2 * fio - the flexible io tester
4 * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
5 * Copyright (C) 2006-2012 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
52 static pthread_t disk_util_thread;
53 static struct fio_mutex *startup_mutex;
54 static struct fio_mutex *writeout_mutex;
55 static struct flist_head *cgroup_list;
56 static char *cgroup_mnt;
57 static int exit_value;
58 static volatile int fio_abort;
60 struct io_log *agg_io_log[2];
63 unsigned int thread_number = 0;
64 unsigned int nr_process = 0;
65 unsigned int nr_thread = 0;
68 unsigned long done_secs = 0;
70 #define PAGE_ALIGN(buf) \
71 (char *) (((uintptr_t) (buf) + page_mask) & ~page_mask)
73 #define JOB_START_TIMEOUT (5 * 1000)
75 static void sig_int(int sig)
79 fio_server_got_signal(sig);
81 log_info("\nfio: terminating on signal %d\n", sig);
86 fio_terminate_threads(TERMINATE_ALL);
90 static void set_sig_handlers(void)
94 memset(&act, 0, sizeof(act));
95 act.sa_handler = sig_int;
96 act.sa_flags = SA_RESTART;
97 sigaction(SIGINT, &act, NULL);
99 memset(&act, 0, sizeof(act));
100 act.sa_handler = sig_int;
101 act.sa_flags = SA_RESTART;
102 sigaction(SIGTERM, &act, NULL);
105 memset(&act, 0, sizeof(act));
106 act.sa_handler = sig_int;
107 act.sa_flags = SA_RESTART;
108 sigaction(SIGPIPE, &act, NULL);
113 * Check if we are above the minimum rate given.
115 static int __check_min_rate(struct thread_data *td, struct timeval *now,
118 unsigned long long bytes = 0;
119 unsigned long iops = 0;
122 unsigned int ratemin = 0;
123 unsigned int rate_iops = 0;
124 unsigned int rate_iops_min = 0;
126 assert(ddir_rw(ddir));
128 if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
132 * allow a 2 second settle period in the beginning
134 if (mtime_since(&td->start, now) < 2000)
137 iops += td->this_io_blocks[ddir];
138 bytes += td->this_io_bytes[ddir];
139 ratemin += td->o.ratemin[ddir];
140 rate_iops += td->o.rate_iops[ddir];
141 rate_iops_min += td->o.rate_iops_min[ddir];
144 * if rate blocks is set, sample is running
146 if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
147 spent = mtime_since(&td->lastrate[ddir], now);
148 if (spent < td->o.ratecycle)
151 if (td->o.rate[ddir]) {
153 * check bandwidth specified rate
155 if (bytes < td->rate_bytes[ddir]) {
156 log_err("%s: min rate %u not met\n", td->o.name,
160 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
161 if (rate < ratemin ||
162 bytes < td->rate_bytes[ddir]) {
163 log_err("%s: min rate %u not met, got"
164 " %luKB/sec\n", td->o.name,
171 * checks iops specified rate
173 if (iops < rate_iops) {
174 log_err("%s: min iops rate %u not met\n",
175 td->o.name, rate_iops);
178 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
179 if (rate < rate_iops_min ||
180 iops < td->rate_blocks[ddir]) {
181 log_err("%s: min iops rate %u not met,"
182 " got %lu\n", td->o.name,
183 rate_iops_min, rate);
189 td->rate_bytes[ddir] = bytes;
190 td->rate_blocks[ddir] = iops;
191 memcpy(&td->lastrate[ddir], now, sizeof(*now));
195 static int check_min_rate(struct thread_data *td, struct timeval *now,
196 unsigned long *bytes_done)
201 ret |= __check_min_rate(td, now, 0);
203 ret |= __check_min_rate(td, now, 1);
209 * When job exits, we can cancel the in-flight IO if we are using async
210 * io. Attempt to do so.
212 static void cleanup_pending_aio(struct thread_data *td)
214 struct flist_head *entry, *n;
219 * get immediately available events, if any
221 r = io_u_queued_complete(td, 0, NULL);
226 * now cancel remaining active events
228 if (td->io_ops->cancel) {
229 flist_for_each_safe(entry, n, &td->io_u_busylist) {
230 io_u = flist_entry(entry, struct io_u, list);
233 * if the io_u isn't in flight, then that generally
234 * means someone leaked an io_u. complain but fix
235 * it up, so we don't stall here.
237 if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
238 log_err("fio: non-busy IO on busy list\n");
241 r = td->io_ops->cancel(td, io_u);
249 r = io_u_queued_complete(td, td->cur_depth, NULL);
253 * Helper to handle the final sync of a file. Works just like the normal
254 * io path, just does everything sync.
256 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
258 struct io_u *io_u = __get_io_u(td);
264 io_u->ddir = DDIR_SYNC;
267 if (td_io_prep(td, io_u)) {
273 ret = td_io_queue(td, io_u);
275 td_verror(td, io_u->error, "td_io_queue");
278 } else if (ret == FIO_Q_QUEUED) {
279 if (io_u_queued_complete(td, 1, NULL) < 0)
281 } else if (ret == FIO_Q_COMPLETED) {
283 td_verror(td, io_u->error, "td_io_queue");
287 if (io_u_sync_complete(td, io_u, NULL) < 0)
289 } else if (ret == FIO_Q_BUSY) {
290 if (td_io_commit(td))
298 static inline void __update_tv_cache(struct thread_data *td)
300 fio_gettime(&td->tv_cache, NULL);
303 static inline void update_tv_cache(struct thread_data *td)
305 if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
306 __update_tv_cache(td);
309 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
311 if (in_ramp_time(td))
315 if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
321 static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
326 if (ret < 0 || td->error) {
334 if (!(td->o.continue_on_error & td_error_type(ddir, err)))
337 if (td_non_fatal_error(err)) {
339 * Continue with the I/Os in case of
342 update_error_count(td, err);
346 } else if (td->o.fill_device && err == ENOSPC) {
348 * We expect to hit this error if
349 * fill_device option is set.
356 * Stop the I/O in case of a fatal
359 update_error_count(td, err);
368 * The main verify engine. Runs over the writes we previously submitted,
369 * reads the blocks back in, and checks the crc/md5 of the data.
371 static void do_verify(struct thread_data *td)
378 dprint(FD_VERIFY, "starting loop\n");
381 * sync io first and invalidate cache, to make sure we really
384 for_each_file(td, f, i) {
385 if (!fio_file_open(f))
387 if (fio_io_sync(td, f))
389 if (file_invalidate_cache(td, f))
396 td_set_runstate(td, TD_VERIFYING);
399 while (!td->terminate) {
404 if (runtime_exceeded(td, &td->tv_cache)) {
405 __update_tv_cache(td);
406 if (runtime_exceeded(td, &td->tv_cache)) {
412 if (flow_threshold_exceeded(td))
415 io_u = __get_io_u(td);
419 if (get_next_verify(td, io_u)) {
424 if (td_io_prep(td, io_u)) {
429 if (td->o.verify_async)
430 io_u->end_io = verify_io_u_async;
432 io_u->end_io = verify_io_u;
434 ret = td_io_queue(td, io_u);
436 case FIO_Q_COMPLETED:
439 clear_io_u(td, io_u);
440 } else if (io_u->resid) {
441 int bytes = io_u->xfer_buflen - io_u->resid;
447 td_verror(td, EIO, "full resid");
452 io_u->xfer_buflen = io_u->resid;
453 io_u->xfer_buf += bytes;
454 io_u->offset += bytes;
456 if (ddir_rw(io_u->ddir))
457 td->ts.short_io_u[io_u->ddir]++;
460 if (io_u->offset == f->real_file_size)
463 requeue_io_u(td, &io_u);
466 ret = io_u_sync_complete(td, io_u, NULL);
474 requeue_io_u(td, &io_u);
475 ret2 = td_io_commit(td);
481 td_verror(td, -ret, "td_io_queue");
485 if (break_on_this_error(td, io_u->ddir, &ret))
489 * if we can queue more, do so. but check if there are
490 * completed io_u's first. Note that we can get BUSY even
491 * without IO queued, if the system is resource starved.
493 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
494 if (full || !td->o.iodepth_batch_complete) {
495 min_events = min(td->o.iodepth_batch_complete,
497 if (full && !min_events && td->o.iodepth_batch_complete != 0)
502 * Reap required number of io units, if any,
503 * and do the verification on them through
504 * the callback handler
506 if (io_u_queued_complete(td, min_events, NULL) < 0) {
510 } while (full && (td->cur_depth > td->o.iodepth_low));
517 min_events = td->cur_depth;
520 ret = io_u_queued_complete(td, min_events, NULL);
522 cleanup_pending_aio(td);
524 td_set_runstate(td, TD_RUNNING);
526 dprint(FD_VERIFY, "exiting loop\n");
529 static int io_bytes_exceeded(struct thread_data *td)
531 unsigned long long bytes;
534 bytes = td->this_io_bytes[0] + td->this_io_bytes[1];
535 else if (td_write(td))
536 bytes = td->this_io_bytes[1];
538 bytes = td->this_io_bytes[0];
540 return bytes >= td->o.size;
544 * Main IO worker function. It retrieves io_u's to process and queues
545 * and reaps them, checking for rate and errors along the way.
547 static void do_io(struct thread_data *td)
552 if (in_ramp_time(td))
553 td_set_runstate(td, TD_RAMP);
555 td_set_runstate(td, TD_RUNNING);
557 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
558 (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td)) {
559 struct timeval comp_time;
560 unsigned long bytes_done[2] = { 0, 0 };
571 if (runtime_exceeded(td, &td->tv_cache)) {
572 __update_tv_cache(td);
573 if (runtime_exceeded(td, &td->tv_cache)) {
579 if (flow_threshold_exceeded(td))
589 * Add verification end_io handler if:
590 * - Asked to verify (!td_rw(td))
591 * - Or the io_u is from our verify list (mixed write/ver)
593 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
594 ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
595 if (td->o.verify_async)
596 io_u->end_io = verify_io_u_async;
598 io_u->end_io = verify_io_u;
599 td_set_runstate(td, TD_VERIFYING);
600 } else if (in_ramp_time(td))
601 td_set_runstate(td, TD_RAMP);
603 td_set_runstate(td, TD_RUNNING);
605 ret = td_io_queue(td, io_u);
607 case FIO_Q_COMPLETED:
610 clear_io_u(td, io_u);
611 } else if (io_u->resid) {
612 int bytes = io_u->xfer_buflen - io_u->resid;
613 struct fio_file *f = io_u->file;
619 td_verror(td, EIO, "full resid");
624 io_u->xfer_buflen = io_u->resid;
625 io_u->xfer_buf += bytes;
626 io_u->offset += bytes;
628 if (ddir_rw(io_u->ddir))
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, ddir, &ret))
671 * See if we need to complete some commands. Note that we
672 * can get BUSY even without IO queued, if the system is
675 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
676 if (full || !td->o.iodepth_batch_complete) {
677 min_evts = min(td->o.iodepth_batch_complete,
679 if (full && !min_evts && td->o.iodepth_batch_complete != 0)
682 if (__should_check_rate(td, 0) ||
683 __should_check_rate(td, 1))
684 fio_gettime(&comp_time, NULL);
687 ret = io_u_queued_complete(td, min_evts, bytes_done);
691 } while (full && (td->cur_depth > td->o.iodepth_low));
696 if (!(bytes_done[0] + bytes_done[1]))
699 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
700 if (check_min_rate(td, &comp_time, bytes_done)) {
701 if (exitall_on_terminate)
702 fio_terminate_threads(td->groupid);
703 td_verror(td, EIO, "check_min_rate");
708 if (td->o.thinktime) {
709 unsigned long long b;
711 b = td->io_blocks[0] + td->io_blocks[1];
712 if (!(b % td->o.thinktime_blocks)) {
715 if (td->o.thinktime_spin)
716 usec_spin(td->o.thinktime_spin);
718 left = td->o.thinktime - td->o.thinktime_spin;
720 usec_sleep(td, left);
725 if (td->trim_entries)
726 log_err("fio: %d trim entries leaked?\n", td->trim_entries);
728 if (td->o.fill_device && td->error == ENOSPC) {
737 ret = io_u_queued_complete(td, i, NULL);
738 if (td->o.fill_device && td->error == ENOSPC)
742 if (should_fsync(td) && td->o.end_fsync) {
743 td_set_runstate(td, TD_FSYNCING);
745 for_each_file(td, f, i) {
746 if (!fio_file_open(f))
752 cleanup_pending_aio(td);
755 * stop job if we failed doing any IO
757 if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
761 static void cleanup_io_u(struct thread_data *td)
763 struct flist_head *entry, *n;
766 flist_for_each_safe(entry, n, &td->io_u_freelist) {
767 io_u = flist_entry(entry, struct io_u, list);
769 flist_del(&io_u->list);
770 fio_memfree(io_u, sizeof(*io_u));
776 static int init_io_u(struct thread_data *td)
779 unsigned int max_bs, min_write;
780 int cl_align, i, max_units;
783 max_units = td->o.iodepth;
784 max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
785 min_write = td->o.min_bs[DDIR_WRITE];
786 td->orig_buffer_size = (unsigned long long) max_bs
787 * (unsigned long long) max_units;
789 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
792 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
793 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
796 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
797 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
801 if (allocate_io_mem(td))
804 if (td->o.odirect || td->o.mem_align ||
805 (td->io_ops->flags & FIO_RAWIO))
806 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
810 cl_align = os_cache_line_size();
812 for (i = 0; i < max_units; i++) {
818 ptr = fio_memalign(cl_align, sizeof(*io_u));
820 log_err("fio: unable to allocate aligned memory\n");
825 memset(io_u, 0, sizeof(*io_u));
826 INIT_FLIST_HEAD(&io_u->list);
827 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
829 if (!(td->io_ops->flags & FIO_NOIO)) {
831 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
834 io_u_fill_buffer(td, io_u, min_write, max_bs);
835 if (td_write(td) && td->o.verify_pattern_bytes) {
837 * Fill the buffer with the pattern if we are
838 * going to be doing writes.
840 fill_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
845 io_u->flags = IO_U_F_FREE;
846 flist_add(&io_u->list, &td->io_u_freelist);
853 static int switch_ioscheduler(struct thread_data *td)
855 char tmp[256], tmp2[128];
859 if (td->io_ops->flags & FIO_DISKLESSIO)
862 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
864 f = fopen(tmp, "r+");
866 if (errno == ENOENT) {
867 log_err("fio: os or kernel doesn't support IO scheduler"
871 td_verror(td, errno, "fopen iosched");
878 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
879 if (ferror(f) || ret != 1) {
880 td_verror(td, errno, "fwrite");
888 * Read back and check that the selected scheduler is now the default.
890 ret = fread(tmp, 1, sizeof(tmp), f);
891 if (ferror(f) || ret < 0) {
892 td_verror(td, errno, "fread");
897 sprintf(tmp2, "[%s]", td->o.ioscheduler);
898 if (!strstr(tmp, tmp2)) {
899 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
900 td_verror(td, EINVAL, "iosched_switch");
909 static int keep_running(struct thread_data *td)
911 unsigned long long io_done;
915 if (td->o.time_based)
922 io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
924 if (io_done < td->o.size)
930 static int exec_string(const char *string)
932 int ret, newlen = strlen(string) + 1 + 8;
935 str = malloc(newlen);
936 sprintf(str, "sh -c %s", string);
940 log_err("fio: exec of cmd <%s> failed\n", str);
947 * Entry point for the thread based jobs. The process based jobs end up
948 * here as well, after a little setup.
950 static void *thread_main(void *data)
952 unsigned long long elapsed;
953 struct thread_data *td = data;
954 pthread_condattr_t attr;
957 if (!td->o.use_thread) {
963 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
965 INIT_FLIST_HEAD(&td->io_u_freelist);
966 INIT_FLIST_HEAD(&td->io_u_busylist);
967 INIT_FLIST_HEAD(&td->io_u_requeues);
968 INIT_FLIST_HEAD(&td->io_log_list);
969 INIT_FLIST_HEAD(&td->io_hist_list);
970 INIT_FLIST_HEAD(&td->verify_list);
971 INIT_FLIST_HEAD(&td->trim_list);
972 pthread_mutex_init(&td->io_u_lock, NULL);
973 td->io_hist_tree = RB_ROOT;
975 pthread_condattr_init(&attr);
976 pthread_cond_init(&td->verify_cond, &attr);
977 pthread_cond_init(&td->free_cond, &attr);
979 td_set_runstate(td, TD_INITIALIZED);
980 dprint(FD_MUTEX, "up startup_mutex\n");
981 fio_mutex_up(startup_mutex);
982 dprint(FD_MUTEX, "wait on td->mutex\n");
983 fio_mutex_down(td->mutex);
984 dprint(FD_MUTEX, "done waiting on td->mutex\n");
987 * the ->mutex mutex is now no longer used, close it to avoid
988 * eating a file descriptor
990 fio_mutex_remove(td->mutex);
993 * A new gid requires privilege, so we need to do this before setting
996 if (td->o.gid != -1U && setgid(td->o.gid)) {
997 td_verror(td, errno, "setgid");
1000 if (td->o.uid != -1U && setuid(td->o.uid)) {
1001 td_verror(td, errno, "setuid");
1006 * If we have a gettimeofday() thread, make sure we exclude that
1007 * thread from this job
1010 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1013 * Set affinity first, in case it has an impact on the memory
1016 if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1017 td_verror(td, errno, "cpu_set_affinity");
1022 * May alter parameters that init_io_u() will use, so we need to
1031 if (td->o.verify_async && verify_async_init(td))
1034 if (td->ioprio_set) {
1035 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1036 td_verror(td, errno, "ioprio_set");
1041 if (td->o.cgroup_weight && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1045 if (nice(td->o.nice) == -1 && errno != 0) {
1046 td_verror(td, errno, "nice");
1050 if (td->o.ioscheduler && switch_ioscheduler(td))
1053 if (!td->o.create_serialize && setup_files(td))
1059 if (init_random_map(td))
1062 if (td->o.exec_prerun) {
1063 if (exec_string(td->o.exec_prerun))
1067 if (td->o.pre_read) {
1068 if (pre_read_files(td) < 0)
1072 fio_gettime(&td->epoch, NULL);
1073 getrusage(RUSAGE_SELF, &td->ru_start);
1076 while (keep_running(td)) {
1077 fio_gettime(&td->start, NULL);
1078 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1079 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1080 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1082 if (td->o.ratemin[0] || td->o.ratemin[1]) {
1083 memcpy(&td->lastrate[0], &td->bw_sample_time,
1084 sizeof(td->bw_sample_time));
1085 memcpy(&td->lastrate[1], &td->bw_sample_time,
1086 sizeof(td->bw_sample_time));
1092 prune_io_piece_log(td);
1098 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1099 elapsed = utime_since_now(&td->start);
1100 td->ts.runtime[DDIR_READ] += elapsed;
1102 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1103 elapsed = utime_since_now(&td->start);
1104 td->ts.runtime[DDIR_WRITE] += elapsed;
1107 if (td->error || td->terminate)
1110 if (!td->o.do_verify ||
1111 td->o.verify == VERIFY_NONE ||
1112 (td->io_ops->flags & FIO_UNIDIR))
1117 fio_gettime(&td->start, NULL);
1121 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1123 if (td->error || td->terminate)
1127 update_rusage_stat(td);
1128 td->ts.runtime[0] = (td->ts.runtime[0] + 999) / 1000;
1129 td->ts.runtime[1] = (td->ts.runtime[1] + 999) / 1000;
1130 td->ts.total_run_time = mtime_since_now(&td->epoch);
1131 td->ts.io_bytes[0] = td->io_bytes[0];
1132 td->ts.io_bytes[1] = td->io_bytes[1];
1134 fio_mutex_down(writeout_mutex);
1136 if (td->o.bw_log_file) {
1137 finish_log_named(td, td->bw_log,
1138 td->o.bw_log_file, "bw");
1140 finish_log(td, td->bw_log, "bw");
1143 if (td->o.lat_log_file) {
1144 finish_log_named(td, td->lat_log,
1145 td->o.lat_log_file, "lat");
1147 finish_log(td, td->lat_log, "lat");
1150 if (td->o.lat_log_file) {
1151 finish_log_named(td, td->slat_log,
1152 td->o.lat_log_file, "slat");
1154 finish_log(td, td->slat_log, "slat");
1157 if (td->o.lat_log_file) {
1158 finish_log_named(td, td->clat_log,
1159 td->o.lat_log_file, "clat");
1161 finish_log(td, td->clat_log, "clat");
1164 if (td->o.iops_log_file) {
1165 finish_log_named(td, td->iops_log,
1166 td->o.iops_log_file, "iops");
1168 finish_log(td, td->iops_log, "iops");
1171 fio_mutex_up(writeout_mutex);
1172 if (td->o.exec_postrun)
1173 exec_string(td->o.exec_postrun);
1175 if (exitall_on_terminate)
1176 fio_terminate_threads(td->groupid);
1180 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1183 if (td->o.verify_async)
1184 verify_async_exit(td);
1186 close_and_free_files(td);
1189 cgroup_shutdown(td, &cgroup_mnt);
1191 if (td->o.cpumask_set) {
1192 int ret = fio_cpuset_exit(&td->o.cpumask);
1194 td_verror(td, ret, "fio_cpuset_exit");
1198 * do this very late, it will log file closing as well
1200 if (td->o.write_iolog_file)
1201 write_iolog_close(td);
1203 td_set_runstate(td, TD_EXITED);
1204 return (void *) (uintptr_t) td->error;
1209 * We cannot pass the td data into a forked process, so attach the td and
1210 * pass it to the thread worker.
1212 static int fork_main(int shmid, int offset)
1214 struct thread_data *td;
1218 data = shmat(shmid, NULL, 0);
1219 if (data == (void *) -1) {
1227 * HP-UX inherits shm mappings?
1232 td = data + offset * sizeof(struct thread_data);
1233 ret = thread_main(td);
1235 return (int) (uintptr_t) ret;
1239 * Run over the job map and reap the threads that have exited, if any.
1241 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1242 unsigned int *m_rate)
1244 struct thread_data *td;
1245 unsigned int cputhreads, realthreads, pending;
1249 * reap exited threads (TD_EXITED -> TD_REAPED)
1251 realthreads = pending = cputhreads = 0;
1252 for_each_td(td, i) {
1256 * ->io_ops is NULL for a thread that has closed its
1259 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1268 if (td->runstate == TD_REAPED)
1270 if (td->o.use_thread) {
1271 if (td->runstate == TD_EXITED) {
1272 td_set_runstate(td, TD_REAPED);
1279 if (td->runstate == TD_EXITED)
1283 * check if someone quit or got killed in an unusual way
1285 ret = waitpid(td->pid, &status, flags);
1287 if (errno == ECHILD) {
1288 log_err("fio: pid=%d disappeared %d\n",
1289 (int) td->pid, td->runstate);
1290 td_set_runstate(td, TD_REAPED);
1294 } else if (ret == td->pid) {
1295 if (WIFSIGNALED(status)) {
1296 int sig = WTERMSIG(status);
1299 log_err("fio: pid=%d, got signal=%d\n",
1300 (int) td->pid, sig);
1301 td_set_runstate(td, TD_REAPED);
1304 if (WIFEXITED(status)) {
1305 if (WEXITSTATUS(status) && !td->error)
1306 td->error = WEXITSTATUS(status);
1308 td_set_runstate(td, TD_REAPED);
1314 * thread is not dead, continue
1320 (*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
1321 (*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
1328 done_secs += mtime_since_now(&td->epoch) / 1000;
1331 if (*nr_running == cputhreads && !pending && realthreads)
1332 fio_terminate_threads(TERMINATE_ALL);
1336 * Main function for kicking off and reaping jobs, as needed.
1338 static void run_threads(void)
1340 struct thread_data *td;
1341 unsigned long spent;
1342 unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
1344 if (fio_pin_memory())
1347 if (fio_gtod_offload && fio_start_gtod_thread())
1352 if (!terse_output) {
1353 log_info("Starting ");
1355 log_info("%d thread%s", nr_thread,
1356 nr_thread > 1 ? "s" : "");
1360 log_info("%d process%s", nr_process,
1361 nr_process > 1 ? "es" : "");
1367 todo = thread_number;
1370 m_rate = t_rate = 0;
1372 for_each_td(td, i) {
1373 print_status_init(td->thread_number - 1);
1375 if (!td->o.create_serialize)
1379 * do file setup here so it happens sequentially,
1380 * we don't want X number of threads getting their
1381 * client data interspersed on disk
1383 if (setup_files(td)) {
1386 log_err("fio: pid=%d, err=%d/%s\n",
1387 (int) td->pid, td->error, td->verror);
1388 td_set_runstate(td, TD_REAPED);
1395 * for sharing to work, each job must always open
1396 * its own files. so close them, if we opened them
1399 for_each_file(td, f, j) {
1400 if (fio_file_open(f))
1401 td_io_close_file(td, f);
1409 struct thread_data *map[REAL_MAX_JOBS];
1410 struct timeval this_start;
1411 int this_jobs = 0, left;
1414 * create threads (TD_NOT_CREATED -> TD_CREATED)
1416 for_each_td(td, i) {
1417 if (td->runstate != TD_NOT_CREATED)
1421 * never got a chance to start, killed by other
1422 * thread for some reason
1424 if (td->terminate) {
1429 if (td->o.start_delay) {
1430 spent = mtime_since_genesis();
1432 if (td->o.start_delay * 1000 > spent)
1436 if (td->o.stonewall && (nr_started || nr_running)) {
1437 dprint(FD_PROCESS, "%s: stonewall wait\n",
1445 * Set state to created. Thread will transition
1446 * to TD_INITIALIZED when it's done setting up.
1448 td_set_runstate(td, TD_CREATED);
1449 map[this_jobs++] = td;
1452 if (td->o.use_thread) {
1455 dprint(FD_PROCESS, "will pthread_create\n");
1456 ret = pthread_create(&td->thread, NULL,
1459 log_err("pthread_create: %s\n",
1464 ret = pthread_detach(td->thread);
1466 log_err("pthread_detach: %s",
1470 dprint(FD_PROCESS, "will fork\n");
1473 int ret = fork_main(shm_id, i);
1476 } else if (i == fio_debug_jobno)
1477 *fio_debug_jobp = pid;
1479 dprint(FD_MUTEX, "wait on startup_mutex\n");
1480 if (fio_mutex_down_timeout(startup_mutex, 10)) {
1481 log_err("fio: job startup hung? exiting.\n");
1482 fio_terminate_threads(TERMINATE_ALL);
1487 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1491 * Wait for the started threads to transition to
1494 fio_gettime(&this_start, NULL);
1496 while (left && !fio_abort) {
1497 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1502 for (i = 0; i < this_jobs; i++) {
1506 if (td->runstate == TD_INITIALIZED) {
1509 } else if (td->runstate >= TD_EXITED) {
1513 nr_running++; /* work-around... */
1519 log_err("fio: %d job%s failed to start\n", left,
1520 left > 1 ? "s" : "");
1521 for (i = 0; i < this_jobs; i++) {
1525 kill(td->pid, SIGTERM);
1531 * start created threads (TD_INITIALIZED -> TD_RUNNING).
1533 for_each_td(td, i) {
1534 if (td->runstate != TD_INITIALIZED)
1537 if (in_ramp_time(td))
1538 td_set_runstate(td, TD_RAMP);
1540 td_set_runstate(td, TD_RUNNING);
1543 m_rate += td->o.ratemin[0] + td->o.ratemin[1];
1544 t_rate += td->o.rate[0] + td->o.rate[1];
1546 fio_mutex_up(td->mutex);
1549 reap_threads(&nr_running, &t_rate, &m_rate);
1553 fio_server_idle_loop();
1559 while (nr_running) {
1560 reap_threads(&nr_running, &t_rate, &m_rate);
1563 fio_server_idle_loop();
1572 static void *disk_thread_main(void *data)
1574 fio_mutex_up(startup_mutex);
1577 usleep(DISK_UTIL_MSEC * 1000);
1583 print_thread_status();
1589 static int create_disk_util_thread(void)
1593 ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
1595 log_err("Can't create disk util thread: %s\n", strerror(ret));
1599 ret = pthread_detach(disk_util_thread);
1601 log_err("Can't detatch disk util thread: %s\n", strerror(ret));
1605 dprint(FD_MUTEX, "wait on startup_mutex\n");
1606 fio_mutex_down(startup_mutex);
1607 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1611 int fio_backend(void)
1613 struct thread_data *td;
1617 if (load_profile(exec_profile))
1620 exec_profile = NULL;
1626 setup_log(&agg_io_log[DDIR_READ], 0, IO_LOG_TYPE_BW);
1627 setup_log(&agg_io_log[DDIR_WRITE], 0, IO_LOG_TYPE_BW);
1630 startup_mutex = fio_mutex_init(0);
1631 if (startup_mutex == NULL)
1633 writeout_mutex = fio_mutex_init(1);
1634 if (writeout_mutex == NULL)
1638 create_disk_util_thread();
1640 cgroup_list = smalloc(sizeof(*cgroup_list));
1641 INIT_FLIST_HEAD(cgroup_list);
1648 __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1649 __finish_log(agg_io_log[DDIR_WRITE],
1650 "agg-write_bw.log");
1655 fio_options_free(td);
1657 cgroup_kill(cgroup_list);
1661 fio_mutex_remove(startup_mutex);
1662 fio_mutex_remove(writeout_mutex);