6f40038799831a4f521af7ac25942b660e56978b
[fio.git] / backend.c
1 /*
2  * fio - the flexible io tester
3  *
4  * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
5  * Copyright (C) 2006-2012 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 <limits.h>
28 #include <signal.h>
29 #include <time.h>
30 #include <locale.h>
31 #include <assert.h>
32 #include <time.h>
33 #include <inttypes.h>
34 #include <sys/stat.h>
35 #include <sys/wait.h>
36 #include <sys/ipc.h>
37 #include <sys/shm.h>
38 #include <sys/mman.h>
39
40 #include "fio.h"
41 #include "hash.h"
42 #include "smalloc.h"
43 #include "verify.h"
44 #include "trim.h"
45 #include "diskutil.h"
46 #include "cgroup.h"
47 #include "profile.h"
48 #include "lib/rand.h"
49 #include "memalign.h"
50 #include "server.h"
51
52 static pthread_t disk_util_thread;
53 static struct fio_mutex *disk_thread_mutex;
54 static struct fio_mutex *startup_mutex;
55 static struct fio_mutex *writeout_mutex;
56 static struct flist_head *cgroup_list;
57 static char *cgroup_mnt;
58 static int exit_value;
59 static volatile int fio_abort;
60 static unsigned int nr_process = 0;
61 static unsigned int nr_thread = 0;
62
63 struct io_log *agg_io_log[DDIR_RWDIR_CNT];
64
65 int groupid = 0;
66 unsigned int thread_number = 0;
67 int shm_id = 0;
68 int temp_stall_ts;
69 unsigned long done_secs = 0;
70
71 #define PAGE_ALIGN(buf) \
72         (char *) (((uintptr_t) (buf) + page_mask) & ~page_mask)
73
74 #define JOB_START_TIMEOUT       (5 * 1000)
75
76 static void sig_int(int sig)
77 {
78         if (threads) {
79                 if (is_backend)
80                         fio_server_got_signal(sig);
81                 else {
82                         log_info("\nfio: terminating on signal %d\n", sig);
83                         fflush(stdout);
84                         exit_value = 128;
85                 }
86
87                 fio_terminate_threads(TERMINATE_ALL);
88         }
89 }
90
91 static void sig_show_status(int sig)
92 {
93         show_running_run_stats();
94 }
95
96 static void set_sig_handlers(void)
97 {
98         struct sigaction act;
99
100         memset(&act, 0, sizeof(act));
101         act.sa_handler = sig_int;
102         act.sa_flags = SA_RESTART;
103         sigaction(SIGINT, &act, NULL);
104
105         memset(&act, 0, sizeof(act));
106         act.sa_handler = sig_int;
107         act.sa_flags = SA_RESTART;
108         sigaction(SIGTERM, &act, NULL);
109
110         memset(&act, 0, sizeof(act));
111         act.sa_handler = sig_show_status;
112         act.sa_flags = SA_RESTART;
113         sigaction(SIGUSR1, &act, NULL);
114
115         if (is_backend) {
116                 memset(&act, 0, sizeof(act));
117                 act.sa_handler = sig_int;
118                 act.sa_flags = SA_RESTART;
119                 sigaction(SIGPIPE, &act, NULL);
120         }
121 }
122
123 /*
124  * Check if we are above the minimum rate given.
125  */
126 static int __check_min_rate(struct thread_data *td, struct timeval *now,
127                             enum fio_ddir ddir)
128 {
129         unsigned long long bytes = 0;
130         unsigned long iops = 0;
131         unsigned long spent;
132         unsigned long rate;
133         unsigned int ratemin = 0;
134         unsigned int rate_iops = 0;
135         unsigned int rate_iops_min = 0;
136
137         assert(ddir_rw(ddir));
138
139         if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
140                 return 0;
141
142         /*
143          * allow a 2 second settle period in the beginning
144          */
145         if (mtime_since(&td->start, now) < 2000)
146                 return 0;
147
148         iops += td->this_io_blocks[ddir];
149         bytes += td->this_io_bytes[ddir];
150         ratemin += td->o.ratemin[ddir];
151         rate_iops += td->o.rate_iops[ddir];
152         rate_iops_min += td->o.rate_iops_min[ddir];
153
154         /*
155          * if rate blocks is set, sample is running
156          */
157         if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
158                 spent = mtime_since(&td->lastrate[ddir], now);
159                 if (spent < td->o.ratecycle)
160                         return 0;
161
162                 if (td->o.rate[ddir]) {
163                         /*
164                          * check bandwidth specified rate
165                          */
166                         if (bytes < td->rate_bytes[ddir]) {
167                                 log_err("%s: min rate %u not met\n", td->o.name,
168                                                                 ratemin);
169                                 return 1;
170                         } else {
171                                 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
172                                 if (rate < ratemin ||
173                                     bytes < td->rate_bytes[ddir]) {
174                                         log_err("%s: min rate %u not met, got"
175                                                 " %luKB/sec\n", td->o.name,
176                                                         ratemin, rate);
177                                         return 1;
178                                 }
179                         }
180                 } else {
181                         /*
182                          * checks iops specified rate
183                          */
184                         if (iops < rate_iops) {
185                                 log_err("%s: min iops rate %u not met\n",
186                                                 td->o.name, rate_iops);
187                                 return 1;
188                         } else {
189                                 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
190                                 if (rate < rate_iops_min ||
191                                     iops < td->rate_blocks[ddir]) {
192                                         log_err("%s: min iops rate %u not met,"
193                                                 " got %lu\n", td->o.name,
194                                                         rate_iops_min, rate);
195                                 }
196                         }
197                 }
198         }
199
200         td->rate_bytes[ddir] = bytes;
201         td->rate_blocks[ddir] = iops;
202         memcpy(&td->lastrate[ddir], now, sizeof(*now));
203         return 0;
204 }
205
206 static int check_min_rate(struct thread_data *td, struct timeval *now,
207                           unsigned long *bytes_done)
208 {
209         int ret = 0;
210
211         if (bytes_done[DDIR_READ])
212                 ret |= __check_min_rate(td, now, DDIR_READ);
213         if (bytes_done[DDIR_WRITE])
214                 ret |= __check_min_rate(td, now, DDIR_WRITE);
215         if (bytes_done[DDIR_TRIM])
216                 ret |= __check_min_rate(td, now, DDIR_TRIM);
217
218         return ret;
219 }
220
221 /*
222  * When job exits, we can cancel the in-flight IO if we are using async
223  * io. Attempt to do so.
224  */
225 static void cleanup_pending_aio(struct thread_data *td)
226 {
227         struct flist_head *entry, *n;
228         struct io_u *io_u;
229         int r;
230
231         /*
232          * get immediately available events, if any
233          */
234         r = io_u_queued_complete(td, 0, NULL);
235         if (r < 0)
236                 return;
237
238         /*
239          * now cancel remaining active events
240          */
241         if (td->io_ops->cancel) {
242                 flist_for_each_safe(entry, n, &td->io_u_busylist) {
243                         io_u = flist_entry(entry, struct io_u, list);
244
245                         /*
246                          * if the io_u isn't in flight, then that generally
247                          * means someone leaked an io_u. complain but fix
248                          * it up, so we don't stall here.
249                          */
250                         if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
251                                 log_err("fio: non-busy IO on busy list\n");
252                                 put_io_u(td, io_u);
253                         } else {
254                                 r = td->io_ops->cancel(td, io_u);
255                                 if (!r)
256                                         put_io_u(td, io_u);
257                         }
258                 }
259         }
260
261         if (td->cur_depth)
262                 r = io_u_queued_complete(td, td->cur_depth, NULL);
263 }
264
265 /*
266  * Helper to handle the final sync of a file. Works just like the normal
267  * io path, just does everything sync.
268  */
269 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
270 {
271         struct io_u *io_u = __get_io_u(td);
272         int ret;
273
274         if (!io_u)
275                 return 1;
276
277         io_u->ddir = DDIR_SYNC;
278         io_u->file = f;
279
280         if (td_io_prep(td, io_u)) {
281                 put_io_u(td, io_u);
282                 return 1;
283         }
284
285 requeue:
286         ret = td_io_queue(td, io_u);
287         if (ret < 0) {
288                 td_verror(td, io_u->error, "td_io_queue");
289                 put_io_u(td, io_u);
290                 return 1;
291         } else if (ret == FIO_Q_QUEUED) {
292                 if (io_u_queued_complete(td, 1, NULL) < 0)
293                         return 1;
294         } else if (ret == FIO_Q_COMPLETED) {
295                 if (io_u->error) {
296                         td_verror(td, io_u->error, "td_io_queue");
297                         return 1;
298                 }
299
300                 if (io_u_sync_complete(td, io_u, NULL) < 0)
301                         return 1;
302         } else if (ret == FIO_Q_BUSY) {
303                 if (td_io_commit(td))
304                         return 1;
305                 goto requeue;
306         }
307
308         return 0;
309 }
310
311 static inline void __update_tv_cache(struct thread_data *td)
312 {
313         fio_gettime(&td->tv_cache, NULL);
314 }
315
316 static inline void update_tv_cache(struct thread_data *td)
317 {
318         if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
319                 __update_tv_cache(td);
320 }
321
322 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
323 {
324         if (in_ramp_time(td))
325                 return 0;
326         if (!td->o.timeout)
327                 return 0;
328         if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
329                 return 1;
330
331         return 0;
332 }
333
334 static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
335                                int *retptr)
336 {
337         int ret = *retptr;
338
339         if (ret < 0 || td->error) {
340                 int err = td->error;
341                 enum error_type_bit eb;
342
343                 if (ret < 0)
344                         err = -ret;
345
346                 eb = td_error_type(ddir, err);
347                 if (!(td->o.continue_on_error & (1 << eb)))
348                         return 1;
349
350                 if (td_non_fatal_error(td, eb, err)) {
351                         /*
352                          * Continue with the I/Os in case of
353                          * a non fatal error.
354                          */
355                         update_error_count(td, err);
356                         td_clear_error(td);
357                         *retptr = 0;
358                         return 0;
359                 } else if (td->o.fill_device && err == ENOSPC) {
360                         /*
361                          * We expect to hit this error if
362                          * fill_device option is set.
363                          */
364                         td_clear_error(td);
365                         td->terminate = 1;
366                         return 1;
367                 } else {
368                         /*
369                          * Stop the I/O in case of a fatal
370                          * error.
371                          */
372                         update_error_count(td, err);
373                         return 1;
374                 }
375         }
376
377         return 0;
378 }
379
380 /*
381  * The main verify engine. Runs over the writes we previously submitted,
382  * reads the blocks back in, and checks the crc/md5 of the data.
383  */
384 static void do_verify(struct thread_data *td)
385 {
386         struct fio_file *f;
387         struct io_u *io_u;
388         int ret, min_events;
389         unsigned int i;
390
391         dprint(FD_VERIFY, "starting loop\n");
392
393         /*
394          * sync io first and invalidate cache, to make sure we really
395          * read from disk.
396          */
397         for_each_file(td, f, i) {
398                 if (!fio_file_open(f))
399                         continue;
400                 if (fio_io_sync(td, f))
401                         break;
402                 if (file_invalidate_cache(td, f))
403                         break;
404         }
405
406         if (td->error)
407                 return;
408
409         td_set_runstate(td, TD_VERIFYING);
410
411         io_u = NULL;
412         while (!td->terminate) {
413                 int ret2, full;
414
415                 update_tv_cache(td);
416
417                 if (runtime_exceeded(td, &td->tv_cache)) {
418                         __update_tv_cache(td);
419                         if (runtime_exceeded(td, &td->tv_cache)) {
420                                 td->terminate = 1;
421                                 break;
422                         }
423                 }
424
425                 if (flow_threshold_exceeded(td))
426                         continue;
427
428                 io_u = __get_io_u(td);
429                 if (!io_u)
430                         break;
431
432                 if (get_next_verify(td, io_u)) {
433                         put_io_u(td, io_u);
434                         break;
435                 }
436
437                 if (td_io_prep(td, io_u)) {
438                         put_io_u(td, io_u);
439                         break;
440                 }
441
442                 if (td->o.verify_async)
443                         io_u->end_io = verify_io_u_async;
444                 else
445                         io_u->end_io = verify_io_u;
446
447                 ret = td_io_queue(td, io_u);
448                 switch (ret) {
449                 case FIO_Q_COMPLETED:
450                         if (io_u->error) {
451                                 ret = -io_u->error;
452                                 clear_io_u(td, io_u);
453                         } else if (io_u->resid) {
454                                 int bytes = io_u->xfer_buflen - io_u->resid;
455
456                                 /*
457                                  * zero read, fail
458                                  */
459                                 if (!bytes) {
460                                         td_verror(td, EIO, "full resid");
461                                         put_io_u(td, io_u);
462                                         break;
463                                 }
464
465                                 io_u->xfer_buflen = io_u->resid;
466                                 io_u->xfer_buf += bytes;
467                                 io_u->offset += bytes;
468
469                                 if (ddir_rw(io_u->ddir))
470                                         td->ts.short_io_u[io_u->ddir]++;
471
472                                 f = io_u->file;
473                                 if (io_u->offset == f->real_file_size)
474                                         goto sync_done;
475
476                                 requeue_io_u(td, &io_u);
477                         } else {
478 sync_done:
479                                 ret = io_u_sync_complete(td, io_u, NULL);
480                                 if (ret < 0)
481                                         break;
482                         }
483                         continue;
484                 case FIO_Q_QUEUED:
485                         break;
486                 case FIO_Q_BUSY:
487                         requeue_io_u(td, &io_u);
488                         ret2 = td_io_commit(td);
489                         if (ret2 < 0)
490                                 ret = ret2;
491                         break;
492                 default:
493                         assert(ret < 0);
494                         td_verror(td, -ret, "td_io_queue");
495                         break;
496                 }
497
498                 if (break_on_this_error(td, io_u->ddir, &ret))
499                         break;
500
501                 /*
502                  * if we can queue more, do so. but check if there are
503                  * completed io_u's first. Note that we can get BUSY even
504                  * without IO queued, if the system is resource starved.
505                  */
506                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
507                 if (full || !td->o.iodepth_batch_complete) {
508                         min_events = min(td->o.iodepth_batch_complete,
509                                          td->cur_depth);
510                         /*
511                          * if the queue is full, we MUST reap at least 1 event
512                          */
513                         if (full && !min_events)
514                                 min_events = 1;
515
516                         do {
517                                 /*
518                                  * Reap required number of io units, if any,
519                                  * and do the verification on them through
520                                  * the callback handler
521                                  */
522                                 if (io_u_queued_complete(td, min_events, NULL) < 0) {
523                                         ret = -1;
524                                         break;
525                                 }
526                         } while (full && (td->cur_depth > td->o.iodepth_low));
527                 }
528                 if (ret < 0)
529                         break;
530         }
531
532         if (!td->error) {
533                 min_events = td->cur_depth;
534
535                 if (min_events)
536                         ret = io_u_queued_complete(td, min_events, NULL);
537         } else
538                 cleanup_pending_aio(td);
539
540         td_set_runstate(td, TD_RUNNING);
541
542         dprint(FD_VERIFY, "exiting loop\n");
543 }
544
545 static int io_bytes_exceeded(struct thread_data *td)
546 {
547         unsigned long long bytes;
548
549         if (td_rw(td))
550                 bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE];
551         else if (td_write(td))
552                 bytes = td->this_io_bytes[DDIR_WRITE];
553         else if (td_read(td))
554                 bytes = td->this_io_bytes[DDIR_READ];
555         else
556                 bytes = td->this_io_bytes[DDIR_TRIM];
557
558         return bytes >= td->o.size;
559 }
560
561 /*
562  * Main IO worker function. It retrieves io_u's to process and queues
563  * and reaps them, checking for rate and errors along the way.
564  */
565 static void do_io(struct thread_data *td)
566 {
567         unsigned int i;
568         int ret = 0;
569
570         if (in_ramp_time(td))
571                 td_set_runstate(td, TD_RAMP);
572         else
573                 td_set_runstate(td, TD_RUNNING);
574
575         while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
576                 (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td) ||
577                 td->o.time_based) {
578                 struct timeval comp_time;
579                 unsigned long bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
580                 int min_evts = 0;
581                 struct io_u *io_u;
582                 int ret2, full;
583                 enum fio_ddir ddir;
584
585                 if (td->terminate)
586                         break;
587
588                 update_tv_cache(td);
589
590                 if (runtime_exceeded(td, &td->tv_cache)) {
591                         __update_tv_cache(td);
592                         if (runtime_exceeded(td, &td->tv_cache)) {
593                                 td->terminate = 1;
594                                 break;
595                         }
596                 }
597
598                 if (flow_threshold_exceeded(td))
599                         continue;
600
601                 io_u = get_io_u(td);
602                 if (!io_u)
603                         break;
604
605                 ddir = io_u->ddir;
606
607                 /*
608                  * Add verification end_io handler if:
609                  *      - Asked to verify (!td_rw(td))
610                  *      - Or the io_u is from our verify list (mixed write/ver)
611                  */
612                 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
613                     ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
614                         if (td->o.verify_async)
615                                 io_u->end_io = verify_io_u_async;
616                         else
617                                 io_u->end_io = verify_io_u;
618                         td_set_runstate(td, TD_VERIFYING);
619                 } else if (in_ramp_time(td))
620                         td_set_runstate(td, TD_RAMP);
621                 else
622                         td_set_runstate(td, TD_RUNNING);
623
624                 ret = td_io_queue(td, io_u);
625                 switch (ret) {
626                 case FIO_Q_COMPLETED:
627                         if (io_u->error) {
628                                 ret = -io_u->error;
629                                 clear_io_u(td, io_u);
630                         } else if (io_u->resid) {
631                                 int bytes = io_u->xfer_buflen - io_u->resid;
632                                 struct fio_file *f = io_u->file;
633
634                                 /*
635                                  * zero read, fail
636                                  */
637                                 if (!bytes) {
638                                         td_verror(td, EIO, "full resid");
639                                         put_io_u(td, io_u);
640                                         break;
641                                 }
642
643                                 io_u->xfer_buflen = io_u->resid;
644                                 io_u->xfer_buf += bytes;
645                                 io_u->offset += bytes;
646
647                                 if (ddir_rw(io_u->ddir))
648                                         td->ts.short_io_u[io_u->ddir]++;
649
650                                 if (io_u->offset == f->real_file_size)
651                                         goto sync_done;
652
653                                 requeue_io_u(td, &io_u);
654                         } else {
655 sync_done:
656                                 if (__should_check_rate(td, DDIR_READ) ||
657                                     __should_check_rate(td, DDIR_WRITE) ||
658                                     __should_check_rate(td, DDIR_TRIM))
659                                         fio_gettime(&comp_time, NULL);
660
661                                 ret = io_u_sync_complete(td, io_u, bytes_done);
662                                 if (ret < 0)
663                                         break;
664                         }
665                         break;
666                 case FIO_Q_QUEUED:
667                         /*
668                          * if the engine doesn't have a commit hook,
669                          * the io_u is really queued. if it does have such
670                          * a hook, it has to call io_u_queued() itself.
671                          */
672                         if (td->io_ops->commit == NULL)
673                                 io_u_queued(td, io_u);
674                         break;
675                 case FIO_Q_BUSY:
676                         requeue_io_u(td, &io_u);
677                         ret2 = td_io_commit(td);
678                         if (ret2 < 0)
679                                 ret = ret2;
680                         break;
681                 default:
682                         assert(ret < 0);
683                         put_io_u(td, io_u);
684                         break;
685                 }
686
687                 if (break_on_this_error(td, ddir, &ret))
688                         break;
689
690                 /*
691                  * See if we need to complete some commands. Note that we
692                  * can get BUSY even without IO queued, if the system is
693                  * resource starved.
694                  */
695                 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
696                 if (full || !td->o.iodepth_batch_complete) {
697                         min_evts = min(td->o.iodepth_batch_complete,
698                                         td->cur_depth);
699                         /*
700                          * if the queue is full, we MUST reap at least 1 event
701                          */
702                         if (full && !min_evts)
703                                 min_evts = 1;
704
705                         if (__should_check_rate(td, DDIR_READ) ||
706                             __should_check_rate(td, DDIR_WRITE) ||
707                             __should_check_rate(td, DDIR_TRIM))
708                                 fio_gettime(&comp_time, NULL);
709
710                         do {
711                                 ret = io_u_queued_complete(td, min_evts, bytes_done);
712                                 if (ret < 0)
713                                         break;
714
715                         } while (full && (td->cur_depth > td->o.iodepth_low));
716                 }
717
718                 if (ret < 0)
719                         break;
720                 if (!ddir_rw_sum(bytes_done))
721                         continue;
722
723                 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
724                         if (check_min_rate(td, &comp_time, bytes_done)) {
725                                 if (exitall_on_terminate)
726                                         fio_terminate_threads(td->groupid);
727                                 td_verror(td, EIO, "check_min_rate");
728                                 break;
729                         }
730                 }
731
732                 if (td->o.thinktime) {
733                         unsigned long long b;
734
735                         b = ddir_rw_sum(td->io_blocks);
736                         if (!(b % td->o.thinktime_blocks)) {
737                                 int left;
738
739                                 if (td->o.thinktime_spin)
740                                         usec_spin(td->o.thinktime_spin);
741
742                                 left = td->o.thinktime - td->o.thinktime_spin;
743                                 if (left)
744                                         usec_sleep(td, left);
745                         }
746                 }
747         }
748
749         if (td->trim_entries)
750                 log_err("fio: %d trim entries leaked?\n", td->trim_entries);
751
752         if (td->o.fill_device && td->error == ENOSPC) {
753                 td->error = 0;
754                 td->terminate = 1;
755         }
756         if (!td->error) {
757                 struct fio_file *f;
758
759                 i = td->cur_depth;
760                 if (i) {
761                         ret = io_u_queued_complete(td, i, NULL);
762                         if (td->o.fill_device && td->error == ENOSPC)
763                                 td->error = 0;
764                 }
765
766                 if (should_fsync(td) && td->o.end_fsync) {
767                         td_set_runstate(td, TD_FSYNCING);
768
769                         for_each_file(td, f, i) {
770                                 if (!fio_file_open(f))
771                                         continue;
772                                 fio_io_sync(td, f);
773                         }
774                 }
775         } else
776                 cleanup_pending_aio(td);
777
778         /*
779          * stop job if we failed doing any IO
780          */
781         if (!ddir_rw_sum(td->this_io_bytes))
782                 td->done = 1;
783 }
784
785 static void cleanup_io_u(struct thread_data *td)
786 {
787         struct flist_head *entry, *n;
788         struct io_u *io_u;
789
790         flist_for_each_safe(entry, n, &td->io_u_freelist) {
791                 io_u = flist_entry(entry, struct io_u, list);
792
793                 flist_del(&io_u->list);
794                 fio_memfree(io_u, sizeof(*io_u));
795         }
796
797         free_io_mem(td);
798 }
799
800 static int init_io_u(struct thread_data *td)
801 {
802         struct io_u *io_u;
803         unsigned int max_bs, min_write;
804         int cl_align, i, max_units;
805         int data_xfer = 1;
806         char *p;
807
808         max_units = td->o.iodepth;
809         max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
810         max_bs = max(td->o.max_bs[DDIR_TRIM], max_bs);
811         min_write = td->o.min_bs[DDIR_WRITE];
812         td->orig_buffer_size = (unsigned long long) max_bs
813                                         * (unsigned long long) max_units;
814
815         if ((td->io_ops->flags & FIO_NOIO) || !td_rw(td))
816                 data_xfer = 0;
817
818         if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
819                 unsigned long bs;
820
821                 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
822                 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
823         }
824
825         if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
826                 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
827                 return 1;
828         }
829
830         if (data_xfer && allocate_io_mem(td))
831                 return 1;
832
833         if (td->o.odirect || td->o.mem_align ||
834             (td->io_ops->flags & FIO_RAWIO))
835                 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
836         else
837                 p = td->orig_buffer;
838
839         cl_align = os_cache_line_size();
840
841         for (i = 0; i < max_units; i++) {
842                 void *ptr;
843
844                 if (td->terminate)
845                         return 1;
846
847                 ptr = fio_memalign(cl_align, sizeof(*io_u));
848                 if (!ptr) {
849                         log_err("fio: unable to allocate aligned memory\n");
850                         break;
851                 }
852
853                 io_u = ptr;
854                 memset(io_u, 0, sizeof(*io_u));
855                 INIT_FLIST_HEAD(&io_u->list);
856                 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
857
858                 if (data_xfer) {
859                         io_u->buf = p;
860                         dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
861
862                         if (td_write(td))
863                                 io_u_fill_buffer(td, io_u, min_write, max_bs);
864                         if (td_write(td) && td->o.verify_pattern_bytes) {
865                                 /*
866                                  * Fill the buffer with the pattern if we are
867                                  * going to be doing writes.
868                                  */
869                                 fill_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
870                         }
871                 }
872
873                 io_u->index = i;
874                 io_u->flags = IO_U_F_FREE;
875                 flist_add(&io_u->list, &td->io_u_freelist);
876                 p += max_bs;
877         }
878
879         return 0;
880 }
881
882 static int switch_ioscheduler(struct thread_data *td)
883 {
884         char tmp[256], tmp2[128];
885         FILE *f;
886         int ret;
887
888         if (td->io_ops->flags & FIO_DISKLESSIO)
889                 return 0;
890
891         sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
892
893         f = fopen(tmp, "r+");
894         if (!f) {
895                 if (errno == ENOENT) {
896                         log_err("fio: os or kernel doesn't support IO scheduler"
897                                 " switching\n");
898                         return 0;
899                 }
900                 td_verror(td, errno, "fopen iosched");
901                 return 1;
902         }
903
904         /*
905          * Set io scheduler.
906          */
907         ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
908         if (ferror(f) || ret != 1) {
909                 td_verror(td, errno, "fwrite");
910                 fclose(f);
911                 return 1;
912         }
913
914         rewind(f);
915
916         /*
917          * Read back and check that the selected scheduler is now the default.
918          */
919         ret = fread(tmp, 1, sizeof(tmp), f);
920         if (ferror(f) || ret < 0) {
921                 td_verror(td, errno, "fread");
922                 fclose(f);
923                 return 1;
924         }
925
926         sprintf(tmp2, "[%s]", td->o.ioscheduler);
927         if (!strstr(tmp, tmp2)) {
928                 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
929                 td_verror(td, EINVAL, "iosched_switch");
930                 fclose(f);
931                 return 1;
932         }
933
934         fclose(f);
935         return 0;
936 }
937
938 static int keep_running(struct thread_data *td)
939 {
940         if (td->done)
941                 return 0;
942         if (td->o.time_based)
943                 return 1;
944         if (td->o.loops) {
945                 td->o.loops--;
946                 return 1;
947         }
948
949         if (ddir_rw_sum(td->io_bytes) < td->o.size)
950                 return 1;
951
952         return 0;
953 }
954
955 static int exec_string(const char *string)
956 {
957         int ret, newlen = strlen(string) + 1 + 8;
958         char *str;
959
960         str = malloc(newlen);
961         sprintf(str, "sh -c %s", string);
962
963         ret = system(str);
964         if (ret == -1)
965                 log_err("fio: exec of cmd <%s> failed\n", str);
966
967         free(str);
968         return ret;
969 }
970
971 /*
972  * Entry point for the thread based jobs. The process based jobs end up
973  * here as well, after a little setup.
974  */
975 static void *thread_main(void *data)
976 {
977         unsigned long long elapsed;
978         struct thread_data *td = data;
979         struct thread_options *o = &td->o;
980         pthread_condattr_t attr;
981         int clear_state;
982         int ret;
983
984         if (!o->use_thread) {
985                 setsid();
986                 td->pid = getpid();
987         } else
988                 td->pid = gettid();
989
990         dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
991
992         if (is_backend)
993                 fio_server_send_start(td);
994
995         INIT_FLIST_HEAD(&td->io_u_freelist);
996         INIT_FLIST_HEAD(&td->io_u_busylist);
997         INIT_FLIST_HEAD(&td->io_u_requeues);
998         INIT_FLIST_HEAD(&td->io_log_list);
999         INIT_FLIST_HEAD(&td->io_hist_list);
1000         INIT_FLIST_HEAD(&td->verify_list);
1001         INIT_FLIST_HEAD(&td->trim_list);
1002         pthread_mutex_init(&td->io_u_lock, NULL);
1003         td->io_hist_tree = RB_ROOT;
1004
1005         pthread_condattr_init(&attr);
1006         pthread_cond_init(&td->verify_cond, &attr);
1007         pthread_cond_init(&td->free_cond, &attr);
1008
1009         td_set_runstate(td, TD_INITIALIZED);
1010         dprint(FD_MUTEX, "up startup_mutex\n");
1011         fio_mutex_up(startup_mutex);
1012         dprint(FD_MUTEX, "wait on td->mutex\n");
1013         fio_mutex_down(td->mutex);
1014         dprint(FD_MUTEX, "done waiting on td->mutex\n");
1015
1016         /*
1017          * the ->mutex mutex is now no longer used, close it to avoid
1018          * eating a file descriptor
1019          */
1020         fio_mutex_remove(td->mutex);
1021         td->mutex = NULL;
1022
1023         /*
1024          * A new gid requires privilege, so we need to do this before setting
1025          * the uid.
1026          */
1027         if (o->gid != -1U && setgid(o->gid)) {
1028                 td_verror(td, errno, "setgid");
1029                 goto err;
1030         }
1031         if (o->uid != -1U && setuid(o->uid)) {
1032                 td_verror(td, errno, "setuid");
1033                 goto err;
1034         }
1035
1036         /*
1037          * If we have a gettimeofday() thread, make sure we exclude that
1038          * thread from this job
1039          */
1040         if (o->gtod_cpu)
1041                 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1042
1043         /*
1044          * Set affinity first, in case it has an impact on the memory
1045          * allocations.
1046          */
1047         if (o->cpumask_set) {
1048                 ret = fio_setaffinity(td->pid, o->cpumask);
1049                 if (ret == -1) {
1050                         td_verror(td, errno, "cpu_set_affinity");
1051                         goto err;
1052                 }
1053         }
1054
1055         if (fio_pin_memory(td))
1056                 goto err;
1057
1058         /*
1059          * May alter parameters that init_io_u() will use, so we need to
1060          * do this first.
1061          */
1062         if (init_iolog(td))
1063                 goto err;
1064
1065         if (init_io_u(td))
1066                 goto err;
1067
1068         if (o->verify_async && verify_async_init(td))
1069                 goto err;
1070
1071         if (o->ioprio) {
1072                 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1073                 if (ret == -1) {
1074                         td_verror(td, errno, "ioprio_set");
1075                         goto err;
1076                 }
1077         }
1078
1079         if (td->o.cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1080                 goto err;
1081
1082         errno = 0;
1083         if (nice(o->nice) == -1 && errno != 0) {
1084                 td_verror(td, errno, "nice");
1085                 goto err;
1086         }
1087
1088         if (o->ioscheduler && switch_ioscheduler(td))
1089                 goto err;
1090
1091         if (!o->create_serialize && setup_files(td))
1092                 goto err;
1093
1094         if (td_io_init(td))
1095                 goto err;
1096
1097         if (init_random_map(td))
1098                 goto err;
1099
1100         if (o->exec_prerun && exec_string(o->exec_prerun))
1101                 goto err;
1102
1103         if (o->pre_read) {
1104                 if (pre_read_files(td) < 0)
1105                         goto err;
1106         }
1107
1108         fio_verify_init(td);
1109
1110         fio_gettime(&td->epoch, NULL);
1111         getrusage(RUSAGE_SELF, &td->ru_start);
1112
1113         clear_state = 0;
1114         while (keep_running(td)) {
1115                 fio_gettime(&td->start, NULL);
1116                 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1117                 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1118                 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1119
1120                 if (td->o.ratemin[DDIR_READ] || td->o.ratemin[DDIR_WRITE] ||
1121                                 td->o.ratemin[DDIR_TRIM]) {
1122                         memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1123                                                 sizeof(td->bw_sample_time));
1124                         memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1125                                                 sizeof(td->bw_sample_time));
1126                         memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1127                                                 sizeof(td->bw_sample_time));
1128                 }
1129
1130                 if (clear_state)
1131                         clear_io_state(td);
1132
1133                 prune_io_piece_log(td);
1134
1135                 do_io(td);
1136
1137                 clear_state = 1;
1138
1139                 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1140                         elapsed = utime_since_now(&td->start);
1141                         td->ts.runtime[DDIR_READ] += elapsed;
1142                 }
1143                 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1144                         elapsed = utime_since_now(&td->start);
1145                         td->ts.runtime[DDIR_WRITE] += elapsed;
1146                 }
1147                 if (td_trim(td) && td->io_bytes[DDIR_TRIM]) {
1148                         elapsed = utime_since_now(&td->start);
1149                         td->ts.runtime[DDIR_TRIM] += elapsed;
1150                 }
1151
1152                 if (td->error || td->terminate)
1153                         break;
1154
1155                 if (!td->o.do_verify ||
1156                     td->o.verify == VERIFY_NONE ||
1157                     (td->io_ops->flags & FIO_UNIDIR))
1158                         continue;
1159
1160                 clear_io_state(td);
1161
1162                 fio_gettime(&td->start, NULL);
1163
1164                 do_verify(td);
1165
1166                 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1167
1168                 if (td->error || td->terminate)
1169                         break;
1170         }
1171
1172         update_rusage_stat(td);
1173         td->ts.runtime[DDIR_READ] = (td->ts.runtime[DDIR_READ] + 999) / 1000;
1174         td->ts.runtime[DDIR_WRITE] = (td->ts.runtime[DDIR_WRITE] + 999) / 1000;
1175         td->ts.runtime[DDIR_TRIM] = (td->ts.runtime[DDIR_TRIM] + 999) / 1000;
1176         td->ts.total_run_time = mtime_since_now(&td->epoch);
1177         td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1178         td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1179         td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1180
1181         fio_unpin_memory(td);
1182
1183         fio_mutex_down(writeout_mutex);
1184         if (td->bw_log) {
1185                 if (td->o.bw_log_file) {
1186                         finish_log_named(td, td->bw_log,
1187                                                 td->o.bw_log_file, "bw");
1188                 } else
1189                         finish_log(td, td->bw_log, "bw");
1190         }
1191         if (td->lat_log) {
1192                 if (td->o.lat_log_file) {
1193                         finish_log_named(td, td->lat_log,
1194                                                 td->o.lat_log_file, "lat");
1195                 } else
1196                         finish_log(td, td->lat_log, "lat");
1197         }
1198         if (td->slat_log) {
1199                 if (td->o.lat_log_file) {
1200                         finish_log_named(td, td->slat_log,
1201                                                 td->o.lat_log_file, "slat");
1202                 } else
1203                         finish_log(td, td->slat_log, "slat");
1204         }
1205         if (td->clat_log) {
1206                 if (td->o.lat_log_file) {
1207                         finish_log_named(td, td->clat_log,
1208                                                 td->o.lat_log_file, "clat");
1209                 } else
1210                         finish_log(td, td->clat_log, "clat");
1211         }
1212         if (td->iops_log) {
1213                 if (td->o.iops_log_file) {
1214                         finish_log_named(td, td->iops_log,
1215                                                 td->o.iops_log_file, "iops");
1216                 } else
1217                         finish_log(td, td->iops_log, "iops");
1218         }
1219
1220         fio_mutex_up(writeout_mutex);
1221         if (td->o.exec_postrun)
1222                 exec_string(td->o.exec_postrun);
1223
1224         if (exitall_on_terminate)
1225                 fio_terminate_threads(td->groupid);
1226
1227 err:
1228         if (td->error)
1229                 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1230                                                         td->verror);
1231
1232         if (td->o.verify_async)
1233                 verify_async_exit(td);
1234
1235         close_and_free_files(td);
1236         close_ioengine(td);
1237         cleanup_io_u(td);
1238         cgroup_shutdown(td, &cgroup_mnt);
1239
1240         if (o->cpumask_set) {
1241                 int ret = fio_cpuset_exit(&o->cpumask);
1242
1243                 td_verror(td, ret, "fio_cpuset_exit");
1244         }
1245
1246         /*
1247          * do this very late, it will log file closing as well
1248          */
1249         if (td->o.write_iolog_file)
1250                 write_iolog_close(td);
1251
1252         td_set_runstate(td, TD_EXITED);
1253         return (void *) (uintptr_t) td->error;
1254 }
1255
1256
1257 /*
1258  * We cannot pass the td data into a forked process, so attach the td and
1259  * pass it to the thread worker.
1260  */
1261 static int fork_main(int shmid, int offset)
1262 {
1263         struct thread_data *td;
1264         void *data, *ret;
1265
1266 #ifndef __hpux
1267         data = shmat(shmid, NULL, 0);
1268         if (data == (void *) -1) {
1269                 int __err = errno;
1270
1271                 perror("shmat");
1272                 return __err;
1273         }
1274 #else
1275         /*
1276          * HP-UX inherits shm mappings?
1277          */
1278         data = threads;
1279 #endif
1280
1281         td = data + offset * sizeof(struct thread_data);
1282         ret = thread_main(td);
1283         shmdt(data);
1284         return (int) (uintptr_t) ret;
1285 }
1286
1287 /*
1288  * Run over the job map and reap the threads that have exited, if any.
1289  */
1290 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1291                          unsigned int *m_rate)
1292 {
1293         struct thread_data *td;
1294         unsigned int cputhreads, realthreads, pending;
1295         int i, status, ret;
1296
1297         /*
1298          * reap exited threads (TD_EXITED -> TD_REAPED)
1299          */
1300         realthreads = pending = cputhreads = 0;
1301         for_each_td(td, i) {
1302                 int flags = 0;
1303
1304                 /*
1305                  * ->io_ops is NULL for a thread that has closed its
1306                  * io engine
1307                  */
1308                 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1309                         cputhreads++;
1310                 else
1311                         realthreads++;
1312
1313                 if (!td->pid) {
1314                         pending++;
1315                         continue;
1316                 }
1317                 if (td->runstate == TD_REAPED)
1318                         continue;
1319                 if (td->o.use_thread) {
1320                         if (td->runstate == TD_EXITED) {
1321                                 td_set_runstate(td, TD_REAPED);
1322                                 goto reaped;
1323                         }
1324                         continue;
1325                 }
1326
1327                 flags = WNOHANG;
1328                 if (td->runstate == TD_EXITED)
1329                         flags = 0;
1330
1331                 /*
1332                  * check if someone quit or got killed in an unusual way
1333                  */
1334                 ret = waitpid(td->pid, &status, flags);
1335                 if (ret < 0) {
1336                         if (errno == ECHILD) {
1337                                 log_err("fio: pid=%d disappeared %d\n",
1338                                                 (int) td->pid, td->runstate);
1339                                 td->sig = ECHILD;
1340                                 td_set_runstate(td, TD_REAPED);
1341                                 goto reaped;
1342                         }
1343                         perror("waitpid");
1344                 } else if (ret == td->pid) {
1345                         if (WIFSIGNALED(status)) {
1346                                 int sig = WTERMSIG(status);
1347
1348                                 if (sig != SIGTERM)
1349                                         log_err("fio: pid=%d, got signal=%d\n",
1350                                                         (int) td->pid, sig);
1351                                 td->sig = sig;
1352                                 td_set_runstate(td, TD_REAPED);
1353                                 goto reaped;
1354                         }
1355                         if (WIFEXITED(status)) {
1356                                 if (WEXITSTATUS(status) && !td->error)
1357                                         td->error = WEXITSTATUS(status);
1358
1359                                 td_set_runstate(td, TD_REAPED);
1360                                 goto reaped;
1361                         }
1362                 }
1363
1364                 /*
1365                  * thread is not dead, continue
1366                  */
1367                 pending++;
1368                 continue;
1369 reaped:
1370                 (*nr_running)--;
1371                 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1372                 (*t_rate) -= ddir_rw_sum(td->o.rate);
1373                 if (!td->pid)
1374                         pending--;
1375
1376                 if (td->error)
1377                         exit_value++;
1378
1379                 done_secs += mtime_since_now(&td->epoch) / 1000;
1380         }
1381
1382         if (*nr_running == cputhreads && !pending && realthreads)
1383                 fio_terminate_threads(TERMINATE_ALL);
1384 }
1385
1386 /*
1387  * Main function for kicking off and reaping jobs, as needed.
1388  */
1389 static void run_threads(void)
1390 {
1391         struct thread_data *td;
1392         unsigned long spent;
1393         unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
1394
1395         if (fio_gtod_offload && fio_start_gtod_thread())
1396                 return;
1397
1398         set_sig_handlers();
1399
1400         nr_thread = nr_process = 0;
1401         for_each_td(td, i) {
1402                 if (td->o.use_thread)
1403                         nr_thread++;
1404                 else
1405                         nr_process++;
1406         }
1407
1408         if (output_format == FIO_OUTPUT_NORMAL) {
1409                 log_info("Starting ");
1410                 if (nr_thread)
1411                         log_info("%d thread%s", nr_thread,
1412                                                 nr_thread > 1 ? "s" : "");
1413                 if (nr_process) {
1414                         if (nr_thread)
1415                                 log_info(" and ");
1416                         log_info("%d process%s", nr_process,
1417                                                 nr_process > 1 ? "es" : "");
1418                 }
1419                 log_info("\n");
1420                 fflush(stdout);
1421         }
1422
1423         todo = thread_number;
1424         nr_running = 0;
1425         nr_started = 0;
1426         m_rate = t_rate = 0;
1427
1428         for_each_td(td, i) {
1429                 print_status_init(td->thread_number - 1);
1430
1431                 if (!td->o.create_serialize)
1432                         continue;
1433
1434                 /*
1435                  * do file setup here so it happens sequentially,
1436                  * we don't want X number of threads getting their
1437                  * client data interspersed on disk
1438                  */
1439                 if (setup_files(td)) {
1440                         exit_value++;
1441                         if (td->error)
1442                                 log_err("fio: pid=%d, err=%d/%s\n",
1443                                         (int) td->pid, td->error, td->verror);
1444                         td_set_runstate(td, TD_REAPED);
1445                         todo--;
1446                 } else {
1447                         struct fio_file *f;
1448                         unsigned int j;
1449
1450                         /*
1451                          * for sharing to work, each job must always open
1452                          * its own files. so close them, if we opened them
1453                          * for creation
1454                          */
1455                         for_each_file(td, f, j) {
1456                                 if (fio_file_open(f))
1457                                         td_io_close_file(td, f);
1458                         }
1459                 }
1460         }
1461
1462         set_genesis_time();
1463
1464         while (todo) {
1465                 struct thread_data *map[REAL_MAX_JOBS];
1466                 struct timeval this_start;
1467                 int this_jobs = 0, left;
1468
1469                 /*
1470                  * create threads (TD_NOT_CREATED -> TD_CREATED)
1471                  */
1472                 for_each_td(td, i) {
1473                         if (td->runstate != TD_NOT_CREATED)
1474                                 continue;
1475
1476                         /*
1477                          * never got a chance to start, killed by other
1478                          * thread for some reason
1479                          */
1480                         if (td->terminate) {
1481                                 todo--;
1482                                 continue;
1483                         }
1484
1485                         if (td->o.start_delay) {
1486                                 spent = mtime_since_genesis();
1487
1488                                 if (td->o.start_delay * 1000 > spent)
1489                                         continue;
1490                         }
1491
1492                         if (td->o.stonewall && (nr_started || nr_running)) {
1493                                 dprint(FD_PROCESS, "%s: stonewall wait\n",
1494                                                         td->o.name);
1495                                 break;
1496                         }
1497
1498                         init_disk_util(td);
1499
1500                         /*
1501                          * Set state to created. Thread will transition
1502                          * to TD_INITIALIZED when it's done setting up.
1503                          */
1504                         td_set_runstate(td, TD_CREATED);
1505                         map[this_jobs++] = td;
1506                         nr_started++;
1507
1508                         if (td->o.use_thread) {
1509                                 int ret;
1510
1511                                 dprint(FD_PROCESS, "will pthread_create\n");
1512                                 ret = pthread_create(&td->thread, NULL,
1513                                                         thread_main, td);
1514                                 if (ret) {
1515                                         log_err("pthread_create: %s\n",
1516                                                         strerror(ret));
1517                                         nr_started--;
1518                                         break;
1519                                 }
1520                                 ret = pthread_detach(td->thread);
1521                                 if (ret)
1522                                         log_err("pthread_detach: %s",
1523                                                         strerror(ret));
1524                         } else {
1525                                 pid_t pid;
1526                                 dprint(FD_PROCESS, "will fork\n");
1527                                 pid = fork();
1528                                 if (!pid) {
1529                                         int ret = fork_main(shm_id, i);
1530
1531                                         _exit(ret);
1532                                 } else if (i == fio_debug_jobno)
1533                                         *fio_debug_jobp = pid;
1534                         }
1535                         dprint(FD_MUTEX, "wait on startup_mutex\n");
1536                         if (fio_mutex_down_timeout(startup_mutex, 10)) {
1537                                 log_err("fio: job startup hung? exiting.\n");
1538                                 fio_terminate_threads(TERMINATE_ALL);
1539                                 fio_abort = 1;
1540                                 nr_started--;
1541                                 break;
1542                         }
1543                         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1544                 }
1545
1546                 /*
1547                  * Wait for the started threads to transition to
1548                  * TD_INITIALIZED.
1549                  */
1550                 fio_gettime(&this_start, NULL);
1551                 left = this_jobs;
1552                 while (left && !fio_abort) {
1553                         if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1554                                 break;
1555
1556                         usleep(100000);
1557
1558                         for (i = 0; i < this_jobs; i++) {
1559                                 td = map[i];
1560                                 if (!td)
1561                                         continue;
1562                                 if (td->runstate == TD_INITIALIZED) {
1563                                         map[i] = NULL;
1564                                         left--;
1565                                 } else if (td->runstate >= TD_EXITED) {
1566                                         map[i] = NULL;
1567                                         left--;
1568                                         todo--;
1569                                         nr_running++; /* work-around... */
1570                                 }
1571                         }
1572                 }
1573
1574                 if (left) {
1575                         log_err("fio: %d job%s failed to start\n", left,
1576                                         left > 1 ? "s" : "");
1577                         for (i = 0; i < this_jobs; i++) {
1578                                 td = map[i];
1579                                 if (!td)
1580                                         continue;
1581                                 kill(td->pid, SIGTERM);
1582                         }
1583                         break;
1584                 }
1585
1586                 /*
1587                  * start created threads (TD_INITIALIZED -> TD_RUNNING).
1588                  */
1589                 for_each_td(td, i) {
1590                         if (td->runstate != TD_INITIALIZED)
1591                                 continue;
1592
1593                         if (in_ramp_time(td))
1594                                 td_set_runstate(td, TD_RAMP);
1595                         else
1596                                 td_set_runstate(td, TD_RUNNING);
1597                         nr_running++;
1598                         nr_started--;
1599                         m_rate += ddir_rw_sum(td->o.ratemin);
1600                         t_rate += ddir_rw_sum(td->o.rate);
1601                         todo--;
1602                         fio_mutex_up(td->mutex);
1603                 }
1604
1605                 reap_threads(&nr_running, &t_rate, &m_rate);
1606
1607                 if (todo)
1608                         usleep(100000);
1609         }
1610
1611         while (nr_running) {
1612                 reap_threads(&nr_running, &t_rate, &m_rate);
1613                 usleep(10000);
1614         }
1615
1616         update_io_ticks();
1617 }
1618
1619 void wait_for_disk_thread_exit(void)
1620 {
1621         fio_mutex_down(disk_thread_mutex);
1622 }
1623
1624 static void *disk_thread_main(void *data)
1625 {
1626         int ret = 0;
1627
1628         fio_mutex_up(startup_mutex);
1629
1630         while (threads && !ret) {
1631                 usleep(DISK_UTIL_MSEC * 1000);
1632                 if (!threads)
1633                         break;
1634                 ret = update_io_ticks();
1635
1636                 if (!is_backend)
1637                         print_thread_status();
1638         }
1639
1640         fio_mutex_up(disk_thread_mutex);
1641         return NULL;
1642 }
1643
1644 static int create_disk_util_thread(void)
1645 {
1646         int ret;
1647
1648         setup_disk_util();
1649
1650         disk_thread_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
1651
1652         ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
1653         if (ret) {
1654                 fio_mutex_remove(disk_thread_mutex);
1655                 log_err("Can't create disk util thread: %s\n", strerror(ret));
1656                 return 1;
1657         }
1658
1659         ret = pthread_detach(disk_util_thread);
1660         if (ret) {
1661                 fio_mutex_remove(disk_thread_mutex);
1662                 log_err("Can't detatch disk util thread: %s\n", strerror(ret));
1663                 return 1;
1664         }
1665
1666         dprint(FD_MUTEX, "wait on startup_mutex\n");
1667         fio_mutex_down(startup_mutex);
1668         dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1669         return 0;
1670 }
1671
1672 int fio_backend(void)
1673 {
1674         struct thread_data *td;
1675         int i;
1676
1677         if (exec_profile) {
1678                 if (load_profile(exec_profile))
1679                         return 1;
1680                 free(exec_profile);
1681                 exec_profile = NULL;
1682         }
1683         if (!thread_number)
1684                 return 0;
1685
1686         if (write_bw_log) {
1687                 setup_log(&agg_io_log[DDIR_READ], 0, IO_LOG_TYPE_BW);
1688                 setup_log(&agg_io_log[DDIR_WRITE], 0, IO_LOG_TYPE_BW);
1689                 setup_log(&agg_io_log[DDIR_TRIM], 0, IO_LOG_TYPE_BW);
1690         }
1691
1692         startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
1693         if (startup_mutex == NULL)
1694                 return 1;
1695         writeout_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1696         if (writeout_mutex == NULL)
1697                 return 1;
1698
1699         set_genesis_time();
1700         create_disk_util_thread();
1701
1702         cgroup_list = smalloc(sizeof(*cgroup_list));
1703         INIT_FLIST_HEAD(cgroup_list);
1704
1705         run_threads();
1706
1707         if (!fio_abort) {
1708                 show_run_stats();
1709                 if (write_bw_log) {
1710                         __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1711                         __finish_log(agg_io_log[DDIR_WRITE],
1712                                         "agg-write_bw.log");
1713                         __finish_log(agg_io_log[DDIR_TRIM],
1714                                         "agg-write_bw.log");
1715                 }
1716         }
1717
1718         for_each_td(td, i)
1719                 fio_options_free(td);
1720
1721         free_disk_util();
1722         cgroup_kill(cgroup_list);
1723         sfree(cgroup_list);
1724         sfree(cgroup_mnt);
1725
1726         fio_mutex_remove(startup_mutex);
1727         fio_mutex_remove(writeout_mutex);
1728         fio_mutex_remove(disk_thread_mutex);
1729         return exit_value;
1730 }