syslet: x86_64 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 "os.h"
39
40 static unsigned long page_mask;
41 #define ALIGN(buf)      \
42         (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
43
44 int groupid = 0;
45 int thread_number = 0;
46 int shm_id = 0;
47 int temp_stall_ts;
48
49 static volatile int startup_sem;
50 static volatile int fio_abort;
51 static int exit_value;
52
53 struct io_log *agg_io_log[2];
54
55 #define TERMINATE_ALL           (-1)
56 #define JOB_START_TIMEOUT       (5 * 1000)
57
58 static inline void td_set_runstate(struct thread_data *td, int runstate)
59 {
60         td->runstate = runstate;
61 }
62
63 static void terminate_threads(int group_id, int forced_kill)
64 {
65         struct thread_data *td;
66         int i;
67
68         for_each_td(td, i) {
69                 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
70                         td->terminate = 1;
71                         td->start_delay = 0;
72                         if (forced_kill)
73                                 td_set_runstate(td, TD_EXITED);
74                 }
75         }
76 }
77
78 static void sig_handler(int sig)
79 {
80         switch (sig) {
81                 case SIGALRM:
82                         update_io_ticks();
83                         disk_util_timer_arm();
84                         print_thread_status();
85                         break;
86                 default:
87                         printf("\nfio: terminating on signal %d\n", sig);
88                         fflush(stdout);
89                         terminate_threads(TERMINATE_ALL, 0);
90                         break;
91         }
92 }
93
94 /*
95  * Check if we are above the minimum rate given.
96  */
97 static int check_min_rate(struct thread_data *td, struct timeval *now)
98 {
99         unsigned long long bytes = 0;
100         unsigned long spent;
101         unsigned long rate;
102
103         /*
104          * allow a 2 second settle period in the beginning
105          */
106         if (mtime_since(&td->start, now) < 2000)
107                 return 0;
108
109         if (td_read(td))
110                 bytes += td->this_io_bytes[DDIR_READ];
111         if (td_write(td))
112                 bytes += td->this_io_bytes[DDIR_WRITE];
113
114         /*
115          * if rate blocks is set, sample is running
116          */
117         if (td->rate_bytes) {
118                 spent = mtime_since(&td->lastrate, now);
119                 if (spent < td->ratecycle)
120                         return 0;
121
122                 if (bytes < td->rate_bytes) {
123                         fprintf(f_out, "%s: min rate %u not met\n", td->name, td->ratemin);
124                         return 1;
125                 } else {
126                         rate = (bytes - td->rate_bytes) / spent;
127                         if (rate < td->ratemin || bytes < td->rate_bytes) {
128                                 fprintf(f_out, "%s: min rate %u not met, got %luKiB/sec\n", td->name, td->ratemin, rate);
129                                 return 1;
130                         }
131                 }
132         }
133
134         td->rate_bytes = bytes;
135         memcpy(&td->lastrate, now, sizeof(*now));
136         return 0;
137 }
138
139 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
140 {
141         if (!td->timeout)
142                 return 0;
143         if (mtime_since(&td->epoch, t) >= td->timeout * 1000)
144                 return 1;
145
146         return 0;
147 }
148
149 /*
150  * When job exits, we can cancel the in-flight IO if we are using async
151  * io. Attempt to do so.
152  */
153 static void cleanup_pending_aio(struct thread_data *td)
154 {
155         struct list_head *entry, *n;
156         struct io_u *io_u;
157         int r;
158
159         /*
160          * get immediately available events, if any
161          */
162         r = io_u_queued_complete(td, 0);
163         if (r < 0)
164                 return;
165
166         /*
167          * now cancel remaining active events
168          */
169         if (td->io_ops->cancel) {
170                 list_for_each_safe(entry, n, &td->io_u_busylist) {
171                         io_u = list_entry(entry, struct io_u, list);
172
173                         /*
174                          * if the io_u isn't in flight, then that generally
175                          * means someone leaked an io_u. complain but fix
176                          * it up, so we don't stall here.
177                          */
178                         if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
179                                 log_err("fio: non-busy IO on busy list\n");
180                                 put_io_u(td, io_u);
181                         } else {
182                                 r = td->io_ops->cancel(td, io_u);
183                                 if (!r)
184                                         put_io_u(td, io_u);
185                         }
186                 }
187         }
188
189         if (td->cur_depth)
190                 r = io_u_queued_complete(td, td->cur_depth);
191 }
192
193 /*
194  * Helper to handle the final sync of a file. Works just like the normal
195  * io path, just does everything sync.
196  */
197 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
198 {
199         struct io_u *io_u = __get_io_u(td);
200         int ret;
201
202         if (!io_u)
203                 return 1;
204
205         io_u->ddir = DDIR_SYNC;
206         io_u->file = f;
207
208         if (td_io_prep(td, io_u)) {
209                 put_io_u(td, io_u);
210                 return 1;
211         }
212
213 requeue:
214         ret = td_io_queue(td, io_u);
215         if (ret < 0) {
216                 td_verror(td, io_u->error, "td_io_queue");
217                 put_io_u(td, io_u);
218                 return 1;
219         } else if (ret == FIO_Q_QUEUED) {
220                 if (io_u_queued_complete(td, 1) < 0)
221                         return 1;
222         } else if (ret == FIO_Q_COMPLETED) {
223                 if (io_u->error) {
224                         td_verror(td, io_u->error, "td_io_queue");
225                         return 1;
226                 }
227
228                 if (io_u_sync_complete(td, io_u) < 0)
229                         return 1;
230         } else if (ret == FIO_Q_BUSY) {
231                 if (td_io_commit(td))
232                         return 1;
233                 goto requeue;
234         }
235
236         return 0;
237 }
238
239 /*
240  * The main verify engine. Runs over the writes we previusly submitted,
241  * reads the blocks back in, and checks the crc/md5 of the data.
242  */
243 static void do_verify(struct thread_data *td)
244 {
245         struct fio_file *f;
246         struct io_u *io_u;
247         int ret, i, min_events;
248
249         /*
250          * sync io first and invalidate cache, to make sure we really
251          * read from disk.
252          */
253         for_each_file(td, f, i) {
254                 if (fio_io_sync(td, f))
255                         break;
256                 if (file_invalidate_cache(td, f))
257                         break;
258         }
259
260         if (td->error)
261                 return;
262
263         td_set_runstate(td, TD_VERIFYING);
264
265         io_u = NULL;
266         while (!td->terminate) {
267                 io_u = __get_io_u(td);
268                 if (!io_u)
269                         break;
270
271                 if (runtime_exceeded(td, &io_u->start_time)) {
272                         put_io_u(td, io_u);
273                         break;
274                 }
275
276                 if (get_next_verify(td, io_u)) {
277                         put_io_u(td, io_u);
278                         break;
279                 }
280
281                 if (td_io_prep(td, io_u)) {
282                         put_io_u(td, io_u);
283                         break;
284                 }
285
286                 io_u->end_io = verify_io_u;
287 requeue:
288                 ret = td_io_queue(td, io_u);
289
290                 switch (ret) {
291                 case FIO_Q_COMPLETED:
292                         if (io_u->error)
293                                 ret = -io_u->error;
294                         if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
295                                 int bytes = io_u->xfer_buflen - io_u->resid;
296
297                                 io_u->xfer_buflen = io_u->resid;
298                                 io_u->xfer_buf += bytes;
299                                 goto requeue;
300                         }
301                         ret = io_u_sync_complete(td, io_u);
302                         if (ret < 0)
303                                 break;
304                         continue;
305                 case FIO_Q_QUEUED:
306                         break;
307                 case FIO_Q_BUSY:
308                         requeue_io_u(td, &io_u);
309                         ret = td_io_commit(td);
310                         break;
311                 default:
312                         assert(ret < 0);
313                         td_verror(td, -ret, "td_io_queue");
314                         break;
315                 }
316
317                 if (ret < 0 || td->error)
318                         break;
319
320                 /*
321                  * if we can queue more, do so. but check if there are
322                  * completed io_u's first.
323                  */
324                 min_events = 0;
325                 if (queue_full(td) || ret == FIO_Q_BUSY) {
326                         min_events = 1;
327
328                         if (td->cur_depth > td->iodepth_low)
329                                 min_events = td->cur_depth - td->iodepth_low;
330                 }
331
332                 /*
333                  * Reap required number of io units, if any, and do the
334                  * verification on them through the callback handler
335                  */
336                 if (io_u_queued_complete(td, min_events) < 0)
337                         break;
338         }
339
340         if (!td->error) {
341                 min_events = td->cur_depth;
342
343                 if (min_events)
344                         ret = io_u_queued_complete(td, min_events);
345         } else
346                 cleanup_pending_aio(td);
347
348         td_set_runstate(td, TD_RUNNING);
349 }
350
351 /*
352  * Not really an io thread, all it does is burn CPU cycles in the specified
353  * manner.
354  */
355 static void do_cpuio(struct thread_data *td)
356 {
357         struct timeval e;
358         int split = 100 / td->cpuload;
359         int i = 0;
360
361         while (!td->terminate) {
362                 fio_gettime(&e, NULL);
363
364                 if (runtime_exceeded(td, &e))
365                         break;
366
367                 if (!(i % split))
368                         __usec_sleep(10000);
369                 else
370                         usec_sleep(td, 10000);
371
372                 i++;
373         }
374 }
375
376 /*
377  * Main IO worker function. It retrieves io_u's to process and queues
378  * and reaps them, checking for rate and errors along the way.
379  */
380 static void do_io(struct thread_data *td)
381 {
382         struct timeval s;
383         unsigned long usec;
384         int i, ret = 0;
385
386         td_set_runstate(td, TD_RUNNING);
387
388         while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->io_size) {
389                 struct timeval comp_time;
390                 long bytes_done = 0;
391                 int min_evts = 0;
392                 struct io_u *io_u;
393
394                 if (td->terminate)
395                         break;
396
397                 io_u = get_io_u(td);
398                 if (!io_u)
399                         break;
400
401                 memcpy(&s, &io_u->start_time, sizeof(s));
402
403                 if (runtime_exceeded(td, &s)) {
404                         put_io_u(td, io_u);
405                         break;
406                 }
407 requeue:
408                 ret = td_io_queue(td, io_u);
409
410                 switch (ret) {
411                 case FIO_Q_COMPLETED:
412                         if (io_u->error) {
413                                 ret = io_u->error;
414                                 break;
415                         }
416                         if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
417                                 int bytes = io_u->xfer_buflen - io_u->resid;
418
419                                 io_u->xfer_buflen = io_u->resid;
420                                 io_u->xfer_buf += bytes;
421                                 goto requeue;
422                         }
423                         fio_gettime(&comp_time, NULL);
424                         bytes_done = io_u_sync_complete(td, io_u);
425                         if (bytes_done < 0)
426                                 ret = bytes_done;
427                         break;
428                 case FIO_Q_QUEUED:
429                         /*
430                          * if the engine doesn't have a commit hook,
431                          * the io_u is really queued. if it does have such
432                          * a hook, it has to call io_u_queued() itself.
433                          */
434                         if (td->io_ops->commit == NULL)
435                                 io_u_queued(td, io_u);
436                         break;
437                 case FIO_Q_BUSY:
438                         requeue_io_u(td, &io_u);
439                         ret = td_io_commit(td);
440                         break;
441                 default:
442                         assert(ret < 0);
443                         put_io_u(td, io_u);
444                         break;
445                 }
446
447                 if (ret < 0 || td->error)
448                         break;
449
450                 /*
451                  * See if we need to complete some commands
452                  */
453                 if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
454                         min_evts = 0;
455                         if (queue_full(td) || ret == FIO_Q_BUSY) {
456                                 min_evts = 1;
457
458                                 if (td->cur_depth > td->iodepth_low)
459                                         min_evts = td->cur_depth - td->iodepth_low;
460                         }
461
462                         fio_gettime(&comp_time, NULL);
463                         bytes_done = io_u_queued_complete(td, min_evts);
464                         if (bytes_done < 0)
465                                 break;
466                 }
467
468                 if (!bytes_done)
469                         continue;
470
471                 /*
472                  * the rate is batched for now, it should work for batches
473                  * of completions except the very first one which may look
474                  * a little bursty
475                  */
476                 usec = utime_since(&s, &comp_time);
477
478                 rate_throttle(td, usec, bytes_done);
479
480                 if (check_min_rate(td, &comp_time)) {
481                         if (exitall_on_terminate)
482                                 terminate_threads(td->groupid, 0);
483                         td_verror(td, ENODATA, "check_min_rate");
484                         break;
485                 }
486
487                 if (td->thinktime) {
488                         unsigned long long b;
489
490                         b = td->io_blocks[0] + td->io_blocks[1];
491                         if (!(b % td->thinktime_blocks)) {
492                                 int left;
493
494                                 if (td->thinktime_spin)
495                                         __usec_sleep(td->thinktime_spin);
496
497                                 left = td->thinktime - td->thinktime_spin;
498                                 if (left)
499                                         usec_sleep(td, left);
500                         }
501                 }
502         }
503
504         if (!td->error) {
505                 struct fio_file *f;
506
507                 i = td->cur_depth;
508                 if (i)
509                         ret = io_u_queued_complete(td, i);
510
511                 if (should_fsync(td) && td->end_fsync) {
512                         td_set_runstate(td, TD_FSYNCING);
513                         for_each_file(td, f, i)
514                                 fio_io_sync(td, f);
515                 }
516         } else
517                 cleanup_pending_aio(td);
518 }
519
520 static void cleanup_io_u(struct thread_data *td)
521 {
522         struct list_head *entry, *n;
523         struct io_u *io_u;
524
525         list_for_each_safe(entry, n, &td->io_u_freelist) {
526                 io_u = list_entry(entry, struct io_u, list);
527
528                 list_del(&io_u->list);
529                 free(io_u);
530         }
531
532         free_io_mem(td);
533 }
534
535 /*
536  * "randomly" fill the buffer contents
537  */
538 static void fill_rand_buf(struct io_u *io_u, int max_bs)
539 {
540         int *ptr = io_u->buf;
541
542         while ((void *) ptr - io_u->buf < max_bs) {
543                 *ptr = rand() * 0x9e370001;
544                 ptr++;
545         }
546 }
547
548 static int init_io_u(struct thread_data *td)
549 {
550         struct io_u *io_u;
551         unsigned int max_bs;
552         int i, max_units;
553         char *p;
554
555         if (td->io_ops->flags & FIO_CPUIO)
556                 return 0;
557
558         if (td->io_ops->flags & FIO_SYNCIO)
559                 max_units = 1;
560         else
561                 max_units = td->iodepth;
562
563         max_bs = max(td->max_bs[DDIR_READ], td->max_bs[DDIR_WRITE]);
564         td->orig_buffer_size = max_bs * max_units;
565
566         if (td->mem_type == MEM_SHMHUGE || td->mem_type == MEM_MMAPHUGE)
567                 td->orig_buffer_size = (td->orig_buffer_size + td->hugepage_size - 1) & ~(td->hugepage_size - 1);
568         else
569                 td->orig_buffer_size += page_mask;
570
571         if (allocate_io_mem(td))
572                 return 1;
573
574         p = ALIGN(td->orig_buffer);
575         for (i = 0; i < max_units; i++) {
576                 io_u = malloc(sizeof(*io_u));
577                 memset(io_u, 0, sizeof(*io_u));
578                 INIT_LIST_HEAD(&io_u->list);
579
580                 io_u->buf = p + max_bs * i;
581                 if (td_write(td) || td_rw(td))
582                         fill_rand_buf(io_u, max_bs);
583
584                 io_u->index = i;
585                 io_u->flags = IO_U_F_FREE;
586                 list_add(&io_u->list, &td->io_u_freelist);
587         }
588
589         io_u_init_timeout();
590
591         return 0;
592 }
593
594 static int switch_ioscheduler(struct thread_data *td)
595 {
596         char tmp[256], tmp2[128];
597         FILE *f;
598         int ret;
599
600         if (td->io_ops->flags & FIO_CPUIO)
601                 return 0;
602
603         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
604
605         f = fopen(tmp, "r+");
606         if (!f) {
607                 td_verror(td, errno, "fopen");
608                 return 1;
609         }
610
611         /*
612          * Set io scheduler.
613          */
614         ret = fwrite(td->ioscheduler, strlen(td->ioscheduler), 1, f);
615         if (ferror(f) || ret != 1) {
616                 td_verror(td, errno, "fwrite");
617                 fclose(f);
618                 return 1;
619         }
620
621         rewind(f);
622
623         /*
624          * Read back and check that the selected scheduler is now the default.
625          */
626         ret = fread(tmp, 1, sizeof(tmp), f);
627         if (ferror(f) || ret < 0) {
628                 td_verror(td, errno, "fread");
629                 fclose(f);
630                 return 1;
631         }
632
633         sprintf(tmp2, "[%s]", td->ioscheduler);
634         if (!strstr(tmp, tmp2)) {
635                 log_err("fio: io scheduler %s not found\n", td->ioscheduler);
636                 td_verror(td, EINVAL, "iosched_switch");
637                 fclose(f);
638                 return 1;
639         }
640
641         fclose(f);
642         return 0;
643 }
644
645 static void clear_io_state(struct thread_data *td)
646 {
647         struct fio_file *f;
648         int i;
649
650         td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
651         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
652         td->zone_bytes = 0;
653
654         td->last_was_sync = 0;
655
656         for_each_file(td, f, i) {
657                 f->last_completed_pos = 0;
658
659                 f->last_pos = 0;
660                 if (td->io_ops->flags & FIO_SYNCIO)
661                         lseek(f->fd, SEEK_SET, 0);
662
663                 if (f->file_map)
664                         memset(f->file_map, 0, f->num_maps * sizeof(long));
665         }
666 }
667
668 /*
669  * Entry point for the thread based jobs. The process based jobs end up
670  * here as well, after a little setup.
671  */
672 static void *thread_main(void *data)
673 {
674         unsigned long long runtime[2];
675         struct thread_data *td = data;
676
677         if (!td->use_thread)
678                 setsid();
679
680         td->pid = getpid();
681
682         INIT_LIST_HEAD(&td->io_u_freelist);
683         INIT_LIST_HEAD(&td->io_u_busylist);
684         INIT_LIST_HEAD(&td->io_u_requeues);
685         INIT_LIST_HEAD(&td->io_hist_list);
686         INIT_LIST_HEAD(&td->io_log_list);
687
688         if (init_io_u(td))
689                 goto err;
690
691         if (fio_setaffinity(td) == -1) {
692                 td_verror(td, errno, "cpu_set_affinity");
693                 goto err;
694         }
695
696         if (init_iolog(td))
697                 goto err;
698
699         if (td->ioprio) {
700                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
701                         td_verror(td, errno, "ioprio_set");
702                         goto err;
703                 }
704         }
705
706         if (nice(td->nice) == -1) {
707                 td_verror(td, errno, "nice");
708                 goto err;
709         }
710
711         if (init_random_state(td))
712                 goto err;
713
714         if (td->ioscheduler && switch_ioscheduler(td))
715                 goto err;
716
717         td_set_runstate(td, TD_INITIALIZED);
718         fio_sem_up(&startup_sem);
719         fio_sem_down(&td->mutex);
720
721         if (!td->create_serialize && setup_files(td))
722                 goto err;
723         if (open_files(td))
724                 goto err;
725
726         /*
727          * Do this late, as some IO engines would like to have the
728          * files setup prior to initializing structures.
729          */
730         if (td_io_init(td))
731                 goto err;
732
733         if (td->exec_prerun) {
734                 if (system(td->exec_prerun) < 0)
735                         goto err;
736         }
737
738         fio_gettime(&td->epoch, NULL);
739         memcpy(&td->timeout_end, &td->epoch, sizeof(td->epoch));
740         getrusage(RUSAGE_SELF, &td->ts.ru_start);
741
742         runtime[0] = runtime[1] = 0;
743         while (td->loops--) {
744                 fio_gettime(&td->start, NULL);
745                 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
746
747                 if (td->ratemin)
748                         memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
749
750                 clear_io_state(td);
751                 prune_io_piece_log(td);
752
753                 if (td->io_ops->flags & FIO_CPUIO)
754                         do_cpuio(td);
755                 else
756                         do_io(td);
757
758                 if (td_read(td) && td->io_bytes[DDIR_READ])
759                         runtime[DDIR_READ] += utime_since_now(&td->start);
760                 if (td_write(td) && td->io_bytes[DDIR_WRITE])
761                         runtime[DDIR_WRITE] += utime_since_now(&td->start);
762                 
763                 if (td->error || td->terminate)
764                         break;
765
766                 if (td->verify == VERIFY_NONE)
767                         continue;
768
769                 clear_io_state(td);
770                 fio_gettime(&td->start, NULL);
771
772                 do_verify(td);
773
774                 runtime[DDIR_READ] += utime_since_now(&td->start);
775
776                 if (td->error || td->terminate)
777                         break;
778         }
779
780         update_rusage_stat(td);
781         fio_gettime(&td->end_time, NULL);
782         td->runtime[0] = runtime[0] / 1000;
783         td->runtime[1] = runtime[1] / 1000;
784
785         if (td->ts.bw_log)
786                 finish_log(td, td->ts.bw_log, "bw");
787         if (td->ts.slat_log)
788                 finish_log(td, td->ts.slat_log, "slat");
789         if (td->ts.clat_log)
790                 finish_log(td, td->ts.clat_log, "clat");
791         if (td->write_iolog_file)
792                 write_iolog_close(td);
793         if (td->exec_postrun) {
794                 if (system(td->exec_postrun) < 0)
795                         log_err("fio: postrun %s failed\n", td->exec_postrun);
796         }
797
798         if (exitall_on_terminate)
799                 terminate_threads(td->groupid, 0);
800
801 err:
802         if (td->error)
803                 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
804         close_files(td);
805         close_ioengine(td);
806         cleanup_io_u(td);
807         td_set_runstate(td, TD_EXITED);
808         return (void *) (unsigned long) td->error;
809 }
810
811 /*
812  * We cannot pass the td data into a forked process, so attach the td and
813  * pass it to the thread worker.
814  */
815 static int fork_main(int shmid, int offset)
816 {
817         struct thread_data *td;
818         void *data, *ret;
819
820         data = shmat(shmid, NULL, 0);
821         if (data == (void *) -1) {
822                 int __err = errno;
823
824                 perror("shmat");
825                 return __err;
826         }
827
828         td = data + offset * sizeof(struct thread_data);
829         ret = thread_main(td);
830         shmdt(data);
831         return (int) (unsigned long) ret;
832 }
833
834 /*
835  * Run over the job map and reap the threads that have exited, if any.
836  */
837 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
838 {
839         struct thread_data *td;
840         int i, cputhreads, pending, status, ret;
841
842         /*
843          * reap exited threads (TD_EXITED -> TD_REAPED)
844          */
845         pending = cputhreads = 0;
846         for_each_td(td, i) {
847                 int flags = 0;
848
849                 /*
850                  * ->io_ops is NULL for a thread that has closed its
851                  * io engine
852                  */
853                 if (td->io_ops && td->io_ops->flags & FIO_CPUIO)
854                         cputhreads++;
855
856                 if (!td->pid || td->runstate == TD_REAPED)
857                         continue;
858                 if (td->use_thread) {
859                         if (td->runstate == TD_EXITED) {
860                                 td_set_runstate(td, TD_REAPED);
861                                 goto reaped;
862                         }
863                         continue;
864                 }
865
866                 flags = WNOHANG;
867                 if (td->runstate == TD_EXITED)
868                         flags = 0;
869
870                 /*
871                  * check if someone quit or got killed in an unusual way
872                  */
873                 ret = waitpid(td->pid, &status, flags);
874                 if (ret < 0) {
875                         if (errno == ECHILD) {
876                                 log_err("fio: pid=%d disappeared %d\n", td->pid, td->runstate);
877                                 td_set_runstate(td, TD_REAPED);
878                                 goto reaped;
879                         }
880                         perror("waitpid");
881                 } else if (ret == td->pid) {
882                         if (WIFSIGNALED(status)) {
883                                 int sig = WTERMSIG(status);
884
885                                 log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
886                                 td_set_runstate(td, TD_REAPED);
887                                 goto reaped;
888                         }
889                         if (WIFEXITED(status)) {
890                                 if (WEXITSTATUS(status) && !td->error)
891                                         td->error = WEXITSTATUS(status);
892
893                                 td_set_runstate(td, TD_REAPED);
894                                 goto reaped;
895                         }
896                 }
897
898                 /*
899                  * thread is not dead, continue
900                  */
901                 continue;
902 reaped:
903                 if (td->use_thread) {
904                         long ret;
905
906                         if (pthread_join(td->thread, (void *) &ret))
907                                 perror("pthread_join");
908                 }
909
910                 (*nr_running)--;
911                 (*m_rate) -= td->ratemin;
912                 (*t_rate) -= td->rate;
913
914                 if (td->error)
915                         exit_value++;
916         }
917
918         if (*nr_running == cputhreads && !pending)
919                 terminate_threads(TERMINATE_ALL, 0);
920 }
921
922 /*
923  * Main function for kicking off and reaping jobs, as needed.
924  */
925 static void run_threads(void)
926 {
927         struct thread_data *td;
928         unsigned long spent;
929         int i, todo, nr_running, m_rate, t_rate, nr_started;
930
931         if (fio_pin_memory())
932                 return;
933
934         if (!terse_output) {
935                 printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
936                 fflush(stdout);
937         }
938
939         signal(SIGINT, sig_handler);
940         signal(SIGALRM, sig_handler);
941
942         todo = thread_number;
943         nr_running = 0;
944         nr_started = 0;
945         m_rate = t_rate = 0;
946
947         for_each_td(td, i) {
948                 print_status_init(td->thread_number - 1);
949
950                 if (!td->create_serialize) {
951                         init_disk_util(td);
952                         continue;
953                 }
954
955                 /*
956                  * do file setup here so it happens sequentially,
957                  * we don't want X number of threads getting their
958                  * client data interspersed on disk
959                  */
960                 if (setup_files(td)) {
961                         exit_value++;
962                         if (td->error)
963                                 log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
964                         td_set_runstate(td, TD_REAPED);
965                         todo--;
966                 }
967
968                 init_disk_util(td);
969         }
970
971         set_genesis_time();
972
973         while (todo) {
974                 struct thread_data *map[MAX_JOBS];
975                 struct timeval this_start;
976                 int this_jobs = 0, left;
977
978                 /*
979                  * create threads (TD_NOT_CREATED -> TD_CREATED)
980                  */
981                 for_each_td(td, i) {
982                         if (td->runstate != TD_NOT_CREATED)
983                                 continue;
984
985                         /*
986                          * never got a chance to start, killed by other
987                          * thread for some reason
988                          */
989                         if (td->terminate) {
990                                 todo--;
991                                 continue;
992                         }
993
994                         if (td->start_delay) {
995                                 spent = mtime_since_genesis();
996
997                                 if (td->start_delay * 1000 > spent)
998                                         continue;
999                         }
1000
1001                         if (td->stonewall && (nr_started || nr_running))
1002                                 break;
1003
1004                         /*
1005                          * Set state to created. Thread will transition
1006                          * to TD_INITIALIZED when it's done setting up.
1007                          */
1008                         td_set_runstate(td, TD_CREATED);
1009                         map[this_jobs++] = td;
1010                         fio_sem_init(&startup_sem, 1);
1011                         nr_started++;
1012
1013                         if (td->use_thread) {
1014                                 if (pthread_create(&td->thread, NULL, thread_main, td)) {
1015                                         perror("thread_create");
1016                                         nr_started--;
1017                                 }
1018                         } else {
1019                                 if (fork())
1020                                         fio_sem_down(&startup_sem);
1021                                 else {
1022                                         int ret = fork_main(shm_id, i);
1023
1024                                         exit(ret);
1025                                 }
1026                         }
1027                 }
1028
1029                 /*
1030                  * Wait for the started threads to transition to
1031                  * TD_INITIALIZED.
1032                  */
1033                 fio_gettime(&this_start, NULL);
1034                 left = this_jobs;
1035                 while (left && !fio_abort) {
1036                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1037                                 break;
1038
1039                         usleep(100000);
1040
1041                         for (i = 0; i < this_jobs; i++) {
1042                                 td = map[i];
1043                                 if (!td)
1044                                         continue;
1045                                 if (td->runstate == TD_INITIALIZED) {
1046                                         map[i] = NULL;
1047                                         left--;
1048                                 } else if (td->runstate >= TD_EXITED) {
1049                                         map[i] = NULL;
1050                                         left--;
1051                                         todo--;
1052                                         nr_running++; /* work-around... */
1053                                 }
1054                         }
1055                 }
1056
1057                 if (left) {
1058                         log_err("fio: %d jobs failed to start\n", left);
1059                         for (i = 0; i < this_jobs; i++) {
1060                                 td = map[i];
1061                                 if (!td)
1062                                         continue;
1063                                 kill(td->pid, SIGTERM);
1064                         }
1065                         break;
1066                 }
1067
1068                 /*
1069                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1070                  */
1071                 for_each_td(td, i) {
1072                         if (td->runstate != TD_INITIALIZED)
1073                                 continue;
1074
1075                         td_set_runstate(td, TD_RUNNING);
1076                         nr_running++;
1077                         nr_started--;
1078                         m_rate += td->ratemin;
1079                         t_rate += td->rate;
1080                         todo--;
1081                         fio_sem_up(&td->mutex);
1082                 }
1083
1084                 reap_threads(&nr_running, &t_rate, &m_rate);
1085
1086                 if (todo)
1087                         usleep(100000);
1088         }
1089
1090         while (nr_running) {
1091                 reap_threads(&nr_running, &t_rate, &m_rate);
1092                 usleep(10000);
1093         }
1094
1095         update_io_ticks();
1096         fio_unpin_memory();
1097 }
1098
1099 int main(int argc, char *argv[])
1100 {
1101         long ps;
1102
1103         /*
1104          * We need locale for number printing, if it isn't set then just
1105          * go with the US format.
1106          */
1107         if (!getenv("LC_NUMERIC"))
1108                 setlocale(LC_NUMERIC, "en_US");
1109
1110         if (parse_options(argc, argv))
1111                 return 1;
1112
1113         if (!thread_number) {
1114                 log_err("Nothing to do\n");
1115                 return 1;
1116         }
1117
1118         ps = sysconf(_SC_PAGESIZE);
1119         if (ps < 0) {
1120                 log_err("Failed to get page size\n");
1121                 return 1;
1122         }
1123
1124         page_mask = ps - 1;
1125
1126         if (write_bw_log) {
1127                 setup_log(&agg_io_log[DDIR_READ]);
1128                 setup_log(&agg_io_log[DDIR_WRITE]);
1129         }
1130
1131         set_genesis_time();
1132
1133         disk_util_timer_arm();
1134
1135         run_threads();
1136
1137         if (!fio_abort) {
1138                 show_run_stats();
1139                 if (write_bw_log) {
1140                         __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1141                         __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");
1142                 }
1143         }
1144
1145         return exit_value;
1146 }