          * Do this late, as some IO engines would like to have the
709          * files setup prior to initializing structures.
710          */
711         if (td_io_init(td))
712                 goto err;
713
714         if (td->exec_prerun) {
715                 if (system(td->exec_prerun) < 0)
716                         goto err;
717         }
718
719         fio_gettime(&td->epoch, NULL);
720         memcpy(&td->timeout_end, &td->epoch, sizeof(td->epoch));
721         getrusage(RUSAGE_SELF, &td->ts.ru_start);
722
723         runtime[0] = runtime[1] = 0;
724         while (td->loops--) {
725                 fio_gettime(&td->start, NULL);
726                 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
727
728                 if (td->ratemin)
729                         memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
730
731                 clear_io_state(td);
732                 prune_io_piece_log(td);
733
734                 if (td->io_ops->flags & FIO_CPUIO)
735                         do_cpuio(td);
736                 else
737                         do_io(td);
738
739                 runtime[td->ddir] += utime_since_now(&td->start);
740                 if (td_rw(td) && td->io_bytes[td->ddir ^ 1])
741                         runtime[td->ddir ^ 1] = runtime[td->ddir];
742
743                 if (td->error || td->terminate)
744                         break;
745
746                 if (td->verify == VERIFY_NONE)
747                         continue;
748
749                 clear_io_state(td);
750                 fio_gettime(&td->start, NULL);
751
752                 do_verify(td);
753
754                 runtime[DDIR_READ] += utime_since_now(&td->start);
755
756                 if (td->error || td->terminate)
757                         break;
758         }
759
760         update_rusage_stat(td);
761         fio_gettime(&td->end_time, NULL);
762         td->runtime[0] = runtime[0] / 1000;
763         td->runtime[1] = runtime[1] / 1000;
764
765         if (td->ts.bw_log)
766                 finish_log(td, td->ts.bw_log, "bw");
767         if (td->ts.slat_log)
768                 finish_log(td, td->ts.slat_log, "slat");
769         if (td->ts.clat_log)
770                 finish_log(td, td->ts.clat_log, "clat");
771         if (td->write_iolog_file)
772                 write_iolog_close(td);
773         if (td->exec_postrun) {
774                 if (system(td->exec_postrun) < 0)
775                         log_err("fio: postrun %s failed\n", td->exec_postrun);
776         }
777
778         if (exitall_on_terminate)
779                 terminate_threads(td->groupid, 0);
780
781 err:
782         if (td->error)
783                 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
784         close_files(td);
785         close_ioengine(td);
786         cleanup_io_u(td);
787         td_set_runstate(td, TD_EXITED);
788         return (void *) (unsigned long) td->error;
789 }
790
791 /*
792  * We cannot pass the td data into a forked process, so attach the td and
793  * pass it to the thread worker.
794  */
795 static int fork_main(int shmid, int offset)
796 {
797         struct thread_data *td;
798         void *data, *ret;
799
800         data = shmat(shmid, NULL, 0);
801         if (data == (void *) -1) {
802                 int __err = errno;
803
804                 perror("shmat");
805                 return __err;
806         }
807
808         td = data + offset * sizeof(struct thread_data);
809         ret = thread_main(td);
810         shmdt(data);
811         return (int) (unsigned long) ret;
812 }
813
814 /*
815  * Run over the job map and reap the threads that have exited, if any.
816  */
817 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
818 {
819         struct thread_data *td;
820         int i, cputhreads, pending, status, ret;
821
822         /*
823          * reap exited threads (TD_EXITED -> TD_REAPED)
824          */
825         pending = cputhreads = 0;
826         for_each_td(td, i) {
827                 int flags = 0;
828
829                 /*
830                  * ->io_ops is NULL for a thread that has closed its
831                  * io engine
832                  */
833                 if (td->io_ops && td->io_ops->flags & FIO_CPUIO)
834                         cputhreads++;
835
836                 if (!td->pid || td->runstate == TD_REAPED)
837                         continue;
838                 if (td->use_thread) {
839                         if (td->runstate == TD_EXITED) {
840                                 td_set_runstate(td, TD_REAPED);
841                                 goto reaped;
842                         }
843                         continue;
844                 }
845
846                 flags = WNOHANG;
847                 if (td->runstate == TD_EXITED)
848                         flags = 0;
849
850                 /*
851                  * check if someone quit or got killed in an unusual way
852                  */
853                 ret = waitpid(td->pid, &status, flags);
854                 if (ret < 0) {
855                         if (errno == ECHILD) {
856                                 log_err("fio: pid=%d disappeared %d\n", td->pid, td->runstate);
857                                 td_set_runstate(td, TD_REAPED);
858                                 goto reaped;
859                         }
860                         perror("waitpid");
861                 } else if (ret == td->pid) {
862                         if (WIFSIGNALED(status)) {
863                                 int sig = WTERMSIG(status);
864
865                                 log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
866                                 td_set_runstate(td, TD_REAPED);
867                                 goto reaped;
868                         }
869                         if (WIFEXITED(status)) {
870                                 if (WEXITSTATUS(status) && !td->error)
871                                         td->error = WEXITSTATUS(status);
872
873                                 td_set_runstate(td, TD_REAPED);
874                                 goto reaped;
875                         }
876                 }
877
878                 /*
879                  * thread is not dead, continue
880                  */
881                 continue;
882 reaped:
883                 if (td->use_thread) {
884                         long ret;
885
886                         if (pthread_join(td->thread, (void *) &ret))
887                                 perror("pthread_join");
888                 }
889
890                 (*nr_running)--;
891                 (*m_rate) -= td->ratemin;
892                 (*t_rate) -= td->rate;
893
894                 if (td->error)
895                         exit_value++;
896         }
897
898         if (*nr_running == cputhreads && !pending)
899                 terminate_threads(TERMINATE_ALL, 0);
900 }
901
902 /*
903  * Main function for kicking off and reaping jobs, as needed.
904  */
905 static void run_threads(void)
906 {
907         struct thread_data *td;
908         unsigned long spent;
909         int i, todo, nr_running, m_rate, t_rate, nr_started;
910
911         if (fio_pin_memory())
912                 return;
913
914         if (!terse_output) {
915                 printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
916                 fflush(stdout);
917         }
918
919         signal(SIGINT, sig_handler);
920         signal(SIGALRM, sig_handler);
921
922         todo = thread_number;
923         nr_running = 0;
924         nr_started = 0;
925         m_rate = t_rate = 0;
926
927         for_each_td(td, i) {
928                 print_status_init(td->thread_number - 1);
929
930                 if (!td->create_serialize) {
931                         init_disk_util(td);
932                         continue;
933                 }
934
935                 /*
936                  * do file setup here so it happens sequentially,
937                  * we don't want X number of threads getting their
938                  * client data interspersed on disk
939                  */
940                 if (setup_files(td)) {
941                         exit_value++;
942                         if (td->error)
943                                 log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
944                         td_set_runstate(td, TD_REAPED);
945                         todo--;
946                 }
947
948                 init_disk_util(td);
949         }
950
951         set_genesis_time();
952
953         while (todo) {
954                 struct thread_data *map[MAX_JOBS];
955                 struct timeval this_start;
956                 int this_jobs = 0, left;
957
958                 /*
959                  * create threads (TD_NOT_CREATED -> TD_CREATED)
960                  */
961                 for_each_td(td, i) {
962                         if (td->runstate != TD_NOT_CREATED)
963                                 continue;
964
965                         /*
966                          * never got a chance to start, killed by other
967                          * thread for some reason
968                          */
969                         if (td->terminate) {
970                                 todo--;
971                                 continue;
972                         }
973
974                         if (td->start_delay) {
975                                 spent = mtime_since_genesis();
976
977                                 if (td->start_delay * 1000 > spent)
978                                         continue;
979                         }
980
981                         if (td->stonewall && (nr_started || nr_running))
982                                 break;
983
984                         /*
985                          * Set state to created. Thread will transition
986                          * to TD_INITIALIZED when it's done setting up.
987                          */
988                         td_set_runstate(td, TD_CREATED);
989                         map[this_jobs++] = td;
990                         fio_sem_init(&startup_sem, 1);
991                         nr_started++;
992
993                         if (td->use_thread) {
994                                 if (pthread_create(&td->thread, NULL, thread_main, td)) {
995                                         perror("thread_create");
996                                         nr_started--;
997                                 }
998                         } else {
999                                 if (fork())
1000                                         fio_sem_down(&startup_sem);
1001                                 else {
1002                                         int ret = fork_main(shm_id, i);
1003
1004                                         exit(ret);
1005                                 }
1006                         }
1007                 }
1008
1009                 /*
1010                  * Wait for the started threads to transition to
1011                  * TD_INITIALIZED.
1012                  */
1013                 fio_gettime(&this_start, NULL);
1014                 left = this_jobs;
1015                 while (left && !fio_abort) {
1016                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1017                                 break;
1018
1019                         usleep(100000);
1020
1021                         for (i = 0; i < this_jobs; i++) {
1022                                 td = map[i];
1023                                 if (!td)
1024                                         continue;
1025                                 if (td->runstate == TD_INITIALIZED) {
1026                                         map[i] = NULL;
1027                                         left--;
1028                                 } else if (td->runstate >= TD_EXITED) {
1029                                         map[i] = NULL;
1030                                         left--;
1031                                         todo--;
1032                                         nr_running++; /* work-around... */
1033                                 }
1034                         }
1035                 }
1036
1037                 if (left) {
1038                         log_err("fio: %d jobs failed to start\n", left);
1039                         for (i = 0; i < this_jobs; i++) {
1040                                 td = map[i];
1041                                 if (!td)
1042                                         continue;
1043                                 kill(td->pid, SIGTERM);
1044                         }
1045                         break;
1046                 }
1047
1048                 /*
1049                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1050                  */
1051                 for_each_td(td, i) {
1052                         if (td->runstate != TD_INITIALIZED)
1053                                 continue;
1054
1055                         td_set_runstate(td, TD_RUNNING);
1056                         nr_running++;
1057                         nr_started--;
1058                         m_rate += td->ratemin;
1059                         t_rate += td->rate;
1060                         todo--;
1061                         fio_sem_up(&td->mutex);
1062                 }
1063
1064                 reap_threads(&nr_running, &t_rate, &m_rate);
1065
1066                 if (todo)
1067                         usleep(100000);
1068         }
1069
1070         while (nr_running) {
1071                 reap_threads(&nr_running, &t_rate, &m_rate);
1072                 usleep(10000);
1073         }
1074
1075         update_io_ticks();
1076         fio_unpin_memory();
1077 }
1078
1079 int main(int argc, char *argv[])
1080 {
1081         long ps;
1082
1083         /*
1084          * We need locale for number printing, if it isn't set then just
1085          * go with the US format.
1086          */
1087         if (!getenv("LC_NUMERIC"))
1088                 setlocale(LC_NUMERIC, "en_US");
1089
1090         if (parse_options(argc, argv))
1091                 return 1;
1092
1093         if (!thread_number) {
1094                 log_err("Nothing to do\n");
1095                 return 1;
1096         }
1097
1098         ps = sysconf(_SC_PAGESIZE);
1099         if (ps < 0) {
1100                 log_err("Failed to get page size\n");
1101                 return 1;
1102         }
1103
1104         page_mask = ps - 1;
1105
1106         if (write_bw_log) {
1107                 setup_log(&agg_io_log[DDIR_READ]);
1108                 setup_log(&agg_io_log[DDIR_WRITE]);
1109         }
1110
1111         set_genesis_time();
1112
1113         disk_util_timer_arm();
1114
1115         run_threads();
1116
1117         if (!fio_abort) {
1118                 show_run_stats();
1119                 if (write_bw_log) {
1120                         __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1121                         __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");
1122                 }
1123         }
1124
1125         return exit_value;
1126 }