GUASI API change
[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
39 unsigned long page_mask;
40 unsigned long page_size;
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 nr_process = 0;
47 int nr_thread = 0;
48 int shm_id = 0;
49 int temp_stall_ts;
50
51 static struct fio_sem *startup_sem;
52 static volatile int fio_abort;
53 static int exit_value;
54
55 struct io_log *agg_io_log[2];
56
57 #define TERMINATE_ALL           (-1)
58 #define JOB_START_TIMEOUT       (5 * 1000)
59
60 static inline void td_set_runstate(struct thread_data *td, int runstate)
61 {
62         td->runstate = runstate;
63 }
64
65 static void terminate_threads(int group_id)
66 {
67         struct thread_data *td;
68         int i;
69
70         for_each_td(td, i) {
71                 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
72                         /*
73                          * if the thread is running, just let it exit
74                          */
75                         if (td->runstate < TD_RUNNING)
76                                 kill(td->pid, SIGQUIT);
77                         td->terminate = 1;
78                         td->o.start_delay = 0;
79                 }
80         }
81 }
82
83 static void sig_handler(int sig)
84 {
85         switch (sig) {
86                 case SIGALRM:
87                         update_io_ticks();
88                         disk_util_timer_arm();
89                         print_thread_status();
90                         break;
91                 default:
92                         printf("\nfio: terminating on signal %d\n", sig);
93                         fflush(stdout);
94                         terminate_threads(TERMINATE_ALL);
95                         break;
96         }
97 }
98
99 /*
100  * Check if we are above the minimum rate given.
101  */
102 static int check_min_rate(struct thread_data *td, struct timeval *now)
103 {
104         unsigned long long bytes = 0;
105         unsigned long iops = 0;
106         unsigned long spent;
107         unsigned long rate;
108
109         /*
110          * No minimum rate set, always ok
111          */
112         if (!td->o.ratemin && !td->o.rate_iops_min)
113                 return 0;
114
115         /*
116          * allow a 2 second settle period in the beginning
117          */
118         if (mtime_since(&td->start, now) < 2000)
119                 return 0;
120
121         if (td_read(td)) {
122                 iops += td->io_blocks[DDIR_READ];
123                 bytes += td->this_io_bytes[DDIR_READ];
124         }
125         if (td_write(td)) {
126                 iops += td->io_blocks[DDIR_WRITE];
127                 bytes += td->this_io_bytes[DDIR_WRITE];
128         }
129
130         /*
131          * if rate blocks is set, sample is running
132          */
133         if (td->rate_bytes || td->rate_blocks) {
134                 spent = mtime_since(&td->lastrate, now);
135                 if (spent < td->o.ratecycle)
136                         return 0;
137
138                 if (td->o.rate) {
139                         /*
140                          * check bandwidth specified rate
141                          */
142                         if (bytes < td->rate_bytes) {
143                                 log_err("%s: min rate %u not met\n", td->o.name, td->o.ratemin);
144                                 return 1;
145                         } else {
146                                 rate = (bytes - td->rate_bytes) / spent;
147                                 if (rate < td->o.ratemin || bytes < td->rate_bytes) {
148                                         log_err("%s: min rate %u not met, got %luKiB/sec\n", td->o.name, td->o.ratemin, rate);
149                                         return 1;
150                                 }
151                         }
152                 } else {
153                         /*
154                          * checks iops specified rate
155                          */
156                         if (iops < td->o.rate_iops) {
157                                 log_err("%s: min iops rate %u not met\n", td->o.name, td->o.rate_iops);
158                                 return 1;
159                         } else {
160                                 rate = (iops - td->rate_blocks) / spent;
161                                 if (rate < td->o.rate_iops_min || iops < td->rate_blocks) {
162                                         log_err("%s: min iops rate %u not met, got %lu\n", td->o.name, td->o.rate_iops_min, rate);
163                                 }
164                         }
165                 }
166         }
167
168         td->rate_bytes = bytes;
169         td->rate_blocks = iops;
170         memcpy(&td->lastrate, now, sizeof(*now));
171         return 0;
172 }
173
174 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
175 {
176         if (!td->o.timeout)
177                 return 0;
178         if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
179                 return 1;
180
181         return 0;
182 }
183
184 /*
185  * When job exits, we can cancel the in-flight IO if we are using async
186  * io. Attempt to do so.
187  */
188 static void cleanup_pending_aio(struct thread_data *td)
189 {
190         struct list_head *entry, *n;
191         struct io_u *io_u;
192         int r;
193
194         /*
195          * get immediately available events, if any
196          */
197         r = io_u_queued_complete(td, 0);
198         if (r < 0)
199                 return;
200
201         /*
202          * now cancel remaining active events
203          */
204         if (td->io_ops->cancel) {
205                 list_for_each_safe(entry, n, &td->io_u_busylist) {
206                         io_u = list_entry(entry, struct io_u, list);
207
208                         /*
209                          * if the io_u isn't in flight, then that generally
210                          * means someone leaked an io_u. complain but fix
211                          * it up, so we don't stall here.
212                          */
213                         if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
214                                 log_err("fio: non-busy IO on busy list\n");
215                                 put_io_u(td, io_u);
216                         } else {
217                                 r = td->io_ops->cancel(td, io_u);
218                                 if (!r)
219                                         put_io_u(td, io_u);
220                         }
221                 }
222         }
223
224         if (td->cur_depth)
225                 r = io_u_queued_complete(td, td->cur_depth);
226 }
227
228 /*
229  * Helper to handle the final sync of a file. Works just like the normal
230  * io path, just does everything sync.
231  */
232 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
233 {
234         struct io_u *io_u = __get_io_u(td);
235         int ret;
236
237         if (!io_u)
238                 return 1;
239
240         io_u->ddir = DDIR_SYNC;
241         io_u->file = f;
242
243         if (td_io_prep(td, io_u)) {
244                 put_io_u(td, io_u);
245                 return 1;
246         }
247
248 requeue:
249         ret = td_io_queue(td, io_u);
250         if (ret < 0) {
251                 td_verror(td, io_u->error, "td_io_queue");
252                 put_io_u(td, io_u);
253                 return 1;
254         } else if (ret == FIO_Q_QUEUED) {
255                 if (io_u_queued_complete(td, 1) < 0)
256                         return 1;
257         } else if (ret == FIO_Q_COMPLETED) {
258                 if (io_u->error) {
259                         td_verror(td, io_u->error, "td_io_queue");
260                         return 1;
261                 }
262
263                 if (io_u_sync_complete(td, io_u) < 0)
264                         return 1;
265         } else if (ret == FIO_Q_BUSY) {
266                 if (td_io_commit(td))
267                         return 1;
268                 goto requeue;
269         }
270
271         return 0;
272 }
273
274 /*
275  * The main verify engine. Runs over the writes we previously submitted,
276  * reads the blocks back in, and checks the crc/md5 of the data.
277  */
278 static void do_verify(struct thread_data *td)
279 {
280         struct fio_file *f;
281         struct io_u *io_u;
282         int ret, min_events;
283         unsigned int i;
284
285         /*
286          * sync io first and invalidate cache, to make sure we really
287          * read from disk.
288          */
289         for_each_file(td, f, i) {
290                 if (!(f->flags & FIO_FILE_OPEN))
291                         continue;
292                 if (fio_io_sync(td, f))
293                         break;
294                 if (file_invalidate_cache(td, f))
295                         break;
296         }
297
298         if (td->error)
299                 return;
300
301         td_set_runstate(td, TD_VERIFYING);
302
303         io_u = NULL;
304         while (!td->terminate) {
305                 int ret2;
306
307                 io_u = __get_io_u(td);
308                 if (!io_u)
309                         break;
310
311                 if (runtime_exceeded(td, &io_u->start_time)) {
312                         put_io_u(td, io_u);
313                         break;
314                 }
315
316                 if (get_next_verify(td, io_u)) {
317                         put_io_u(td, io_u);
318                         break;
319                 }
320
321                 if (td_io_prep(td, io_u)) {
322                         put_io_u(td, io_u);
323                         break;
324                 }
325
326                 io_u->end_io = verify_io_u;
327
328                 ret = td_io_queue(td, io_u);
329                 switch (ret) {
330                 case FIO_Q_COMPLETED:
331                         if (io_u->error)
332                                 ret = -io_u->error;
333                         else if (io_u->resid) {
334                                 int bytes = io_u->xfer_buflen - io_u->resid;
335
336                                 /*
337                                  * zero read, fail
338                                  */
339                                 if (!bytes) {
340                                         td_verror(td, ENODATA, "full resid");
341                                         put_io_u(td, io_u);
342                                         break;
343                                 }
344
345                                 io_u->xfer_buflen = io_u->resid;
346                                 io_u->xfer_buf += bytes;
347                                 io_u->offset += bytes;
348
349                                 if (io_u->offset == io_u->file->real_file_size)
350                                         goto sync_done;
351
352                                 requeue_io_u(td, &io_u);
353                         } else {
354 sync_done:
355                                 ret = io_u_sync_complete(td, io_u);
356                                 if (ret < 0)
357                                         break;
358                         }
359                         continue;
360                 case FIO_Q_QUEUED:
361                         break;
362                 case FIO_Q_BUSY:
363                         requeue_io_u(td, &io_u);
364                         ret2 = td_io_commit(td);
365                         if (ret2 < 0)
366                                 ret = ret2;
367                         break;
368                 default:
369                         assert(ret < 0);
370                         td_verror(td, -ret, "td_io_queue");
371                         break;
372                 }
373
374                 if (ret < 0 || td->error)
375                         break;
376
377                 /*
378                  * if we can queue more, do so. but check if there are
379                  * completed io_u's first.
380                  */
381                 min_events = 0;
382                 if (queue_full(td) || ret == FIO_Q_BUSY) {
383                         min_events = 1;
384
385                         if (td->cur_depth > td->o.iodepth_low)
386                                 min_events = td->cur_depth - td->o.iodepth_low;
387                 }
388
389                 /*
390                  * Reap required number of io units, if any, and do the
391                  * verification on them through the callback handler
392                  */
393                 if (io_u_queued_complete(td, min_events) < 0)
394                         break;
395         }
396
397         if (!td->error) {
398                 min_events = td->cur_depth;
399
400                 if (min_events)
401                         ret = io_u_queued_complete(td, min_events);
402         } else
403                 cleanup_pending_aio(td);
404
405         td_set_runstate(td, TD_RUNNING);
406 }
407
408 /*
409  * Main IO worker function. It retrieves io_u's to process and queues
410  * and reaps them, checking for rate and errors along the way.
411  */
412 static void do_io(struct thread_data *td)
413 {
414         struct timeval s;
415         unsigned long usec;
416         unsigned int i;
417         int ret = 0;
418
419         td_set_runstate(td, TD_RUNNING);
420
421         while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) {
422                 struct timeval comp_time;
423                 long bytes_done = 0;
424                 int min_evts = 0;
425                 struct io_u *io_u;
426                 int ret2;
427
428                 if (td->terminate)
429                         break;
430
431                 io_u = get_io_u(td);
432                 if (!io_u)
433                         break;
434
435                 memcpy(&s, &io_u->start_time, sizeof(s));
436
437                 if (runtime_exceeded(td, &s)) {
438                         put_io_u(td, io_u);
439                         break;
440                 }
441
442                 ret = td_io_queue(td, io_u);
443                 switch (ret) {
444                 case FIO_Q_COMPLETED:
445                         if (io_u->error)
446                                 ret = -io_u->error;
447                         else if (io_u->resid) {
448                                 int bytes = io_u->xfer_buflen - io_u->resid;
449
450                                 /*
451                                  * zero read, fail
452                                  */
453                                 if (!bytes) {
454                                         td_verror(td, ENODATA, "full resid");
455                                         put_io_u(td, io_u);
456                                         break;
457                                 }
458
459                                 io_u->xfer_buflen = io_u->resid;
460                                 io_u->xfer_buf += bytes;
461                                 io_u->offset += bytes;
462
463                                 if (io_u->offset == io_u->file->real_file_size)
464                                         goto sync_done;
465
466                                 requeue_io_u(td, &io_u);
467                         } else {
468 sync_done:
469                                 fio_gettime(&comp_time, NULL);
470                                 bytes_done = io_u_sync_complete(td, io_u);
471                                 if (bytes_done < 0)
472                                         ret = bytes_done;
473                         }
474                         break;
475                 case FIO_Q_QUEUED:
476                         /*
477                          * if the engine doesn't have a commit hook,
478                          * the io_u is really queued. if it does have such
479                          * a hook, it has to call io_u_queued() itself.
480                          */
481                         if (td->io_ops->commit == NULL)
482                                 io_u_queued(td, io_u);
483                         break;
484                 case FIO_Q_BUSY:
485                         requeue_io_u(td, &io_u);
486                         ret2 = td_io_commit(td);
487                         if (ret2 < 0)
488                                 ret = ret2;
489                         break;
490                 default:
491                         assert(ret < 0);
492                         put_io_u(td, io_u);
493                         break;
494                 }
495
496                 if (ret < 0 || td->error)
497                         break;
498
499                 /*
500                  * See if we need to complete some commands
501                  */
502                 if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
503                         min_evts = 0;
504                         if (queue_full(td) || ret == FIO_Q_BUSY) {
505                                 min_evts = 1;
506
507                                 if (td->cur_depth > td->o.iodepth_low)
508                                         min_evts = td->cur_depth - td->o.iodepth_low;
509                         }
510
511                         fio_gettime(&comp_time, NULL);
512                         bytes_done = io_u_queued_complete(td, min_evts);
513                         if (bytes_done < 0)
514                                 break;
515                 }
516
517                 if (!bytes_done)
518                         continue;
519
520                 /*
521                  * the rate is batched for now, it should work for batches
522                  * of completions except the very first one which may look
523                  * a little bursty
524                  */
525                 usec = utime_since(&s, &comp_time);
526
527                 rate_throttle(td, usec, bytes_done);
528
529                 if (check_min_rate(td, &comp_time)) {
530                         if (exitall_on_terminate)
531                                 terminate_threads(td->groupid);
532                         td_verror(td, ENODATA, "check_min_rate");
533                         break;
534                 }
535
536                 if (td->o.thinktime) {
537                         unsigned long long b;
538
539                         b = td->io_blocks[0] + td->io_blocks[1];
540                         if (!(b % td->o.thinktime_blocks)) {
541                                 int left;
542
543                                 if (td->o.thinktime_spin)
544                                         __usec_sleep(td->o.thinktime_spin);
545
546                                 left = td->o.thinktime - td->o.thinktime_spin;
547                                 if (left)
548                                         usec_sleep(td, left);
549                         }
550                 }
551         }
552
553         if (!td->error) {
554                 struct fio_file *f;
555
556                 i = td->cur_depth;
557                 if (i)
558                         ret = io_u_queued_complete(td, i);
559
560                 if (should_fsync(td) && td->o.end_fsync) {
561                         td_set_runstate(td, TD_FSYNCING);
562
563                         for_each_file(td, f, i) {
564                                 if (!(f->flags & FIO_FILE_OPEN))
565                                         continue;
566                                 fio_io_sync(td, f);
567                         }
568                 }
569         } else
570                 cleanup_pending_aio(td);
571 }
572
573 static void cleanup_io_u(struct thread_data *td)
574 {
575         struct list_head *entry, *n;
576         struct io_u *io_u;
577
578         list_for_each_safe(entry, n, &td->io_u_freelist) {
579                 io_u = list_entry(entry, struct io_u, list);
580
581                 list_del(&io_u->list);
582                 free(io_u);
583         }
584
585         free_io_mem(td);
586 }
587
588 /*
589  * "randomly" fill the buffer contents
590  */
591 static void fill_rand_buf(struct io_u *io_u, int max_bs)
592 {
593         int *ptr = io_u->buf;
594
595         while ((void *) ptr - io_u->buf < max_bs) {
596                 *ptr = rand() * 0x9e370001;
597                 ptr++;
598         }
599 }
600
601 static int init_io_u(struct thread_data *td)
602 {
603         unsigned long long buf_size;
604         struct io_u *io_u;
605         unsigned int max_bs;
606         int i, max_units;
607         char *p;
608
609         if (td->io_ops->flags & FIO_SYNCIO)
610                 max_units = 1;
611         else
612                 max_units = td->o.iodepth;
613
614         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
615         buf_size = (unsigned long long) max_bs * (unsigned long long) max_units;
616         buf_size += page_mask;
617         if (buf_size != (size_t) buf_size) {
618                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
619                 return 1;
620         }
621
622         td->orig_buffer_size = buf_size;
623
624         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE)
625                 td->orig_buffer_size = (td->orig_buffer_size + td->o.hugepage_size - 1) & ~(td->o.hugepage_size - 1);
626         else if (td->orig_buffer_size & page_mask)
627                 td->orig_buffer_size = (td->orig_buffer_size + page_mask) & ~page_mask;
628
629         if (allocate_io_mem(td))
630                 return 1;
631
632         p = ALIGN(td->orig_buffer);
633         for (i = 0; i < max_units; i++) {
634                 io_u = malloc(sizeof(*io_u));
635                 memset(io_u, 0, sizeof(*io_u));
636                 INIT_LIST_HEAD(&io_u->list);
637
638                 io_u->buf = p + max_bs * i;
639                 if (td_write(td) || td_rw(td))
640                         fill_rand_buf(io_u, max_bs);
641
642                 io_u->index = i;
643                 io_u->flags = IO_U_F_FREE;
644                 list_add(&io_u->list, &td->io_u_freelist);
645         }
646
647         io_u_init_timeout();
648
649         return 0;
650 }
651
652 static int switch_ioscheduler(struct thread_data *td)
653 {
654         char tmp[256], tmp2[128];
655         FILE *f;
656         int ret;
657
658         if (td->io_ops->flags & FIO_DISKLESSIO)
659                 return 0;
660
661         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
662
663         f = fopen(tmp, "r+");
664         if (!f) {
665                 td_verror(td, errno, "fopen");
666                 return 1;
667         }
668
669         /*
670          * Set io scheduler.
671          */
672         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
673         if (ferror(f) || ret != 1) {
674                 td_verror(td, errno, "fwrite");
675                 fclose(f);
676                 return 1;
677         }
678
679         rewind(f);
680
681         /*
682          * Read back and check that the selected scheduler is now the default.
683          */
684         ret = fread(tmp, 1, sizeof(tmp), f);
685         if (ferror(f) || ret < 0) {
686                 td_verror(td, errno, "fread");
687                 fclose(f);
688                 return 1;
689         }
690
691         sprintf(tmp2, "[%s]", td->o.ioscheduler);
692         if (!strstr(tmp, tmp2)) {
693                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
694                 td_verror(td, EINVAL, "iosched_switch");
695                 fclose(f);
696                 return 1;
697         }
698
699         fclose(f);
700         return 0;
701 }
702
703 static int clear_io_state(struct thread_data *td)
704 {
705         struct fio_file *f;
706         unsigned int i;
707         int ret;
708
709         td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
710         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
711         td->zone_bytes = 0;
712         td->rate_bytes = 0;
713         td->rate_blocks = 0;
714         td->rw_end_set[0] = td->rw_end_set[1] = 0;
715
716         td->last_was_sync = 0;
717
718         for_each_file(td, f, i)
719                 td_io_close_file(td, f);
720
721         ret = 0;
722         for_each_file(td, f, i) {
723                 ret = td_io_open_file(td, f);
724                 if (ret)
725                         break;
726         }
727
728         return ret;
729 }
730
731 /*
732  * Entry point for the thread based jobs. The process based jobs end up
733  * here as well, after a little setup.
734  */
735 static void *thread_main(void *data)
736 {
737         unsigned long long runtime[2];
738         struct thread_data *td = data;
739         unsigned long elapsed;
740         int clear_state;
741
742         if (!td->o.use_thread)
743                 setsid();
744
745         td->pid = getpid();
746
747         INIT_LIST_HEAD(&td->io_u_freelist);
748         INIT_LIST_HEAD(&td->io_u_busylist);
749         INIT_LIST_HEAD(&td->io_u_requeues);
750         INIT_LIST_HEAD(&td->io_log_list);
751         INIT_LIST_HEAD(&td->io_hist_list);
752         td->io_hist_tree = RB_ROOT;
753
754         if (init_io_u(td))
755                 goto err_sem;
756
757         if (fio_setaffinity(td) == -1) {
758                 td_verror(td, errno, "cpu_set_affinity");
759                 goto err_sem;
760         }
761
762         if (init_iolog(td))
763                 goto err_sem;
764
765         if (td->ioprio) {
766                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
767                         td_verror(td, errno, "ioprio_set");
768                         goto err_sem;
769                 }
770         }
771
772         if (nice(td->o.nice) == -1) {
773                 td_verror(td, errno, "nice");
774                 goto err_sem;
775         }
776
777         if (td->o.ioscheduler && switch_ioscheduler(td))
778                 goto err_sem;
779
780         td_set_runstate(td, TD_INITIALIZED);
781         fio_sem_up(startup_sem);
782         fio_sem_down(td->mutex);
783
784         /*
785          * the ->mutex semaphore is now no longer used, close it to avoid
786          * eating a file descriptor
787          */
788         fio_sem_remove(td->mutex);
789
790         if (!td->o.create_serialize && setup_files(td))
791                 goto err;
792
793         if (td_io_init(td))
794                 goto err;
795
796         if (open_files(td))
797                 goto err;
798
799         if (init_random_map(td))
800                 goto err;
801
802         if (td->o.exec_prerun) {
803                 if (system(td->o.exec_prerun) < 0)
804                         goto err;
805         }
806
807         fio_gettime(&td->epoch, NULL);
808         memcpy(&td->timeout_end, &td->epoch, sizeof(td->epoch));
809         getrusage(RUSAGE_SELF, &td->ts.ru_start);
810
811         runtime[0] = runtime[1] = 0;
812         clear_state = 0;
813         while (td->o.loops--) {
814                 fio_gettime(&td->start, NULL);
815                 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
816
817                 if (td->o.ratemin)
818                         memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
819
820                 if (clear_state && clear_io_state(td))
821                         break;
822
823                 prune_io_piece_log(td);
824
825                 do_io(td);
826
827                 clear_state = 1;
828
829                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
830                         if (td->rw_end_set[DDIR_READ])
831                                 elapsed = utime_since(&td->start, &td->rw_end[DDIR_READ]);
832                         else
833                                 elapsed = utime_since_now(&td->start);
834
835                         runtime[DDIR_READ] += elapsed;
836                 }
837                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
838                         if (td->rw_end_set[DDIR_WRITE])
839                                 elapsed = utime_since(&td->start, &td->rw_end[DDIR_WRITE]);
840                         else
841                                 elapsed = utime_since_now(&td->start);
842
843                         runtime[DDIR_WRITE] += elapsed;
844                 }
845                 
846                 if (td->error || td->terminate)
847                         break;
848
849                 if (td->o.verify == VERIFY_NONE)
850                         continue;
851
852                 if (clear_io_state(td))
853                         break;
854
855                 fio_gettime(&td->start, NULL);
856
857                 do_verify(td);
858
859                 runtime[DDIR_READ] += utime_since_now(&td->start);
860
861                 if (td->error || td->terminate)
862                         break;
863         }
864
865         update_rusage_stat(td);
866         td->ts.runtime[0] = runtime[0] / 1000;
867         td->ts.runtime[1] = runtime[1] / 1000;
868         td->ts.total_run_time = mtime_since_now(&td->epoch);
869         td->ts.io_bytes[0] = td->io_bytes[0];
870         td->ts.io_bytes[1] = td->io_bytes[1];
871
872         if (td->ts.bw_log)
873                 finish_log(td, td->ts.bw_log, "bw");
874         if (td->ts.slat_log)
875                 finish_log(td, td->ts.slat_log, "slat");
876         if (td->ts.clat_log)
877                 finish_log(td, td->ts.clat_log, "clat");
878         if (td->o.write_iolog_file)
879                 write_iolog_close(td);
880         if (td->o.exec_postrun) {
881                 if (system(td->o.exec_postrun) < 0)
882                         log_err("fio: postrun %s failed\n", td->o.exec_postrun);
883         }
884
885         if (exitall_on_terminate)
886                 terminate_threads(td->groupid);
887
888 err:
889         if (td->error)
890                 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
891         close_files(td);
892         close_ioengine(td);
893         cleanup_io_u(td);
894         options_mem_free(td);
895         td_set_runstate(td, TD_EXITED);
896         return (void *) (unsigned long) td->error;
897 err_sem:
898         fio_sem_up(startup_sem);
899         goto err;
900 }
901
902 /*
903  * We cannot pass the td data into a forked process, so attach the td and
904  * pass it to the thread worker.
905  */
906 static int fork_main(int shmid, int offset)
907 {
908         struct thread_data *td;
909         void *data, *ret;
910
911         data = shmat(shmid, NULL, 0);
912         if (data == (void *) -1) {
913                 int __err = errno;
914
915                 perror("shmat");
916                 return __err;
917         }
918
919         td = data + offset * sizeof(struct thread_data);
920         ret = thread_main(td);
921         shmdt(data);
922         return (int) (unsigned long) ret;
923 }
924
925 /*
926  * Run over the job map and reap the threads that have exited, if any.
927  */
928 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
929 {
930         struct thread_data *td;
931         int i, cputhreads, pending, status, ret;
932
933         /*
934          * reap exited threads (TD_EXITED -> TD_REAPED)
935          */
936         pending = cputhreads = 0;
937         for_each_td(td, i) {
938                 int flags = 0;
939
940                 /*
941                  * ->io_ops is NULL for a thread that has closed its
942                  * io engine
943                  */
944                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
945                         cputhreads++;
946
947                 if (!td->pid || td->runstate == TD_REAPED)
948                         continue;
949                 if (td->o.use_thread) {
950                         if (td->runstate == TD_EXITED) {
951                                 td_set_runstate(td, TD_REAPED);
952                                 goto reaped;
953                         }
954                         continue;
955                 }
956
957                 flags = WNOHANG;
958                 if (td->runstate == TD_EXITED)
959                         flags = 0;
960
961                 /*
962                  * check if someone quit or got killed in an unusual way
963                  */
964                 ret = waitpid(td->pid, &status, flags);
965                 if (ret < 0) {
966                         if (errno == ECHILD) {
967                                 log_err("fio: pid=%d disappeared %d\n", td->pid, td->runstate);
968                                 td_set_runstate(td, TD_REAPED);
969                                 goto reaped;
970                         }
971                         perror("waitpid");
972                 } else if (ret == td->pid) {
973                         if (WIFSIGNALED(status)) {
974                                 int sig = WTERMSIG(status);
975
976                                 if (sig != SIGQUIT)
977                                         log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
978                                 td_set_runstate(td, TD_REAPED);
979                                 goto reaped;
980                         }
981                         if (WIFEXITED(status)) {
982                                 if (WEXITSTATUS(status) && !td->error)
983                                         td->error = WEXITSTATUS(status);
984
985                                 td_set_runstate(td, TD_REAPED);
986                                 goto reaped;
987                         }
988                 }
989
990                 /*
991                  * thread is not dead, continue
992                  */
993                 continue;
994 reaped:
995                 if (td->o.use_thread) {
996                         long ret;
997
998                         if (pthread_join(td->thread, (void *) &ret))
999                                 perror("pthread_join");
1000                 }
1001
1002                 (*nr_running)--;
1003                 (*m_rate) -= td->o.ratemin;
1004                 (*t_rate) -= td->o.rate;
1005
1006                 if (td->error)
1007                         exit_value++;
1008         }
1009
1010         if (*nr_running == cputhreads && !pending)
1011                 terminate_threads(TERMINATE_ALL);
1012 }
1013
1014 /*
1015  * Main function for kicking off and reaping jobs, as needed.
1016  */
1017 static void run_threads(void)
1018 {
1019         struct thread_data *td;
1020         unsigned long spent;
1021         int i, todo, nr_running, m_rate, t_rate, nr_started;
1022
1023         if (fio_pin_memory())
1024                 return;
1025
1026         if (!terse_output) {
1027                 printf("Starting ");
1028                 if (nr_thread)
1029                         printf("%d thread%s", nr_thread, nr_thread > 1 ? "s" : "");
1030                 if (nr_process) {
1031                         if (nr_thread)
1032                                 printf(" and ");
1033                         printf("%d process%s", nr_process, nr_process > 1 ? "es" : "");
1034                 }
1035                 printf("\n");
1036                 fflush(stdout);
1037         }
1038
1039         signal(SIGINT, sig_handler);
1040         signal(SIGALRM, sig_handler);
1041
1042         todo = thread_number;
1043         nr_running = 0;
1044         nr_started = 0;
1045         m_rate = t_rate = 0;
1046
1047         for_each_td(td, i) {
1048                 print_status_init(td->thread_number - 1);
1049
1050                 if (!td->o.create_serialize) {
1051                         init_disk_util(td);
1052                         continue;
1053                 }
1054
1055                 /*
1056                  * do file setup here so it happens sequentially,
1057                  * we don't want X number of threads getting their
1058                  * client data interspersed on disk
1059                  */
1060                 if (setup_files(td)) {
1061                         exit_value++;
1062                         if (td->error)
1063                                 log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
1064                         td_set_runstate(td, TD_REAPED);
1065                         todo--;
1066                 }
1067
1068                 init_disk_util(td);
1069         }
1070
1071         set_genesis_time();
1072
1073         while (todo) {
1074                 struct thread_data *map[MAX_JOBS];
1075                 struct timeval this_start;
1076                 int this_jobs = 0, left;
1077
1078                 /*
1079                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1080                  */
1081                 for_each_td(td, i) {
1082                         if (td->runstate != TD_NOT_CREATED)
1083                                 continue;
1084
1085                         /*
1086                          * never got a chance to start, killed by other
1087                          * thread for some reason
1088                          */
1089                         if (td->terminate) {
1090                                 todo--;
1091                                 continue;
1092                         }
1093
1094                         if (td->o.start_delay) {
1095                                 spent = mtime_since_genesis();
1096
1097                                 if (td->o.start_delay * 1000 > spent)
1098                                         continue;
1099                         }
1100
1101                         if (td->o.stonewall && (nr_started || nr_running))
1102                                 break;
1103
1104                         /*
1105                          * Set state to created. Thread will transition
1106                          * to TD_INITIALIZED when it's done setting up.
1107                          */
1108                         td_set_runstate(td, TD_CREATED);
1109                         map[this_jobs++] = td;
1110                         nr_started++;
1111
1112                         if (td->o.use_thread) {
1113                                 if (pthread_create(&td->thread, NULL, thread_main, td)) {
1114                                         perror("thread_create");
1115                                         nr_started--;
1116                                         break;
1117                                 }
1118                         } else {
1119                                 if (!fork()) {
1120                                         int ret = fork_main(shm_id, i);
1121
1122                                         exit(ret);
1123                                 }
1124                         }
1125                         fio_sem_down(startup_sem);
1126                 }
1127
1128                 /*
1129                  * Wait for the started threads to transition to
1130                  * TD_INITIALIZED.
1131                  */
1132                 fio_gettime(&this_start, NULL);
1133                 left = this_jobs;
1134                 while (left && !fio_abort) {
1135                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1136                                 break;
1137
1138                         usleep(100000);
1139
1140                         for (i = 0; i < this_jobs; i++) {
1141                                 td = map[i];
1142                                 if (!td)
1143                                         continue;
1144                                 if (td->runstate == TD_INITIALIZED) {
1145                                         map[i] = NULL;
1146                                         left--;
1147                                 } else if (td->runstate >= TD_EXITED) {
1148                                         map[i] = NULL;
1149                                         left--;
1150                                         todo--;
1151                                         nr_running++; /* work-around... */
1152                                 }
1153                         }
1154                 }
1155
1156                 if (left) {
1157                         log_err("fio: %d jobs failed to start\n", left);
1158                         for (i = 0; i < this_jobs; i++) {
1159                                 td = map[i];
1160                                 if (!td)
1161                                         continue;
1162                                 kill(td->pid, SIGTERM);
1163                         }
1164                         break;
1165                 }
1166
1167                 /*
1168                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1169                  */
1170                 for_each_td(td, i) {
1171                         if (td->runstate != TD_INITIALIZED)
1172                                 continue;
1173
1174                         td_set_runstate(td, TD_RUNNING);
1175                         nr_running++;
1176                         nr_started--;
1177                         m_rate += td->o.ratemin;
1178                         t_rate += td->o.rate;
1179                         todo--;
1180                         fio_sem_up(td->mutex);
1181                 }
1182
1183                 reap_threads(&nr_running, &t_rate, &m_rate);
1184
1185                 if (todo)
1186                         usleep(100000);
1187         }
1188
1189         while (nr_running) {
1190                 reap_threads(&nr_running, &t_rate, &m_rate);
1191                 usleep(10000);
1192         }
1193
1194         update_io_ticks();
1195         fio_unpin_memory();
1196 }
1197
1198 int main(int argc, char *argv[])
1199 {
1200         long ps;
1201
1202         /*
1203          * We need locale for number printing, if it isn't set then just
1204          * go with the US format.
1205          */
1206         if (!getenv("LC_NUMERIC"))
1207                 setlocale(LC_NUMERIC, "en_US");
1208
1209         if (parse_options(argc, argv))
1210                 return 1;
1211
1212         if (!thread_number)
1213                 return 0;
1214
1215         ps = sysconf(_SC_PAGESIZE);
1216         if (ps < 0) {
1217                 log_err("Failed to get page size\n");
1218                 return 1;
1219         }
1220
1221         page_size = ps;
1222         page_mask = ps - 1;
1223
1224         if (write_bw_log) {
1225                 setup_log(&agg_io_log[DDIR_READ]);
1226                 setup_log(&agg_io_log[DDIR_WRITE]);
1227         }
1228
1229         startup_sem = fio_sem_init(0);
1230
1231         set_genesis_time();
1232
1233         disk_util_timer_arm();
1234
1235         run_threads();
1236
1237         if (!fio_abort) {
1238                 show_run_stats();
1239                 if (write_bw_log) {
1240                         __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1241                         __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");
1242                 }
1243         }
1244
1245         fio_sem_remove(startup_sem);
1246         return exit_value;
1247 }