Missed FIO_UNIDIR io ops flags on net engine
[fio.git] / fio.c
1 /*
2  * fio - the flexible io tester
3  *
4  * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
5  * Copyright (C) 2006 Jens Axboe <axboe@kernel.dk>
6  *
7  * The license below covers all files distributed with fio unless otherwise
8  * noted in the file itself.
9  *
10  *  This program is free software; you can redistribute it and/or modify
11  *  it under the terms of the GNU General Public License version 2 as
12  *  published by the Free Software Foundation.
13  *
14  *  This program is distributed in the hope that it will be useful,
15  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
16  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
17  *  GNU General Public License for more details.
18  *
19  *  You should have received a copy of the GNU General Public License
20  *  along with this program; if not, write to the Free Software
21  *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
22  *
23  */
24 #include <unistd.h>
25 #include <fcntl.h>
26 #include <string.h>
27 #include <signal.h>
28 #include <time.h>
29 #include <locale.h>
30 #include <assert.h>
31 #include <sys/stat.h>
32 #include <sys/wait.h>
33 #include <sys/ipc.h>
34 #include <sys/shm.h>
35 #include <sys/mman.h>
36
37 #include "fio.h"
38 #include "hash.h"
39
40 unsigned long page_mask;
41 unsigned long page_size;
42 #define ALIGN(buf)      \
43         (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
44
45 int groupid = 0;
46 int thread_number = 0;
47 int nr_process = 0;
48 int nr_thread = 0;
49 int shm_id = 0;
50 int temp_stall_ts;
51
52 static struct fio_sem *startup_sem;
53 static volatile int fio_abort;
54 static int exit_value;
55
56 struct io_log *agg_io_log[2];
57
58 #define TERMINATE_ALL           (-1)
59 #define JOB_START_TIMEOUT       (5 * 1000)
60
61 static inline void td_set_runstate(struct thread_data *td, int runstate)
62 {
63         td->runstate = runstate;
64 }
65
66 static void terminate_threads(int group_id)
67 {
68         struct thread_data *td;
69         int i;
70
71         for_each_td(td, i) {
72                 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
73                         /*
74                          * if the thread is running, just let it exit
75                          */
76                         if (td->runstate < TD_RUNNING)
77                                 kill(td->pid, SIGQUIT);
78                         td->terminate = 1;
79                         td->o.start_delay = 0;
80                 }
81         }
82 }
83
84 static void sig_handler(int sig)
85 {
86         switch (sig) {
87                 case SIGALRM:
88                         update_io_ticks();
89                         disk_util_timer_arm();
90                         print_thread_status();
91                         break;
92                 default:
93                         printf("\nfio: terminating on signal %d\n", sig);
94                         fflush(stdout);
95                         terminate_threads(TERMINATE_ALL);
96                         break;
97         }
98 }
99
100 /*
101  * Check if we are above the minimum rate given.
102  */
103 static int check_min_rate(struct thread_data *td, struct timeval *now)
104 {
105         unsigned long long bytes = 0;
106         unsigned long iops = 0;
107         unsigned long spent;
108         unsigned long rate;
109
110         /*
111          * No minimum rate set, always ok
112          */
113         if (!td->o.ratemin && !td->o.rate_iops_min)
114                 return 0;
115
116         /*
117          * allow a 2 second settle period in the beginning
118          */
119         if (mtime_since(&td->start, now) < 2000)
120                 return 0;
121
122         if (td_read(td)) {
123                 iops += td->io_blocks[DDIR_READ];
124                 bytes += td->this_io_bytes[DDIR_READ];
125         }
126         if (td_write(td)) {
127                 iops += td->io_blocks[DDIR_WRITE];
128                 bytes += td->this_io_bytes[DDIR_WRITE];
129         }
130
131         /*
132          * if rate blocks is set, sample is running
133          */
134         if (td->rate_bytes || td->rate_blocks) {
135                 spent = mtime_since(&td->lastrate, now);
136                 if (spent < td->o.ratecycle)
137                         return 0;
138
139                 if (td->o.rate) {
140                         /*
141                          * check bandwidth specified rate
142                          */
143                         if (bytes < td->rate_bytes) {
144                                 log_err("%s: min rate %u not met\n", td->o.name, td->o.ratemin);
145                                 return 1;
146                         } else {
147                                 rate = (bytes - td->rate_bytes) / spent;
148                                 if (rate < td->o.ratemin || bytes < td->rate_bytes) {
149                                         log_err("%s: min rate %u not met, got %luKiB/sec\n", td->o.name, td->o.ratemin, rate);
150                                         return 1;
151                                 }
152                         }
153                 } else {
154                         /*
155                          * checks iops specified rate
156                          */
157                         if (iops < td->o.rate_iops) {
158                                 log_err("%s: min iops rate %u not met\n", td->o.name, td->o.rate_iops);
159                                 return 1;
160                         } else {
161                                 rate = (iops - td->rate_blocks) / spent;
162                                 if (rate < td->o.rate_iops_min || iops < td->rate_blocks) {
163                                         log_err("%s: min iops rate %u not met, got %lu\n", td->o.name, td->o.rate_iops_min, rate);
164                                 }
165                         }
166                 }
167         }
168
169         td->rate_bytes = bytes;
170         td->rate_blocks = iops;
171         memcpy(&td->lastrate, now, sizeof(*now));
172         return 0;
173 }
174
175 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
176 {
177         if (!td->o.timeout)
178                 return 0;
179         if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
180                 return 1;
181
182         return 0;
183 }
184
185 /*
186  * When job exits, we can cancel the in-flight IO if we are using async
187  * io. Attempt to do so.
188  */
189 static void cleanup_pending_aio(struct thread_data *td)
190 {
191         struct list_head *entry, *n;
192         struct io_u *io_u;
193         int r;
194
195         /*
196          * get immediately available events, if any
197          */
198         r = io_u_queued_complete(td, 0);
199         if (r < 0)
200                 return;
201
202         /*
203          * now cancel remaining active events
204          */
205         if (td->io_ops->cancel) {
206                 list_for_each_safe(entry, n, &td->io_u_busylist) {
207                         io_u = list_entry(entry, struct io_u, list);
208
209                         /*
210                          * if the io_u isn't in flight, then that generally
211                          * means someone leaked an io_u. complain but fix
212                          * it up, so we don't stall here.
213                          */
214                         if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
215                                 log_err("fio: non-busy IO on busy list\n");
216                                 put_io_u(td, io_u);
217                         } else {
218                                 r = td->io_ops->cancel(td, io_u);
219                                 if (!r)
220                                         put_io_u(td, io_u);
221                         }
222                 }
223         }
224
225         if (td->cur_depth)
226                 r = io_u_queued_complete(td, td->cur_depth);
227 }
228
229 /*
230  * Helper to handle the final sync of a file. Works just like the normal
231  * io path, just does everything sync.
232  */
233 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
234 {
235         struct io_u *io_u = __get_io_u(td);
236         int ret;
237
238         if (!io_u)
239                 return 1;
240
241         io_u->ddir = DDIR_SYNC;
242         io_u->file = f;
243
244         if (td_io_prep(td, io_u)) {
245                 put_io_u(td, io_u);
246                 return 1;
247         }
248
249 requeue:
250         ret = td_io_queue(td, io_u);
251         if (ret < 0) {
252                 td_verror(td, io_u->error, "td_io_queue");
253                 put_io_u(td, io_u);
254                 return 1;
255         } else if (ret == FIO_Q_QUEUED) {
256                 if (io_u_queued_complete(td, 1) < 0)
257                         return 1;
258         } else if (ret == FIO_Q_COMPLETED) {
259                 if (io_u->error) {
260                         td_verror(td, io_u->error, "td_io_queue");
261                         return 1;
262                 }
263
264                 if (io_u_sync_complete(td, io_u) < 0)
265                         return 1;
266         } else if (ret == FIO_Q_BUSY) {
267                 if (td_io_commit(td))
268                         return 1;
269                 goto requeue;
270         }
271
272         return 0;
273 }
274
275 /*
276  * The main verify engine. Runs over the writes we previously submitted,
277  * reads the blocks back in, and checks the crc/md5 of the data.
278  */
279 static void do_verify(struct thread_data *td)
280 {
281         struct fio_file *f;
282         struct io_u *io_u;
283         int ret, min_events;
284         unsigned int i;
285
286         /*
287          * sync io first and invalidate cache, to make sure we really
288          * read from disk.
289          */
290         for_each_file(td, f, i) {
291                 if (!(f->flags & FIO_FILE_OPEN))
292                         continue;
293                 if (fio_io_sync(td, f))
294                         break;
295                 if (file_invalidate_cache(td, f))
296                         break;
297         }
298
299         if (td->error)
300                 return;
301
302         td_set_runstate(td, TD_VERIFYING);
303
304         io_u = NULL;
305         while (!td->terminate) {
306                 int ret2;
307
308                 io_u = __get_io_u(td);
309                 if (!io_u)
310                         break;
311
312                 if (runtime_exceeded(td, &io_u->start_time)) {
313                         put_io_u(td, io_u);
314                         td->terminate = 1;
315                         break;
316                 }
317
318                 if (get_next_verify(td, io_u)) {
319                         put_io_u(td, io_u);
320                         break;
321                 }
322
323                 if (td_io_prep(td, io_u)) {
324                         put_io_u(td, io_u);
325                         break;
326                 }
327
328                 io_u->end_io = verify_io_u;
329
330                 ret = td_io_queue(td, io_u);
331                 switch (ret) {
332                 case FIO_Q_COMPLETED:
333                         if (io_u->error)
334                                 ret = -io_u->error;
335                         else if (io_u->resid) {
336                                 int bytes = io_u->xfer_buflen - io_u->resid;
337                                 struct fio_file *f = io_u->file;
338
339                                 /*
340                                  * zero read, fail
341                                  */
342                                 if (!bytes) {
343                                         td_verror(td, ENODATA, "full resid");
344                                         put_io_u(td, io_u);
345                                         break;
346                                 }
347
348                                 io_u->xfer_buflen = io_u->resid;
349                                 io_u->xfer_buf += bytes;
350                                 io_u->offset += bytes;
351                                 f->last_completed_pos = io_u->offset;
352
353                                 td->ts.short_io_u[io_u->ddir]++;
354
355                                 if (io_u->offset == f->real_file_size)
356                                         goto sync_done;
357
358                                 requeue_io_u(td, &io_u);
359                         } else {
360 sync_done:
361                                 ret = io_u_sync_complete(td, io_u);
362                                 if (ret < 0)
363                                         break;
364                         }
365                         continue;
366                 case FIO_Q_QUEUED:
367                         break;
368                 case FIO_Q_BUSY:
369                         requeue_io_u(td, &io_u);
370                         ret2 = td_io_commit(td);
371                         if (ret2 < 0)
372                                 ret = ret2;
373                         break;
374                 default:
375                         assert(ret < 0);
376                         td_verror(td, -ret, "td_io_queue");
377                         break;
378                 }
379
380                 if (ret < 0 || td->error)
381                         break;
382
383                 /*
384                  * if we can queue more, do so. but check if there are
385                  * completed io_u's first.
386                  */
387                 min_events = 0;
388                 if (queue_full(td) || ret == FIO_Q_BUSY) {
389                         min_events = 1;
390
391                         if (td->cur_depth > td->o.iodepth_low)
392                                 min_events = td->cur_depth - td->o.iodepth_low;
393                 }
394
395                 /*
396                  * Reap required number of io units, if any, and do the
397                  * verification on them through the callback handler
398                  */
399                 if (io_u_queued_complete(td, min_events) < 0)
400                         break;
401         }
402
403         if (!td->error) {
404                 min_events = td->cur_depth;
405
406                 if (min_events)
407                         ret = io_u_queued_complete(td, min_events);
408         } else
409                 cleanup_pending_aio(td);
410
411         td_set_runstate(td, TD_RUNNING);
412 }
413
414 /*
415  * Main IO worker function. It retrieves io_u's to process and queues
416  * and reaps them, checking for rate and errors along the way.
417  */
418 static void do_io(struct thread_data *td)
419 {
420         struct timeval s;
421         unsigned long usec;
422         unsigned int i;
423         int ret = 0;
424
425         td_set_runstate(td, TD_RUNNING);
426
427         while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) {
428                 struct timeval comp_time;
429                 long bytes_done = 0;
430                 int min_evts = 0;
431                 struct io_u *io_u;
432                 int ret2;
433
434                 if (td->terminate)
435                         break;
436
437                 io_u = get_io_u(td);
438                 if (!io_u)
439                         break;
440
441                 memcpy(&s, &io_u->start_time, sizeof(s));
442
443                 if (runtime_exceeded(td, &s)) {
444                         put_io_u(td, io_u);
445                         td->terminate = 1;
446                         break;
447                 }
448
449                 /*
450                  * Add verification end_io handler, if asked to verify
451                  * a previously written file.
452                  */
453                 if (td->o.verify != VERIFY_NONE)
454                         io_u->end_io = verify_io_u;
455
456                 ret = td_io_queue(td, io_u);
457                 switch (ret) {
458                 case FIO_Q_COMPLETED:
459                         if (io_u->error)
460                                 ret = -io_u->error;
461                         else if (io_u->resid) {
462                                 int bytes = io_u->xfer_buflen - io_u->resid;
463                                 struct fio_file *f = io_u->file;
464
465                                 /*
466                                  * zero read, fail
467                                  */
468                                 if (!bytes) {
469                                         td_verror(td, ENODATA, "full resid");
470                                         put_io_u(td, io_u);
471                                         break;
472                                 }
473
474                                 io_u->xfer_buflen = io_u->resid;
475                                 io_u->xfer_buf += bytes;
476                                 io_u->offset += bytes;
477                                 f->last_completed_pos = io_u->offset;
478
479                                 td->ts.short_io_u[io_u->ddir]++;
480
481                                 if (io_u->offset == f->real_file_size)
482                                         goto sync_done;
483
484                                 requeue_io_u(td, &io_u);
485                         } else {
486 sync_done:
487                                 fio_gettime(&comp_time, NULL);
488                                 bytes_done = io_u_sync_complete(td, io_u);
489                                 if (bytes_done < 0)
490                                         ret = bytes_done;
491                         }
492                         break;
493                 case FIO_Q_QUEUED:
494                         /*
495                          * if the engine doesn't have a commit hook,
496                          * the io_u is really queued. if it does have such
497                          * a hook, it has to call io_u_queued() itself.
498                          */
499                         if (td->io_ops->commit == NULL)
500                                 io_u_queued(td, io_u);
501                         break;
502                 case FIO_Q_BUSY:
503                         requeue_io_u(td, &io_u);
504                         ret2 = td_io_commit(td);
505                         if (ret2 < 0)
506                                 ret = ret2;
507                         break;
508                 default:
509                         assert(ret < 0);
510                         put_io_u(td, io_u);
511                         break;
512                 }
513
514                 if (ret < 0 || td->error)
515                         break;
516
517                 /*
518                  * See if we need to complete some commands
519                  */
520                 if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
521                         min_evts = 0;
522                         if (queue_full(td) || ret == FIO_Q_BUSY) {
523                                 min_evts = 1;
524
525                                 if (td->cur_depth > td->o.iodepth_low)
526                                         min_evts = td->cur_depth - td->o.iodepth_low;
527                         }
528
529                         fio_gettime(&comp_time, NULL);
530                         bytes_done = io_u_queued_complete(td, min_evts);
531                         if (bytes_done < 0)
532                                 break;
533                 }
534
535                 if (!bytes_done)
536                         continue;
537
538                 /*
539                  * the rate is batched for now, it should work for batches
540                  * of completions except the very first one which may look
541                  * a little bursty
542                  */
543                 usec = utime_since(&s, &comp_time);
544
545                 rate_throttle(td, usec, bytes_done);
546
547                 if (check_min_rate(td, &comp_time)) {
548                         if (exitall_on_terminate)
549                                 terminate_threads(td->groupid);
550                         td_verror(td, ENODATA, "check_min_rate");
551                         break;
552                 }
553
554                 if (td->o.thinktime) {
555                         unsigned long long b;
556
557                         b = td->io_blocks[0] + td->io_blocks[1];
558                         if (!(b % td->o.thinktime_blocks)) {
559                                 int left;
560
561                                 if (td->o.thinktime_spin)
562                                         __usec_sleep(td->o.thinktime_spin);
563
564                                 left = td->o.thinktime - td->o.thinktime_spin;
565                                 if (left)
566                                         usec_sleep(td, left);
567                         }
568                 }
569         }
570
571         if (!td->error) {
572                 struct fio_file *f;
573
574                 i = td->cur_depth;
575                 if (i)
576                         ret = io_u_queued_complete(td, i);
577
578                 if (should_fsync(td) && td->o.end_fsync) {
579                         td_set_runstate(td, TD_FSYNCING);
580
581                         for_each_file(td, f, i) {
582                                 if (!(f->flags & FIO_FILE_OPEN))
583                                         continue;
584                                 fio_io_sync(td, f);
585                         }
586                 }
587         } else
588                 cleanup_pending_aio(td);
589 }
590
591 static void cleanup_io_u(struct thread_data *td)
592 {
593         struct list_head *entry, *n;
594         struct io_u *io_u;
595
596         list_for_each_safe(entry, n, &td->io_u_freelist) {
597                 io_u = list_entry(entry, struct io_u, list);
598
599                 list_del(&io_u->list);
600                 free(io_u);
601         }
602
603         free_io_mem(td);
604 }
605
606 /*
607  * "randomly" fill the buffer contents
608  */
609 static void fill_io_buf(struct thread_data *td, struct io_u *io_u, int max_bs)
610 {
611         long *ptr = io_u->buf;
612
613         if (!td->o.zero_buffers) {
614                 while ((void *) ptr - io_u->buf < max_bs) {
615                         *ptr = rand() * GOLDEN_RATIO_PRIME;
616                         ptr++;
617                 }
618         } else
619                 memset(ptr, 0, max_bs);
620 }
621
622 static int init_io_u(struct thread_data *td)
623 {
624         struct io_u *io_u;
625         unsigned int max_bs;
626         int i, max_units;
627         char *p;
628
629         if (td->io_ops->flags & FIO_SYNCIO)
630                 max_units = 1;
631         else
632                 max_units = td->o.iodepth;
633
634         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
635         td->orig_buffer_size = (unsigned long long) max_bs * (unsigned long long) max_units;
636
637         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE)
638                 td->orig_buffer_size = (td->orig_buffer_size + td->o.hugepage_size - 1) & ~(td->o.hugepage_size - 1);
639
640         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
641                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
642                 return 1;
643         }
644
645         if (allocate_io_mem(td))
646                 return 1;
647
648         if (td->o.odirect)
649                 p = ALIGN(td->orig_buffer);
650         else
651                 p = td->orig_buffer;
652
653         for (i = 0; i < max_units; i++) {
654                 if (td->terminate)
655                         return 1;
656                 io_u = malloc(sizeof(*io_u));
657                 memset(io_u, 0, sizeof(*io_u));
658                 INIT_LIST_HEAD(&io_u->list);
659
660                 io_u->buf = p + max_bs * i;
661
662                 if (td_write(td))
663                         fill_io_buf(td, io_u, max_bs);
664
665                 io_u->index = i;
666                 io_u->flags = IO_U_F_FREE;
667                 list_add(&io_u->list, &td->io_u_freelist);
668         }
669
670         io_u_init_timeout();
671
672         return 0;
673 }
674
675 static int switch_ioscheduler(struct thread_data *td)
676 {
677         char tmp[256], tmp2[128];
678         FILE *f;
679         int ret;
680
681         if (td->io_ops->flags & FIO_DISKLESSIO)
682                 return 0;
683
684         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
685
686         f = fopen(tmp, "r+");
687         if (!f) {
688                 if (errno == ENOENT) {
689                         log_err("fio: os or kernel doesn't support IO scheduler switching\n");
690                         return 0;
691                 }
692                 td_verror(td, errno, "fopen iosched");
693                 return 1;
694         }
695
696         /*
697          * Set io scheduler.
698          */
699         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
700         if (ferror(f) || ret != 1) {
701                 td_verror(td, errno, "fwrite");
702                 fclose(f);
703                 return 1;
704         }
705
706         rewind(f);
707
708         /*
709          * Read back and check that the selected scheduler is now the default.
710          */
711         ret = fread(tmp, 1, sizeof(tmp), f);
712         if (ferror(f) || ret < 0) {
713                 td_verror(td, errno, "fread");
714                 fclose(f);
715                 return 1;
716         }
717
718         sprintf(tmp2, "[%s]", td->o.ioscheduler);
719         if (!strstr(tmp, tmp2)) {
720                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
721                 td_verror(td, EINVAL, "iosched_switch");
722                 fclose(f);
723                 return 1;
724         }
725
726         fclose(f);
727         return 0;
728 }
729
730 static int keep_running(struct thread_data *td)
731 {
732         unsigned long long io_done;
733
734         if (td->done)
735                 return 0;
736         if (td->o.time_based)
737                 return 1;
738         if (td->o.loops) {
739                 td->o.loops--;
740                 return 1;
741         }
742
743         io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE] + td->io_skip_bytes;
744         if (io_done < td->o.size)
745                 return 1;
746
747         return 0;
748 }
749
750 static int clear_io_state(struct thread_data *td)
751 {
752         struct fio_file *f;
753         unsigned int i;
754         int ret;
755
756         td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
757         td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
758         td->zone_bytes = 0;
759         td->rate_bytes = 0;
760         td->rate_blocks = 0;
761         td->rw_end_set[0] = td->rw_end_set[1] = 0;
762
763         td->last_was_sync = 0;
764
765         if (td->o.time_based)
766                 td->nr_done_files = 0;
767
768         for_each_file(td, f, i)
769                 td_io_close_file(td, f);
770
771         ret = 0;
772         for_each_file(td, f, i) {
773                 f->flags &= ~FIO_FILE_DONE;
774                 ret = td_io_open_file(td, f);
775                 if (ret)
776                         break;
777         }
778
779         return ret;
780 }
781
782 /*
783  * Entry point for the thread based jobs. The process based jobs end up
784  * here as well, after a little setup.
785  */
786 static void *thread_main(void *data)
787 {
788         unsigned long long runtime[2];
789         struct thread_data *td = data;
790         unsigned long elapsed;
791         int clear_state;
792
793         if (!td->o.use_thread)
794                 setsid();
795
796         td->pid = getpid();
797
798         INIT_LIST_HEAD(&td->io_u_freelist);
799         INIT_LIST_HEAD(&td->io_u_busylist);
800         INIT_LIST_HEAD(&td->io_u_requeues);
801         INIT_LIST_HEAD(&td->io_log_list);
802         INIT_LIST_HEAD(&td->io_hist_list);
803         td->io_hist_tree = RB_ROOT;
804
805         td_set_runstate(td, TD_INITIALIZED);
806         fio_sem_up(startup_sem);
807         fio_sem_down(td->mutex);
808
809         /*
810          * the ->mutex semaphore is now no longer used, close it to avoid
811          * eating a file descriptor
812          */
813         fio_sem_remove(td->mutex);
814
815         /*
816          * May alter parameters that init_io_u() will use, so we need to
817          * do this first.
818          */
819         if (init_iolog(td))
820                 goto err;
821
822         if (init_io_u(td))
823                 goto err;
824
825         if (td->o.cpumask_set && fio_setaffinity(td) == -1) {
826                 td_verror(td, errno, "cpu_set_affinity");
827                 goto err;
828         }
829
830         if (td->ioprio) {
831                 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
832                         td_verror(td, errno, "ioprio_set");
833                         goto err;
834                 }
835         }
836
837         if (nice(td->o.nice) == -1) {
838                 td_verror(td, errno, "nice");
839                 goto err;
840         }
841
842         if (td->o.ioscheduler && switch_ioscheduler(td))
843                 goto err;
844
845         if (!td->o.create_serialize && setup_files(td))
846                 goto err;
847
848         if (td_io_init(td))
849                 goto err;
850
851         if (open_files(td))
852                 goto err;
853
854         if (init_random_map(td))
855                 goto err;
856
857         if (td->o.exec_prerun) {
858                 if (system(td->o.exec_prerun) < 0)
859                         goto err;
860         }
861
862         fio_gettime(&td->epoch, NULL);
863         memcpy(&td->timeout_end, &td->epoch, sizeof(td->epoch));
864         getrusage(RUSAGE_SELF, &td->ts.ru_start);
865
866         runtime[0] = runtime[1] = 0;
867         clear_state = 0;
868         while (keep_running(td)) {
869                 fio_gettime(&td->start, NULL);
870                 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
871
872                 if (td->o.ratemin)
873                         memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
874
875                 if (clear_state && clear_io_state(td))
876                         break;
877
878                 prune_io_piece_log(td);
879
880                 do_io(td);
881
882                 clear_state = 1;
883
884                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
885                         if (td->rw_end_set[DDIR_READ])
886                                 elapsed = utime_since(&td->start, &td->rw_end[DDIR_READ]);
887                         else
888                                 elapsed = utime_since_now(&td->start);
889
890                         runtime[DDIR_READ] += elapsed;
891                 }
892                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
893                         if (td->rw_end_set[DDIR_WRITE])
894                                 elapsed = utime_since(&td->start, &td->rw_end[DDIR_WRITE]);
895                         else
896                                 elapsed = utime_since_now(&td->start);
897
898                         runtime[DDIR_WRITE] += elapsed;
899                 }
900                 
901                 if (td->error || td->terminate)
902                         break;
903
904                 if (td->o.verify == VERIFY_NONE ||
905                     (td->io_ops->flags & FIO_UNIDIR))
906                         continue;
907
908                 if (clear_io_state(td))
909                         break;
910
911                 fio_gettime(&td->start, NULL);
912
913                 do_verify(td);
914
915                 runtime[DDIR_READ] += utime_since_now(&td->start);
916
917                 if (td->error || td->terminate)
918                         break;
919         }
920
921         update_rusage_stat(td);
922         td->ts.runtime[0] = (runtime[0] + 999) / 1000;
923         td->ts.runtime[1] = (runtime[1] + 999) / 1000;
924         td->ts.total_run_time = mtime_since_now(&td->epoch);
925         td->ts.io_bytes[0] = td->io_bytes[0];
926         td->ts.io_bytes[1] = td->io_bytes[1];
927
928         if (td->ts.bw_log)
929                 finish_log(td, td->ts.bw_log, "bw");
930         if (td->ts.slat_log)
931                 finish_log(td, td->ts.slat_log, "slat");
932         if (td->ts.clat_log)
933                 finish_log(td, td->ts.clat_log, "clat");
934         if (td->o.exec_postrun) {
935                 if (system(td->o.exec_postrun) < 0)
936                         log_err("fio: postrun %s failed\n", td->o.exec_postrun);
937         }
938
939         if (exitall_on_terminate)
940                 terminate_threads(td->groupid);
941
942 err:
943         if (td->error)
944                 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
945         close_files(td);
946         close_ioengine(td);
947         cleanup_io_u(td);
948
949         /*
950          * do this very late, it will log file closing as well
951          */
952         if (td->o.write_iolog_file)
953                 write_iolog_close(td);
954
955         options_mem_free(td);
956         td_set_runstate(td, TD_EXITED);
957         return (void *) (unsigned long) td->error;
958 }
959
960 /*
961  * We cannot pass the td data into a forked process, so attach the td and
962  * pass it to the thread worker.
963  */
964 static int fork_main(int shmid, int offset)
965 {
966         struct thread_data *td;
967         void *data, *ret;
968
969         data = shmat(shmid, NULL, 0);
970         if (data == (void *) -1) {
971                 int __err = errno;
972
973                 perror("shmat");
974                 return __err;
975         }
976
977         td = data + offset * sizeof(struct thread_data);
978         ret = thread_main(td);
979         shmdt(data);
980         return (int) (unsigned long) ret;
981 }
982
983 /*
984  * Run over the job map and reap the threads that have exited, if any.
985  */
986 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
987 {
988         struct thread_data *td;
989         int i, cputhreads, pending, status, ret;
990
991         /*
992          * reap exited threads (TD_EXITED -> TD_REAPED)
993          */
994         pending = cputhreads = 0;
995         for_each_td(td, i) {
996                 int flags = 0;
997
998                 /*
999                  * ->io_ops is NULL for a thread that has closed its
1000                  * io engine
1001                  */
1002                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1003                         cputhreads++;
1004
1005                 if (!td->pid || td->runstate == TD_REAPED)
1006                         continue;
1007                 if (td->o.use_thread) {
1008                         if (td->runstate == TD_EXITED) {
1009                                 td_set_runstate(td, TD_REAPED);
1010                                 goto reaped;
1011                         }
1012                         continue;
1013                 }
1014
1015                 flags = WNOHANG;
1016                 if (td->runstate == TD_EXITED)
1017                         flags = 0;
1018
1019                 /*
1020                  * check if someone quit or got killed in an unusual way
1021                  */
1022                 ret = waitpid(td->pid, &status, flags);
1023                 if (ret < 0) {
1024                         if (errno == ECHILD) {
1025                                 log_err("fio: pid=%d disappeared %d\n", td->pid, td->runstate);
1026                                 td_set_runstate(td, TD_REAPED);
1027                                 goto reaped;
1028                         }
1029                         perror("waitpid");
1030                 } else if (ret == td->pid) {
1031                         if (WIFSIGNALED(status)) {
1032                                 int sig = WTERMSIG(status);
1033
1034                                 if (sig != SIGQUIT)
1035                                         log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
1036                                 td_set_runstate(td, TD_REAPED);
1037                                 goto reaped;
1038                         }
1039                         if (WIFEXITED(status)) {
1040                                 if (WEXITSTATUS(status) && !td->error)
1041                                         td->error = WEXITSTATUS(status);
1042
1043                                 td_set_runstate(td, TD_REAPED);
1044                                 goto reaped;
1045                         }
1046                 }
1047
1048                 /*
1049                  * thread is not dead, continue
1050                  */
1051                 pending++;
1052                 continue;
1053 reaped:
1054                 if (td->o.use_thread) {
1055                         long ret;
1056
1057                         if (pthread_join(td->thread, (void *) &ret))
1058                                 perror("pthread_join");
1059                 }
1060
1061                 (*nr_running)--;
1062                 (*m_rate) -= td->o.ratemin;
1063                 (*t_rate) -= td->o.rate;
1064                 pending--;
1065
1066                 if (td->error)
1067                         exit_value++;
1068         }
1069
1070         if (*nr_running == cputhreads && !pending)
1071                 terminate_threads(TERMINATE_ALL);
1072 }
1073
1074 /*
1075  * Main function for kicking off and reaping jobs, as needed.
1076  */
1077 static void run_threads(void)
1078 {
1079         struct thread_data *td;
1080         unsigned long spent;
1081         int i, todo, nr_running, m_rate, t_rate, nr_started;
1082
1083         if (fio_pin_memory())
1084                 return;
1085
1086         if (!terse_output) {
1087                 printf("Starting ");
1088                 if (nr_thread)
1089                         printf("%d thread%s", nr_thread, nr_thread > 1 ? "s" : "");
1090                 if (nr_process) {
1091                         if (nr_thread)
1092                                 printf(" and ");
1093                         printf("%d process%s", nr_process, nr_process > 1 ? "es" : "");
1094                 }
1095                 printf("\n");
1096                 fflush(stdout);
1097         }
1098
1099         signal(SIGINT, sig_handler);
1100         signal(SIGALRM, sig_handler);
1101
1102         todo = thread_number;
1103         nr_running = 0;
1104         nr_started = 0;
1105         m_rate = t_rate = 0;
1106
1107         for_each_td(td, i) {
1108                 print_status_init(td->thread_number - 1);
1109
1110                 if (!td->o.create_serialize) {
1111                         init_disk_util(td);
1112                         continue;
1113                 }
1114
1115                 /*
1116                  * do file setup here so it happens sequentially,
1117                  * we don't want X number of threads getting their
1118                  * client data interspersed on disk
1119                  */
1120                 if (setup_files(td)) {
1121                         exit_value++;
1122                         if (td->error)
1123                                 log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
1124                         td_set_runstate(td, TD_REAPED);
1125                         todo--;
1126                 }
1127
1128                 init_disk_util(td);
1129         }
1130
1131         set_genesis_time();
1132
1133         while (todo) {
1134                 struct thread_data *map[MAX_JOBS];
1135                 struct timeval this_start;
1136                 int this_jobs = 0, left;
1137
1138                 /*
1139                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1140                  */
1141                 for_each_td(td, i) {
1142                         if (td->runstate != TD_NOT_CREATED)
1143                                 continue;
1144
1145                         /*
1146                          * never got a chance to start, killed by other
1147                          * thread for some reason
1148                          */
1149                         if (td->terminate) {
1150                                 todo--;
1151                                 continue;
1152                         }
1153
1154                         if (td->o.start_delay) {
1155                                 spent = mtime_since_genesis();
1156
1157                                 if (td->o.start_delay * 1000 > spent)
1158                                         continue;
1159                         }
1160
1161                         if (td->o.stonewall && (nr_started || nr_running))
1162                                 break;
1163
1164                         /*
1165                          * Set state to created. Thread will transition
1166                          * to TD_INITIALIZED when it's done setting up.
1167                          */
1168                         td_set_runstate(td, TD_CREATED);
1169                         map[this_jobs++] = td;
1170                         nr_started++;
1171
1172                         if (td->o.use_thread) {
1173                                 if (pthread_create(&td->thread, NULL, thread_main, td)) {
1174                                         perror("thread_create");
1175                                         nr_started--;
1176                                         break;
1177                                 }
1178                         } else {
1179                                 if (!fork()) {
1180                                         int ret = fork_main(shm_id, i);
1181
1182                                         exit(ret);
1183                                 }
1184                         }
1185                         fio_sem_down(startup_sem);
1186                 }
1187
1188                 /*
1189                  * Wait for the started threads to transition to
1190                  * TD_INITIALIZED.
1191                  */
1192                 fio_gettime(&this_start, NULL);
1193                 left = this_jobs;
1194                 while (left && !fio_abort) {
1195                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1196                                 break;
1197
1198                         usleep(100000);
1199
1200                         for (i = 0; i < this_jobs; i++) {
1201                                 td = map[i];
1202                                 if (!td)
1203                                         continue;
1204                                 if (td->runstate == TD_INITIALIZED) {
1205                                         map[i] = NULL;
1206                                         left--;
1207                                 } else if (td->runstate >= TD_EXITED) {
1208                                         map[i] = NULL;
1209                                         left--;
1210                                         todo--;
1211                                         nr_running++; /* work-around... */
1212                                 }
1213                         }
1214                 }
1215
1216                 if (left) {
1217                         log_err("fio: %d jobs failed to start\n", left);
1218                         for (i = 0; i < this_jobs; i++) {
1219                                 td = map[i];
1220                                 if (!td)
1221                                         continue;
1222                                 kill(td->pid, SIGTERM);
1223                         }
1224                         break;
1225                 }
1226
1227                 /*
1228                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1229                  */
1230                 for_each_td(td, i) {
1231                         if (td->runstate != TD_INITIALIZED)
1232                                 continue;
1233
1234                         td_set_runstate(td, TD_RUNNING);
1235                         nr_running++;
1236                         nr_started--;
1237                         m_rate += td->o.ratemin;
1238                         t_rate += td->o.rate;
1239                         todo--;
1240                         fio_sem_up(td->mutex);
1241                 }
1242
1243                 reap_threads(&nr_running, &t_rate, &m_rate);
1244
1245                 if (todo)
1246                         usleep(100000);
1247         }
1248
1249         while (nr_running) {
1250                 reap_threads(&nr_running, &t_rate, &m_rate);
1251                 usleep(10000);
1252         }
1253
1254         update_io_ticks();
1255         fio_unpin_memory();
1256 }
1257
1258 int main(int argc, char *argv[])
1259 {
1260         long ps;
1261
1262         /*
1263          * We need locale for number printing, if it isn't set then just
1264          * go with the US format.
1265          */
1266         if (!getenv("LC_NUMERIC"))
1267                 setlocale(LC_NUMERIC, "en_US");
1268
1269         if (parse_options(argc, argv))
1270                 return 1;
1271
1272         if (!thread_number)
1273                 return 0;
1274
1275         ps = sysconf(_SC_PAGESIZE);
1276         if (ps < 0) {
1277                 log_err("Failed to get page size\n");
1278                 return 1;
1279         }
1280
1281         page_size = ps;
1282         page_mask = ps - 1;
1283
1284         if (write_bw_log) {
1285                 setup_log(&agg_io_log[DDIR_READ]);
1286                 setup_log(&agg_io_log[DDIR_WRITE]);
1287         }
1288
1289         startup_sem = fio_sem_init(0);
1290
1291         set_genesis_time();
1292
1293         disk_util_timer_arm();
1294
1295         run_threads();
1296
1297         if (!fio_abort) {
1298                 show_run_stats();
1299                 if (write_bw_log) {
1300                         __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1301                         __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");
1302                 }
1303         }
1304
1305         fio_sem_remove(startup_sem);
1306         return exit_value;
1307 }