2 * fio - the flexible io tester
4 * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
5 * Copyright (C) 2006 Jens Axboe <axboe@kernel.dk>
7 * The license below covers all files distributed with fio unless otherwise
8 * noted in the file itself.
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.
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.
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
43 unsigned long page_mask;
44 unsigned long page_size;
46 #define PAGE_ALIGN(buf) \
47 (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
50 int thread_number = 0;
55 unsigned long done_secs = 0;
57 static struct fio_mutex *startup_mutex;
58 static struct fio_mutex *writeout_mutex;
59 static volatile int fio_abort;
60 static int exit_value;
61 static struct itimerval itimer;
62 static pthread_t gtod_thread;
64 struct io_log *agg_io_log[2];
66 #define TERMINATE_ALL (-1)
67 #define JOB_START_TIMEOUT (5 * 1000)
69 void td_set_runstate(struct thread_data *td, int runstate)
71 if (td->runstate == runstate)
74 dprint(FD_PROCESS, "pid=%d: runstate %d -> %d\n", (int) td->pid,
75 td->runstate, runstate);
76 td->runstate = runstate;
79 static void terminate_threads(int group_id)
81 struct thread_data *td;
84 dprint(FD_PROCESS, "terminate group_id=%d\n", group_id);
87 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
88 dprint(FD_PROCESS, "setting terminate on %s/%d\n",
89 td->o.name, (int) td->pid);
91 td->o.start_delay = 0;
94 * if the thread is running, just let it exit
96 if (td->runstate < TD_RUNNING)
97 kill(td->pid, SIGQUIT);
99 struct ioengine_ops *ops = td->io_ops;
101 if (ops && (ops->flags & FIO_SIGQUIT))
102 kill(td->pid, SIGQUIT);
108 static void status_timer_arm(void)
110 itimer.it_value.tv_sec = 0;
111 itimer.it_value.tv_usec = DISK_UTIL_MSEC * 1000;
112 setitimer(ITIMER_REAL, &itimer, NULL);
115 static void sig_alrm(int fio_unused sig)
119 print_thread_status();
125 * Happens on thread runs with ctrl-c, ignore our own SIGQUIT
127 static void sig_quit(int sig)
131 static void sig_int(int sig)
134 printf("\nfio: terminating on signal %d\n", sig);
136 terminate_threads(TERMINATE_ALL);
140 static void sig_ill(int fio_unused sig)
145 log_err("fio: illegal instruction. your cpu does not support "
146 "the sse4.2 instruction for crc32c\n");
147 terminate_threads(TERMINATE_ALL);
151 static void set_sig_handlers(void)
153 struct sigaction act;
155 memset(&act, 0, sizeof(act));
156 act.sa_handler = sig_alrm;
157 act.sa_flags = SA_RESTART;
158 sigaction(SIGALRM, &act, NULL);
160 memset(&act, 0, sizeof(act));
161 act.sa_handler = sig_int;
162 act.sa_flags = SA_RESTART;
163 sigaction(SIGINT, &act, NULL);
165 memset(&act, 0, sizeof(act));
166 act.sa_handler = sig_ill;
167 act.sa_flags = SA_RESTART;
168 sigaction(SIGILL, &act, NULL);
170 memset(&act, 0, sizeof(act));
171 act.sa_handler = sig_quit;
172 act.sa_flags = SA_RESTART;
173 sigaction(SIGQUIT, &act, NULL);
177 * Check if we are above the minimum rate given.
179 static int __check_min_rate(struct thread_data *td, struct timeval *now,
182 unsigned long long bytes = 0;
183 unsigned long iops = 0;
186 unsigned int ratemin = 0;
187 unsigned int rate_iops = 0;
188 unsigned int rate_iops_min = 0;
190 if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
194 * allow a 2 second settle period in the beginning
196 if (mtime_since(&td->start, now) < 2000)
199 iops += td->io_blocks[ddir];
200 bytes += td->this_io_bytes[ddir];
201 ratemin += td->o.ratemin[ddir];
202 rate_iops += td->o.rate_iops[ddir];
203 rate_iops_min += td->o.rate_iops_min[ddir];
206 * if rate blocks is set, sample is running
208 if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
209 spent = mtime_since(&td->lastrate[ddir], now);
210 if (spent < td->o.ratecycle)
213 if (td->o.rate[ddir]) {
215 * check bandwidth specified rate
217 if (bytes < td->rate_bytes[ddir]) {
218 log_err("%s: min rate %u not met\n", td->o.name,
222 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
223 if (rate < ratemin ||
224 bytes < td->rate_bytes[ddir]) {
225 log_err("%s: min rate %u not met, got"
226 " %luKiB/sec\n", td->o.name,
233 * checks iops specified rate
235 if (iops < rate_iops) {
236 log_err("%s: min iops rate %u not met\n",
237 td->o.name, rate_iops);
240 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
241 if (rate < rate_iops_min ||
242 iops < td->rate_blocks[ddir]) {
243 log_err("%s: min iops rate %u not met,"
244 " got %lu\n", td->o.name,
245 rate_iops_min, rate);
251 td->rate_bytes[ddir] = bytes;
252 td->rate_blocks[ddir] = iops;
253 memcpy(&td->lastrate[ddir], now, sizeof(*now));
257 static int check_min_rate(struct thread_data *td, struct timeval *now,
258 unsigned long *bytes_done)
263 ret |= __check_min_rate(td, now, 0);
265 ret |= __check_min_rate(td, now, 1);
270 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
274 if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
281 * When job exits, we can cancel the in-flight IO if we are using async
282 * io. Attempt to do so.
284 static void cleanup_pending_aio(struct thread_data *td)
286 struct flist_head *entry, *n;
291 * get immediately available events, if any
293 r = io_u_queued_complete(td, 0, NULL);
298 * now cancel remaining active events
300 if (td->io_ops->cancel) {
301 flist_for_each_safe(entry, n, &td->io_u_busylist) {
302 io_u = flist_entry(entry, struct io_u, list);
305 * if the io_u isn't in flight, then that generally
306 * means someone leaked an io_u. complain but fix
307 * it up, so we don't stall here.
309 if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
310 log_err("fio: non-busy IO on busy list\n");
313 r = td->io_ops->cancel(td, io_u);
321 r = io_u_queued_complete(td, td->cur_depth, NULL);
325 * Helper to handle the final sync of a file. Works just like the normal
326 * io path, just does everything sync.
328 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
330 struct io_u *io_u = __get_io_u(td);
336 io_u->ddir = DDIR_SYNC;
339 if (td_io_prep(td, io_u)) {
345 ret = td_io_queue(td, io_u);
347 td_verror(td, io_u->error, "td_io_queue");
350 } else if (ret == FIO_Q_QUEUED) {
351 if (io_u_queued_complete(td, 1, NULL) < 0)
353 } else if (ret == FIO_Q_COMPLETED) {
355 td_verror(td, io_u->error, "td_io_queue");
359 if (io_u_sync_complete(td, io_u, NULL) < 0)
361 } else if (ret == FIO_Q_BUSY) {
362 if (td_io_commit(td))
370 static inline void update_tv_cache(struct thread_data *td)
372 if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
373 fio_gettime(&td->tv_cache, NULL);
376 static int break_on_this_error(struct thread_data *td, int *retptr)
380 if (ret < 0 || td->error) {
383 if (!td->o.continue_on_error)
391 update_error_count(td, err);
393 if (td_non_fatal_error(err)) {
395 * Continue with the I/Os in case of
403 * Stop the I/O in case of a fatal
414 * The main verify engine. Runs over the writes we previously submitted,
415 * reads the blocks back in, and checks the crc/md5 of the data.
417 static void do_verify(struct thread_data *td)
425 * sync io first and invalidate cache, to make sure we really
428 for_each_file(td, f, i) {
429 if (!fio_file_open(f))
431 if (fio_io_sync(td, f))
433 if (file_invalidate_cache(td, f))
440 td_set_runstate(td, TD_VERIFYING);
443 while (!td->terminate) {
448 if (runtime_exceeded(td, &td->tv_cache)) {
453 io_u = __get_io_u(td);
457 if (get_next_verify(td, io_u)) {
462 if (td_io_prep(td, io_u)) {
467 io_u->end_io = verify_io_u;
469 ret = td_io_queue(td, io_u);
471 case FIO_Q_COMPLETED:
474 clear_io_u(td, io_u);
475 } else if (io_u->resid) {
476 int bytes = io_u->xfer_buflen - io_u->resid;
477 struct fio_file *f = io_u->file;
483 td_verror(td, EIO, "full resid");
488 io_u->xfer_buflen = io_u->resid;
489 io_u->xfer_buf += bytes;
490 io_u->offset += bytes;
492 td->ts.short_io_u[io_u->ddir]++;
494 if (io_u->offset == f->real_file_size)
497 requeue_io_u(td, &io_u);
500 ret = io_u_sync_complete(td, io_u, NULL);
508 requeue_io_u(td, &io_u);
509 ret2 = td_io_commit(td);
515 td_verror(td, -ret, "td_io_queue");
519 if (break_on_this_error(td, &ret))
523 * if we can queue more, do so. but check if there are
524 * completed io_u's first.
526 full = queue_full(td) || ret == FIO_Q_BUSY;
527 if (full || !td->o.iodepth_batch_complete) {
528 min_events = td->o.iodepth_batch_complete;
529 if (full && !min_events)
534 * Reap required number of io units, if any,
535 * and do the verification on them through
536 * the callback handler
538 if (io_u_queued_complete(td, min_events, NULL) < 0) {
542 } while (full && (td->cur_depth > td->o.iodepth_low));
549 min_events = td->cur_depth;
552 ret = io_u_queued_complete(td, min_events, NULL);
554 cleanup_pending_aio(td);
556 td_set_runstate(td, TD_RUNNING);
560 * Main IO worker function. It retrieves io_u's to process and queues
561 * and reaps them, checking for rate and errors along the way.
563 static void do_io(struct thread_data *td)
568 if (in_ramp_time(td))
569 td_set_runstate(td, TD_RAMP);
571 td_set_runstate(td, TD_RUNNING);
573 while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) {
574 struct timeval comp_time;
575 unsigned long bytes_done[2] = { 0, 0 };
585 if (runtime_exceeded(td, &td->tv_cache)) {
595 * Add verification end_io handler, if asked to verify
596 * a previously written file.
598 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ) {
599 io_u->end_io = verify_io_u;
600 td_set_runstate(td, TD_VERIFYING);
601 } else if (in_ramp_time(td))
602 td_set_runstate(td, TD_RAMP);
604 td_set_runstate(td, TD_RUNNING);
606 ret = td_io_queue(td, io_u);
608 case FIO_Q_COMPLETED:
611 clear_io_u(td, io_u);
612 } else if (io_u->resid) {
613 int bytes = io_u->xfer_buflen - io_u->resid;
614 struct fio_file *f = io_u->file;
620 td_verror(td, EIO, "full resid");
625 io_u->xfer_buflen = io_u->resid;
626 io_u->xfer_buf += bytes;
627 io_u->offset += bytes;
629 td->ts.short_io_u[io_u->ddir]++;
631 if (io_u->offset == f->real_file_size)
634 requeue_io_u(td, &io_u);
637 if (__should_check_rate(td, 0) ||
638 __should_check_rate(td, 1))
639 fio_gettime(&comp_time, NULL);
641 ret = io_u_sync_complete(td, io_u, bytes_done);
648 * if the engine doesn't have a commit hook,
649 * the io_u is really queued. if it does have such
650 * a hook, it has to call io_u_queued() itself.
652 if (td->io_ops->commit == NULL)
653 io_u_queued(td, io_u);
656 requeue_io_u(td, &io_u);
657 ret2 = td_io_commit(td);
667 if (break_on_this_error(td, &ret))
671 * See if we need to complete some commands
673 full = queue_full(td) || ret == FIO_Q_BUSY;
674 if (full || !td->o.iodepth_batch_complete) {
675 min_evts = td->o.iodepth_batch_complete;
676 if (full && !min_evts)
679 if (__should_check_rate(td, 0) ||
680 __should_check_rate(td, 1))
681 fio_gettime(&comp_time, NULL);
684 ret = io_u_queued_complete(td, min_evts, bytes_done);
688 } while (full && (td->cur_depth > td->o.iodepth_low));
693 if (!(bytes_done[0] + bytes_done[1]))
696 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
697 if (check_min_rate(td, &comp_time, bytes_done)) {
698 if (exitall_on_terminate)
699 terminate_threads(td->groupid);
700 td_verror(td, EIO, "check_min_rate");
705 if (td->o.thinktime) {
706 unsigned long long b;
708 b = td->io_blocks[0] + td->io_blocks[1];
709 if (!(b % td->o.thinktime_blocks)) {
712 if (td->o.thinktime_spin)
713 usec_spin(td->o.thinktime_spin);
715 left = td->o.thinktime - td->o.thinktime_spin;
717 usec_sleep(td, left);
722 if (td->o.fill_device && td->error == ENOSPC) {
731 ret = io_u_queued_complete(td, i, NULL);
733 if (should_fsync(td) && td->o.end_fsync) {
734 td_set_runstate(td, TD_FSYNCING);
736 for_each_file(td, f, i) {
737 if (!fio_file_open(f))
743 cleanup_pending_aio(td);
746 * stop job if we failed doing any IO
748 if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
752 static void cleanup_io_u(struct thread_data *td)
754 struct flist_head *entry, *n;
757 flist_for_each_safe(entry, n, &td->io_u_freelist) {
758 io_u = flist_entry(entry, struct io_u, list);
760 flist_del(&io_u->list);
767 static int init_io_u(struct thread_data *td)
771 int cl_align, i, max_units;
774 max_units = td->o.iodepth;
775 max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
776 td->orig_buffer_size = (unsigned long long) max_bs
777 * (unsigned long long) max_units;
779 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
782 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
783 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
786 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
787 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
791 if (allocate_io_mem(td))
795 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
799 cl_align = os_cache_line_size();
801 for (i = 0; i < max_units; i++) {
807 if (posix_memalign(&ptr, cl_align, sizeof(*io_u))) {
808 log_err("fio: posix_memalign=%s\n", strerror(errno));
813 memset(io_u, 0, sizeof(*io_u));
814 INIT_FLIST_HEAD(&io_u->list);
815 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
817 if (!(td->io_ops->flags & FIO_NOIO)) {
818 io_u->buf = p + max_bs * i;
819 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
821 if (td_write(td) && !td->o.refill_buffers)
822 io_u_fill_buffer(td, io_u, max_bs);
826 io_u->flags = IO_U_F_FREE;
827 flist_add(&io_u->list, &td->io_u_freelist);
833 static int switch_ioscheduler(struct thread_data *td)
835 char tmp[256], tmp2[128];
839 if (td->io_ops->flags & FIO_DISKLESSIO)
842 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
844 f = fopen(tmp, "r+");
846 if (errno == ENOENT) {
847 log_err("fio: os or kernel doesn't support IO scheduler"
851 td_verror(td, errno, "fopen iosched");
858 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
859 if (ferror(f) || ret != 1) {
860 td_verror(td, errno, "fwrite");
868 * Read back and check that the selected scheduler is now the default.
870 ret = fread(tmp, 1, sizeof(tmp), f);
871 if (ferror(f) || ret < 0) {
872 td_verror(td, errno, "fread");
877 sprintf(tmp2, "[%s]", td->o.ioscheduler);
878 if (!strstr(tmp, tmp2)) {
879 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
880 td_verror(td, EINVAL, "iosched_switch");
889 static int keep_running(struct thread_data *td)
891 unsigned long long io_done;
895 if (td->o.time_based)
902 io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
904 if (io_done < td->o.size)
910 static void reset_io_counters(struct thread_data *td)
912 td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
913 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
915 td->rate_bytes[0] = td->rate_bytes[1] = 0;
916 td->rate_blocks[0] = td->rate_blocks[1] = 0;
918 td->last_was_sync = 0;
921 * reset file done count if we are to start over
923 if (td->o.time_based || td->o.loops)
924 td->nr_done_files = 0;
927 * Set the same seed to get repeatable runs
929 td_fill_rand_seeds(td);
932 void reset_all_stats(struct thread_data *td)
937 reset_io_counters(td);
939 for (i = 0; i < 2; i++) {
941 td->io_blocks[i] = 0;
942 td->io_issues[i] = 0;
943 td->ts.total_io_u[i] = 0;
946 fio_gettime(&tv, NULL);
947 memcpy(&td->epoch, &tv, sizeof(tv));
948 memcpy(&td->start, &tv, sizeof(tv));
951 static void clear_io_state(struct thread_data *td)
956 reset_io_counters(td);
959 for_each_file(td, f, i)
960 fio_file_clear_done(f);
963 static int exec_string(const char *string)
965 int ret, newlen = strlen(string) + 1 + 8;
968 str = malloc(newlen);
969 sprintf(str, "sh -c %s", string);
973 log_err("fio: exec of cmd <%s> failed\n", str);
980 * Entry point for the thread based jobs. The process based jobs end up
981 * here as well, after a little setup.
983 static void *thread_main(void *data)
985 unsigned long long runtime[2], elapsed;
986 struct thread_data *td = data;
989 if (!td->o.use_thread)
994 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
996 INIT_FLIST_HEAD(&td->io_u_freelist);
997 INIT_FLIST_HEAD(&td->io_u_busylist);
998 INIT_FLIST_HEAD(&td->io_u_requeues);
999 INIT_FLIST_HEAD(&td->io_log_list);
1000 INIT_FLIST_HEAD(&td->io_hist_list);
1001 td->io_hist_tree = RB_ROOT;
1003 td_set_runstate(td, TD_INITIALIZED);
1004 dprint(FD_MUTEX, "up startup_mutex\n");
1005 fio_mutex_up(startup_mutex);
1006 dprint(FD_MUTEX, "wait on td->mutex\n");
1007 fio_mutex_down(td->mutex);
1008 dprint(FD_MUTEX, "done waiting on td->mutex\n");
1011 * the ->mutex mutex is now no longer used, close it to avoid
1012 * eating a file descriptor
1014 fio_mutex_remove(td->mutex);
1017 * May alter parameters that init_io_u() will use, so we need to
1026 if (td->o.cpumask_set && fio_setaffinity(td) == -1) {
1027 td_verror(td, errno, "cpu_set_affinity");
1032 * If we have a gettimeofday() thread, make sure we exclude that
1033 * thread from this job
1035 if (td->o.gtod_cpu) {
1036 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1037 if (fio_setaffinity(td) == -1) {
1038 td_verror(td, errno, "cpu_set_affinity");
1043 if (td->ioprio_set) {
1044 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1045 td_verror(td, errno, "ioprio_set");
1050 if (nice(td->o.nice) == -1) {
1051 td_verror(td, errno, "nice");
1055 if (td->o.ioscheduler && switch_ioscheduler(td))
1058 if (!td->o.create_serialize && setup_files(td))
1064 if (init_random_map(td))
1067 if (td->o.exec_prerun) {
1068 if (exec_string(td->o.exec_prerun))
1072 if (td->o.pre_read) {
1073 if (pre_read_files(td) < 0)
1077 fio_gettime(&td->epoch, NULL);
1078 getrusage(RUSAGE_SELF, &td->ts.ru_start);
1080 runtime[0] = runtime[1] = 0;
1082 while (keep_running(td)) {
1083 fio_gettime(&td->start, NULL);
1084 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
1085 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1087 if (td->o.ratemin[0] || td->o.ratemin[1])
1088 memcpy(&td->lastrate, &td->ts.stat_sample_time,
1089 sizeof(td->lastrate));
1094 prune_io_piece_log(td);
1100 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1101 elapsed = utime_since_now(&td->start);
1102 runtime[DDIR_READ] += elapsed;
1104 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1105 elapsed = utime_since_now(&td->start);
1106 runtime[DDIR_WRITE] += elapsed;
1109 if (td->error || td->terminate)
1112 if (!td->o.do_verify ||
1113 td->o.verify == VERIFY_NONE ||
1114 (td->io_ops->flags & FIO_UNIDIR))
1119 fio_gettime(&td->start, NULL);
1123 runtime[DDIR_READ] += utime_since_now(&td->start);
1125 if (td->error || td->terminate)
1129 update_rusage_stat(td);
1130 td->ts.runtime[0] = (runtime[0] + 999) / 1000;
1131 td->ts.runtime[1] = (runtime[1] + 999) / 1000;
1132 td->ts.total_run_time = mtime_since_now(&td->epoch);
1133 td->ts.io_bytes[0] = td->io_bytes[0];
1134 td->ts.io_bytes[1] = td->io_bytes[1];
1136 fio_mutex_down(writeout_mutex);
1137 if (td->ts.bw_log) {
1138 if (td->o.bw_log_file) {
1139 finish_log_named(td, td->ts.bw_log,
1140 td->o.bw_log_file, "bw");
1142 finish_log(td, td->ts.bw_log, "bw");
1144 if (td->ts.slat_log) {
1145 if (td->o.lat_log_file) {
1146 finish_log_named(td, td->ts.slat_log,
1147 td->o.lat_log_file, "slat");
1149 finish_log(td, td->ts.slat_log, "slat");
1151 if (td->ts.clat_log) {
1152 if (td->o.lat_log_file) {
1153 finish_log_named(td, td->ts.clat_log,
1154 td->o.lat_log_file, "clat");
1156 finish_log(td, td->ts.clat_log, "clat");
1158 fio_mutex_up(writeout_mutex);
1159 if (td->o.exec_postrun)
1160 exec_string(td->o.exec_postrun);
1162 if (exitall_on_terminate)
1163 terminate_threads(td->groupid);
1167 printf("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1169 close_and_free_files(td);
1173 if (td->o.cpumask_set) {
1174 int ret = fio_cpuset_exit(&td->o.cpumask);
1176 td_verror(td, ret, "fio_cpuset_exit");
1180 * do this very late, it will log file closing as well
1182 if (td->o.write_iolog_file)
1183 write_iolog_close(td);
1185 options_mem_free(td);
1186 td_set_runstate(td, TD_EXITED);
1187 return (void *) (unsigned long) td->error;
1191 * We cannot pass the td data into a forked process, so attach the td and
1192 * pass it to the thread worker.
1194 static int fork_main(int shmid, int offset)
1196 struct thread_data *td;
1199 data = shmat(shmid, NULL, 0);
1200 if (data == (void *) -1) {
1207 td = data + offset * sizeof(struct thread_data);
1208 ret = thread_main(td);
1210 return (int) (unsigned long) ret;
1214 * Run over the job map and reap the threads that have exited, if any.
1216 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1218 struct thread_data *td;
1219 int i, cputhreads, realthreads, pending, status, ret;
1222 * reap exited threads (TD_EXITED -> TD_REAPED)
1224 realthreads = pending = cputhreads = 0;
1225 for_each_td(td, i) {
1229 * ->io_ops is NULL for a thread that has closed its
1232 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1241 if (td->runstate == TD_REAPED)
1243 if (td->o.use_thread) {
1244 if (td->runstate == TD_EXITED) {
1245 td_set_runstate(td, TD_REAPED);
1252 if (td->runstate == TD_EXITED)
1256 * check if someone quit or got killed in an unusual way
1258 ret = waitpid(td->pid, &status, flags);
1260 if (errno == ECHILD) {
1261 log_err("fio: pid=%d disappeared %d\n",
1262 (int) td->pid, td->runstate);
1263 td_set_runstate(td, TD_REAPED);
1267 } else if (ret == td->pid) {
1268 if (WIFSIGNALED(status)) {
1269 int sig = WTERMSIG(status);
1272 log_err("fio: pid=%d, got signal=%d\n",
1273 (int) td->pid, sig);
1274 td_set_runstate(td, TD_REAPED);
1277 if (WIFEXITED(status)) {
1278 if (WEXITSTATUS(status) && !td->error)
1279 td->error = WEXITSTATUS(status);
1281 td_set_runstate(td, TD_REAPED);
1287 * thread is not dead, continue
1293 (*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
1294 (*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
1301 done_secs += mtime_since_now(&td->epoch) / 1000;
1304 if (*nr_running == cputhreads && !pending && realthreads)
1305 terminate_threads(TERMINATE_ALL);
1308 static void *gtod_thread_main(void *data)
1310 fio_mutex_up(startup_mutex);
1313 * As long as we have jobs around, update the clock. It would be nice
1314 * to have some way of NOT hammering that CPU with gettimeofday(),
1315 * but I'm not sure what to use outside of a simple CPU nop to relax
1316 * it - we don't want to lose precision.
1326 static int fio_start_gtod_thread(void)
1330 ret = pthread_create(>od_thread, NULL, gtod_thread_main, NULL);
1332 log_err("Can't create gtod thread: %s\n", strerror(ret));
1336 ret = pthread_detach(gtod_thread);
1338 log_err("Can't detatch gtod thread: %s\n", strerror(ret));
1342 dprint(FD_MUTEX, "wait on startup_mutex\n");
1343 fio_mutex_down(startup_mutex);
1344 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1349 * Main function for kicking off and reaping jobs, as needed.
1351 static void run_threads(void)
1353 struct thread_data *td;
1354 unsigned long spent;
1355 int i, todo, nr_running, m_rate, t_rate, nr_started;
1357 if (fio_pin_memory())
1360 if (fio_gtod_offload && fio_start_gtod_thread())
1363 if (!terse_output) {
1364 printf("Starting ");
1366 printf("%d thread%s", nr_thread,
1367 nr_thread > 1 ? "s" : "");
1371 printf("%d process%s", nr_process,
1372 nr_process > 1 ? "es" : "");
1380 todo = thread_number;
1383 m_rate = t_rate = 0;
1385 for_each_td(td, i) {
1386 print_status_init(td->thread_number - 1);
1388 if (!td->o.create_serialize) {
1394 * do file setup here so it happens sequentially,
1395 * we don't want X number of threads getting their
1396 * client data interspersed on disk
1398 if (setup_files(td)) {
1401 log_err("fio: pid=%d, err=%d/%s\n",
1402 (int) td->pid, td->error, td->verror);
1403 td_set_runstate(td, TD_REAPED);
1410 * for sharing to work, each job must always open
1411 * its own files. so close them, if we opened them
1414 for_each_file(td, f, i) {
1415 if (fio_file_open(f))
1416 td_io_close_file(td, f);
1426 struct thread_data *map[MAX_JOBS];
1427 struct timeval this_start;
1428 int this_jobs = 0, left;
1431 * create threads (TD_NOT_CREATED -> TD_CREATED)
1433 for_each_td(td, i) {
1434 if (td->runstate != TD_NOT_CREATED)
1438 * never got a chance to start, killed by other
1439 * thread for some reason
1441 if (td->terminate) {
1446 if (td->o.start_delay) {
1447 spent = mtime_since_genesis();
1449 if (td->o.start_delay * 1000 > spent)
1453 if (td->o.stonewall && (nr_started || nr_running)) {
1454 dprint(FD_PROCESS, "%s: stonewall wait\n",
1460 * Set state to created. Thread will transition
1461 * to TD_INITIALIZED when it's done setting up.
1463 td_set_runstate(td, TD_CREATED);
1464 map[this_jobs++] = td;
1467 if (td->o.use_thread) {
1470 dprint(FD_PROCESS, "will pthread_create\n");
1471 ret = pthread_create(&td->thread, NULL,
1474 log_err("pthread_create: %s\n",
1479 ret = pthread_detach(td->thread);
1481 log_err("pthread_detach: %s",
1485 dprint(FD_PROCESS, "will fork\n");
1488 int ret = fork_main(shm_id, i);
1491 } else if (i == fio_debug_jobno)
1492 *fio_debug_jobp = pid;
1494 dprint(FD_MUTEX, "wait on startup_mutex\n");
1495 fio_mutex_down(startup_mutex);
1496 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1500 * Wait for the started threads to transition to
1503 fio_gettime(&this_start, NULL);
1505 while (left && !fio_abort) {
1506 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1511 for (i = 0; i < this_jobs; i++) {
1515 if (td->runstate == TD_INITIALIZED) {
1518 } else if (td->runstate >= TD_EXITED) {
1522 nr_running++; /* work-around... */
1528 log_err("fio: %d jobs failed to start\n", left);
1529 for (i = 0; i < this_jobs; i++) {
1533 kill(td->pid, SIGTERM);
1539 * start created threads (TD_INITIALIZED -> TD_RUNNING).
1541 for_each_td(td, i) {
1542 if (td->runstate != TD_INITIALIZED)
1545 if (in_ramp_time(td))
1546 td_set_runstate(td, TD_RAMP);
1548 td_set_runstate(td, TD_RUNNING);
1551 m_rate += td->o.ratemin[0] + td->o.ratemin[1];
1552 t_rate += td->o.rate[0] + td->o.rate[1];
1554 fio_mutex_up(td->mutex);
1557 reap_threads(&nr_running, &t_rate, &m_rate);
1563 while (nr_running) {
1564 reap_threads(&nr_running, &t_rate, &m_rate);
1572 int main(int argc, char *argv[])
1579 * We need locale for number printing, if it isn't set then just
1580 * go with the US format.
1582 if (!getenv("LC_NUMERIC"))
1583 setlocale(LC_NUMERIC, "en_US");
1585 if (parse_options(argc, argv))
1591 ps = sysconf(_SC_PAGESIZE);
1593 log_err("Failed to get page size\n");
1601 setup_log(&agg_io_log[DDIR_READ]);
1602 setup_log(&agg_io_log[DDIR_WRITE]);
1605 startup_mutex = fio_mutex_init(0);
1606 writeout_mutex = fio_mutex_init(1);
1617 __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1618 __finish_log(agg_io_log[DDIR_WRITE],
1619 "agg-write_bw.log");
1623 fio_mutex_remove(startup_mutex);
1624 fio_mutex_remove(writeout_mutex);