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 (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
49 int thread_number = 0;
54 unsigned long done_secs = 0;
56 static struct fio_mutex *startup_mutex;
57 static struct fio_mutex *writeout_mutex;
58 static volatile int fio_abort;
59 static int exit_value;
60 static struct itimerval itimer;
61 static pthread_t gtod_thread;
63 struct io_log *agg_io_log[2];
65 #define TERMINATE_ALL (-1)
66 #define JOB_START_TIMEOUT (5 * 1000)
68 void td_set_runstate(struct thread_data *td, int runstate)
70 if (td->runstate == runstate)
73 dprint(FD_PROCESS, "pid=%d: runstate %d -> %d\n", (int) td->pid,
74 td->runstate, runstate);
75 td->runstate = runstate;
78 static void terminate_threads(int group_id)
80 struct thread_data *td;
83 dprint(FD_PROCESS, "terminate group_id=%d\n", group_id);
86 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
87 dprint(FD_PROCESS, "setting terminate on %s/%d\n",
88 td->o.name, (int) td->pid);
90 td->o.start_delay = 0;
93 * if the thread is running, just let it exit
95 if (td->runstate < TD_RUNNING)
96 kill(td->pid, SIGQUIT);
98 struct ioengine_ops *ops = td->io_ops;
100 if (ops && (ops->flags & FIO_SIGQUIT))
101 kill(td->pid, SIGQUIT);
107 static void status_timer_arm(void)
109 itimer.it_value.tv_sec = 0;
110 itimer.it_value.tv_usec = DISK_UTIL_MSEC * 1000;
111 setitimer(ITIMER_REAL, &itimer, NULL);
114 static void sig_alrm(int fio_unused sig)
118 print_thread_status();
124 * Happens on thread runs with ctrl-c, ignore our own SIGQUIT
126 static void sig_quit(int sig)
130 static void sig_int(int sig)
133 printf("\nfio: terminating on signal %d\n", sig);
135 terminate_threads(TERMINATE_ALL);
139 static void sig_ill(int fio_unused sig)
144 log_err("fio: illegal instruction. your cpu does not support "
145 "the sse4.2 instruction for crc32c\n");
146 terminate_threads(TERMINATE_ALL);
150 static void set_sig_handlers(void)
152 struct sigaction act;
154 memset(&act, 0, sizeof(act));
155 act.sa_handler = sig_alrm;
156 act.sa_flags = SA_RESTART;
157 sigaction(SIGALRM, &act, NULL);
159 memset(&act, 0, sizeof(act));
160 act.sa_handler = sig_int;
161 act.sa_flags = SA_RESTART;
162 sigaction(SIGINT, &act, NULL);
164 memset(&act, 0, sizeof(act));
165 act.sa_handler = sig_ill;
166 act.sa_flags = SA_RESTART;
167 sigaction(SIGILL, &act, NULL);
169 memset(&act, 0, sizeof(act));
170 act.sa_handler = sig_quit;
171 act.sa_flags = SA_RESTART;
172 sigaction(SIGQUIT, &act, NULL);
176 * Check if we are above the minimum rate given.
178 static int __check_min_rate(struct thread_data *td, struct timeval *now,
181 unsigned long long bytes = 0;
182 unsigned long iops = 0;
185 unsigned int ratemin = 0;
186 unsigned int rate_iops = 0;
187 unsigned int rate_iops_min = 0;
189 if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
193 * allow a 2 second settle period in the beginning
195 if (mtime_since(&td->start, now) < 2000)
198 iops += td->io_blocks[ddir];
199 bytes += td->this_io_bytes[ddir];
200 ratemin += td->o.ratemin[ddir];
201 rate_iops += td->o.rate_iops[ddir];
202 rate_iops_min += td->o.rate_iops_min[ddir];
205 * if rate blocks is set, sample is running
207 if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
208 spent = mtime_since(&td->lastrate[ddir], now);
209 if (spent < td->o.ratecycle)
212 if (td->o.rate[ddir]) {
214 * check bandwidth specified rate
216 if (bytes < td->rate_bytes[ddir]) {
217 log_err("%s: min rate %u not met\n", td->o.name,
221 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
222 if (rate < ratemin ||
223 bytes < td->rate_bytes[ddir]) {
224 log_err("%s: min rate %u not met, got"
225 " %luKiB/sec\n", td->o.name,
232 * checks iops specified rate
234 if (iops < rate_iops) {
235 log_err("%s: min iops rate %u not met\n",
236 td->o.name, rate_iops);
239 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
240 if (rate < rate_iops_min ||
241 iops < td->rate_blocks[ddir]) {
242 log_err("%s: min iops rate %u not met,"
243 " got %lu\n", td->o.name,
244 rate_iops_min, rate);
250 td->rate_bytes[ddir] = bytes;
251 td->rate_blocks[ddir] = iops;
252 memcpy(&td->lastrate[ddir], now, sizeof(*now));
256 static int check_min_rate(struct thread_data *td, struct timeval *now,
257 unsigned long *bytes_done)
262 ret |= __check_min_rate(td, now, 0);
264 ret |= __check_min_rate(td, now, 1);
269 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
273 if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
280 * When job exits, we can cancel the in-flight IO if we are using async
281 * io. Attempt to do so.
283 static void cleanup_pending_aio(struct thread_data *td)
285 struct flist_head *entry, *n;
290 * get immediately available events, if any
292 r = io_u_queued_complete(td, 0, NULL);
297 * now cancel remaining active events
299 if (td->io_ops->cancel) {
300 flist_for_each_safe(entry, n, &td->io_u_busylist) {
301 io_u = flist_entry(entry, struct io_u, list);
304 * if the io_u isn't in flight, then that generally
305 * means someone leaked an io_u. complain but fix
306 * it up, so we don't stall here.
308 if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
309 log_err("fio: non-busy IO on busy list\n");
312 r = td->io_ops->cancel(td, io_u);
320 r = io_u_queued_complete(td, td->cur_depth, NULL);
324 * Helper to handle the final sync of a file. Works just like the normal
325 * io path, just does everything sync.
327 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
329 struct io_u *io_u = __get_io_u(td);
335 io_u->ddir = DDIR_SYNC;
338 if (td_io_prep(td, io_u)) {
344 ret = td_io_queue(td, io_u);
346 td_verror(td, io_u->error, "td_io_queue");
349 } else if (ret == FIO_Q_QUEUED) {
350 if (io_u_queued_complete(td, 1, NULL) < 0)
352 } else if (ret == FIO_Q_COMPLETED) {
354 td_verror(td, io_u->error, "td_io_queue");
358 if (io_u_sync_complete(td, io_u, NULL) < 0)
360 } else if (ret == FIO_Q_BUSY) {
361 if (td_io_commit(td))
369 static inline void update_tv_cache(struct thread_data *td)
371 if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
372 fio_gettime(&td->tv_cache, NULL);
375 static int break_on_this_error(struct thread_data *td, int *retptr)
379 if (ret < 0 || td->error) {
382 if (!td->o.continue_on_error);
390 update_error_count(td, err);
392 if (td_non_fatal_error(err)) {
394 * Continue with the I/Os in case of
402 * Stop the I/O in case of a fatal
413 * The main verify engine. Runs over the writes we previously submitted,
414 * reads the blocks back in, and checks the crc/md5 of the data.
416 static void do_verify(struct thread_data *td)
424 * sync io first and invalidate cache, to make sure we really
427 for_each_file(td, f, i) {
428 if (!fio_file_open(f))
430 if (fio_io_sync(td, f))
432 if (file_invalidate_cache(td, f))
439 td_set_runstate(td, TD_VERIFYING);
442 while (!td->terminate) {
445 io_u = __get_io_u(td);
451 if (runtime_exceeded(td, &td->tv_cache)) {
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 };
589 if (runtime_exceeded(td, &td->tv_cache)) {
596 * Add verification end_io handler, if asked to verify
597 * a previously written file.
599 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ) {
600 io_u->end_io = verify_io_u;
601 td_set_runstate(td, TD_VERIFYING);
602 } else if (in_ramp_time(td))
603 td_set_runstate(td, TD_RAMP);
605 td_set_runstate(td, TD_RUNNING);
607 ret = td_io_queue(td, io_u);
609 case FIO_Q_COMPLETED:
612 clear_io_u(td, io_u);
613 } else if (io_u->resid) {
614 int bytes = io_u->xfer_buflen - io_u->resid;
615 struct fio_file *f = io_u->file;
621 td_verror(td, EIO, "full resid");
626 io_u->xfer_buflen = io_u->resid;
627 io_u->xfer_buf += bytes;
628 io_u->offset += bytes;
630 td->ts.short_io_u[io_u->ddir]++;
632 if (io_u->offset == f->real_file_size)
635 requeue_io_u(td, &io_u);
638 if (__should_check_rate(td, 0) ||
639 __should_check_rate(td, 1))
640 fio_gettime(&comp_time, NULL);
642 ret = io_u_sync_complete(td, io_u, bytes_done);
649 * if the engine doesn't have a commit hook,
650 * the io_u is really queued. if it does have such
651 * a hook, it has to call io_u_queued() itself.
653 if (td->io_ops->commit == NULL)
654 io_u_queued(td, io_u);
657 requeue_io_u(td, &io_u);
658 ret2 = td_io_commit(td);
668 if (break_on_this_error(td, &ret))
672 * See if we need to complete some commands
674 full = queue_full(td) || ret == FIO_Q_BUSY;
675 if (full || !td->o.iodepth_batch_complete) {
676 min_evts = td->o.iodepth_batch_complete;
677 if (full && !min_evts)
680 if (__should_check_rate(td, 0) ||
681 __should_check_rate(td, 1))
682 fio_gettime(&comp_time, NULL);
685 ret = io_u_queued_complete(td, min_evts, bytes_done);
689 } while (full && (td->cur_depth > td->o.iodepth_low));
694 if (!(bytes_done[0] + bytes_done[1]))
697 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
698 if (check_min_rate(td, &comp_time, bytes_done)) {
699 if (exitall_on_terminate)
700 terminate_threads(td->groupid);
701 td_verror(td, EIO, "check_min_rate");
706 if (td->o.thinktime) {
707 unsigned long long b;
709 b = td->io_blocks[0] + td->io_blocks[1];
710 if (!(b % td->o.thinktime_blocks)) {
713 if (td->o.thinktime_spin)
714 usec_spin(td->o.thinktime_spin);
716 left = td->o.thinktime - td->o.thinktime_spin;
718 usec_sleep(td, left);
723 if (td->o.fill_device && td->error == ENOSPC) {
732 ret = io_u_queued_complete(td, i, NULL);
734 if (should_fsync(td) && td->o.end_fsync) {
735 td_set_runstate(td, TD_FSYNCING);
737 for_each_file(td, f, i) {
738 if (!fio_file_open(f))
744 cleanup_pending_aio(td);
747 * stop job if we failed doing any IO
749 if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
753 static void cleanup_io_u(struct thread_data *td)
755 struct flist_head *entry, *n;
758 flist_for_each_safe(entry, n, &td->io_u_freelist) {
759 io_u = flist_entry(entry, struct io_u, list);
761 flist_del(&io_u->list);
768 static int init_io_u(struct thread_data *td)
772 int cl_align, i, max_units;
775 max_units = td->o.iodepth;
776 max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
777 td->orig_buffer_size = (unsigned long long) max_bs
778 * (unsigned long long) max_units;
780 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
783 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
784 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
787 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
788 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
792 if (allocate_io_mem(td))
796 p = ALIGN(td->orig_buffer);
800 cl_align = os_cache_line_size();
802 for (i = 0; i < max_units; i++) {
808 if (posix_memalign(&ptr, cl_align, sizeof(*io_u))) {
809 log_err("fio: posix_memalign=%s\n", strerror(errno));
814 memset(io_u, 0, sizeof(*io_u));
815 INIT_FLIST_HEAD(&io_u->list);
817 if (!(td->io_ops->flags & FIO_NOIO)) {
818 io_u->buf = p + max_bs * i;
820 if (td_write(td) && !td->o.refill_buffers)
821 io_u_fill_buffer(td, io_u, max_bs);
825 io_u->flags = IO_U_F_FREE;
826 flist_add(&io_u->list, &td->io_u_freelist);
832 static int switch_ioscheduler(struct thread_data *td)
834 char tmp[256], tmp2[128];
838 if (td->io_ops->flags & FIO_DISKLESSIO)
841 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
843 f = fopen(tmp, "r+");
845 if (errno == ENOENT) {
846 log_err("fio: os or kernel doesn't support IO scheduler"
850 td_verror(td, errno, "fopen iosched");
857 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
858 if (ferror(f) || ret != 1) {
859 td_verror(td, errno, "fwrite");
867 * Read back and check that the selected scheduler is now the default.
869 ret = fread(tmp, 1, sizeof(tmp), f);
870 if (ferror(f) || ret < 0) {
871 td_verror(td, errno, "fread");
876 sprintf(tmp2, "[%s]", td->o.ioscheduler);
877 if (!strstr(tmp, tmp2)) {
878 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
879 td_verror(td, EINVAL, "iosched_switch");
888 static int keep_running(struct thread_data *td)
890 unsigned long long io_done;
894 if (td->o.time_based)
901 io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
903 if (io_done < td->o.size)
909 static void reset_io_counters(struct thread_data *td)
911 td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
912 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
914 td->rate_bytes[0] = td->rate_bytes[1] = 0;
915 td->rate_blocks[0] = td->rate_blocks[1] = 0;
917 td->last_was_sync = 0;
920 * reset file done count if we are to start over
922 if (td->o.time_based || td->o.loops)
923 td->nr_done_files = 0;
926 * Set the same seed to get repeatable runs
928 td_fill_rand_seeds(td);
931 void reset_all_stats(struct thread_data *td)
936 reset_io_counters(td);
938 for (i = 0; i < 2; i++) {
940 td->io_blocks[i] = 0;
941 td->io_issues[i] = 0;
942 td->ts.total_io_u[i] = 0;
945 fio_gettime(&tv, NULL);
946 memcpy(&td->epoch, &tv, sizeof(tv));
947 memcpy(&td->start, &tv, sizeof(tv));
950 static void clear_io_state(struct thread_data *td)
955 reset_io_counters(td);
958 for_each_file(td, f, i)
959 fio_file_clear_done(f);
962 static int exec_string(const char *string)
964 int ret, newlen = strlen(string) + 1 + 8;
967 str = malloc(newlen);
968 sprintf(str, "sh -c %s", string);
972 log_err("fio: exec of cmd <%s> failed\n", str);
979 * Entry point for the thread based jobs. The process based jobs end up
980 * here as well, after a little setup.
982 static void *thread_main(void *data)
984 unsigned long long runtime[2], elapsed;
985 struct thread_data *td = data;
988 if (!td->o.use_thread)
993 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
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 td->io_hist_tree = RB_ROOT;
1002 td_set_runstate(td, TD_INITIALIZED);
1003 dprint(FD_MUTEX, "up startup_mutex\n");
1004 fio_mutex_up(startup_mutex);
1005 dprint(FD_MUTEX, "wait on td->mutex\n");
1006 fio_mutex_down(td->mutex);
1007 dprint(FD_MUTEX, "done waiting on td->mutex\n");
1010 * the ->mutex mutex is now no longer used, close it to avoid
1011 * eating a file descriptor
1013 fio_mutex_remove(td->mutex);
1016 * May alter parameters that init_io_u() will use, so we need to
1025 if (td->o.cpumask_set && fio_setaffinity(td) == -1) {
1026 td_verror(td, errno, "cpu_set_affinity");
1031 * If we have a gettimeofday() thread, make sure we exclude that
1032 * thread from this job
1034 if (td->o.gtod_cpu) {
1035 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1036 if (fio_setaffinity(td) == -1) {
1037 td_verror(td, errno, "cpu_set_affinity");
1042 if (td->ioprio_set) {
1043 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1044 td_verror(td, errno, "ioprio_set");
1049 if (nice(td->o.nice) == -1) {
1050 td_verror(td, errno, "nice");
1054 if (td->o.ioscheduler && switch_ioscheduler(td))
1057 if (!td->o.create_serialize && setup_files(td))
1063 if (init_random_map(td))
1066 if (td->o.exec_prerun) {
1067 if (exec_string(td->o.exec_prerun))
1071 if (td->o.pre_read) {
1072 if (pre_read_files(td) < 0)
1076 fio_gettime(&td->epoch, NULL);
1077 getrusage(RUSAGE_SELF, &td->ts.ru_start);
1079 runtime[0] = runtime[1] = 0;
1081 while (keep_running(td)) {
1082 fio_gettime(&td->start, NULL);
1083 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
1084 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1086 if (td->o.ratemin[0] || td->o.ratemin[1])
1087 memcpy(&td->lastrate, &td->ts.stat_sample_time,
1088 sizeof(td->lastrate));
1093 prune_io_piece_log(td);
1099 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1100 elapsed = utime_since_now(&td->start);
1101 runtime[DDIR_READ] += elapsed;
1103 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1104 elapsed = utime_since_now(&td->start);
1105 runtime[DDIR_WRITE] += elapsed;
1108 if (td->error || td->terminate)
1111 if (!td->o.do_verify ||
1112 td->o.verify == VERIFY_NONE ||
1113 (td->io_ops->flags & FIO_UNIDIR))
1118 fio_gettime(&td->start, NULL);
1122 runtime[DDIR_READ] += utime_since_now(&td->start);
1124 if (td->error || td->terminate)
1128 update_rusage_stat(td);
1129 td->ts.runtime[0] = (runtime[0] + 999) / 1000;
1130 td->ts.runtime[1] = (runtime[1] + 999) / 1000;
1131 td->ts.total_run_time = mtime_since_now(&td->epoch);
1132 td->ts.io_bytes[0] = td->io_bytes[0];
1133 td->ts.io_bytes[1] = td->io_bytes[1];
1135 fio_mutex_down(writeout_mutex);
1136 if (td->ts.bw_log) {
1137 if (td->o.bw_log_file) {
1138 finish_log_named(td, td->ts.bw_log,
1139 td->o.bw_log_file, "bw");
1141 finish_log(td, td->ts.bw_log, "bw");
1143 if (td->ts.slat_log) {
1144 if (td->o.lat_log_file) {
1145 finish_log_named(td, td->ts.slat_log,
1146 td->o.lat_log_file, "slat");
1148 finish_log(td, td->ts.slat_log, "slat");
1150 if (td->ts.clat_log) {
1151 if (td->o.lat_log_file) {
1152 finish_log_named(td, td->ts.clat_log,
1153 td->o.lat_log_file, "clat");
1155 finish_log(td, td->ts.clat_log, "clat");
1157 fio_mutex_up(writeout_mutex);
1158 if (td->o.exec_postrun)
1159 exec_string(td->o.exec_postrun);
1161 if (exitall_on_terminate)
1162 terminate_threads(td->groupid);
1166 printf("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1168 close_and_free_files(td);
1172 if (td->o.cpumask_set) {
1173 int ret = fio_cpuset_exit(&td->o.cpumask);
1175 td_verror(td, ret, "fio_cpuset_exit");
1179 * do this very late, it will log file closing as well
1181 if (td->o.write_iolog_file)
1182 write_iolog_close(td);
1184 options_mem_free(td);
1185 td_set_runstate(td, TD_EXITED);
1186 return (void *) (unsigned long) td->error;
1190 * We cannot pass the td data into a forked process, so attach the td and
1191 * pass it to the thread worker.
1193 static int fork_main(int shmid, int offset)
1195 struct thread_data *td;
1198 data = shmat(shmid, NULL, 0);
1199 if (data == (void *) -1) {
1206 td = data + offset * sizeof(struct thread_data);
1207 ret = thread_main(td);
1209 return (int) (unsigned long) ret;
1213 * Run over the job map and reap the threads that have exited, if any.
1215 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1217 struct thread_data *td;
1218 int i, cputhreads, realthreads, pending, status, ret;
1221 * reap exited threads (TD_EXITED -> TD_REAPED)
1223 realthreads = pending = cputhreads = 0;
1224 for_each_td(td, i) {
1228 * ->io_ops is NULL for a thread that has closed its
1231 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1240 if (td->runstate == TD_REAPED)
1242 if (td->o.use_thread) {
1243 if (td->runstate == TD_EXITED) {
1244 td_set_runstate(td, TD_REAPED);
1251 if (td->runstate == TD_EXITED)
1255 * check if someone quit or got killed in an unusual way
1257 ret = waitpid(td->pid, &status, flags);
1259 if (errno == ECHILD) {
1260 log_err("fio: pid=%d disappeared %d\n",
1261 (int) td->pid, td->runstate);
1262 td_set_runstate(td, TD_REAPED);
1266 } else if (ret == td->pid) {
1267 if (WIFSIGNALED(status)) {
1268 int sig = WTERMSIG(status);
1271 log_err("fio: pid=%d, got signal=%d\n",
1272 (int) td->pid, sig);
1273 td_set_runstate(td, TD_REAPED);
1276 if (WIFEXITED(status)) {
1277 if (WEXITSTATUS(status) && !td->error)
1278 td->error = WEXITSTATUS(status);
1280 td_set_runstate(td, TD_REAPED);
1286 * thread is not dead, continue
1292 (*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
1293 (*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
1300 done_secs += mtime_since_now(&td->epoch) / 1000;
1303 if (*nr_running == cputhreads && !pending && realthreads)
1304 terminate_threads(TERMINATE_ALL);
1307 static void *gtod_thread_main(void *data)
1309 fio_mutex_up(startup_mutex);
1312 * As long as we have jobs around, update the clock. It would be nice
1313 * to have some way of NOT hammering that CPU with gettimeofday(),
1314 * but I'm not sure what to use outside of a simple CPU nop to relax
1315 * it - we don't want to lose precision.
1325 static int fio_start_gtod_thread(void)
1329 ret = pthread_create(>od_thread, NULL, gtod_thread_main, NULL);
1331 log_err("Can't create gtod thread: %s\n", strerror(ret));
1335 ret = pthread_detach(gtod_thread);
1337 log_err("Can't detatch gtod thread: %s\n", strerror(ret));
1341 dprint(FD_MUTEX, "wait on startup_mutex\n");
1342 fio_mutex_down(startup_mutex);
1343 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1348 * Main function for kicking off and reaping jobs, as needed.
1350 static void run_threads(void)
1352 struct thread_data *td;
1353 unsigned long spent;
1354 int i, todo, nr_running, m_rate, t_rate, nr_started;
1356 if (fio_pin_memory())
1359 if (fio_gtod_offload && fio_start_gtod_thread())
1362 if (!terse_output) {
1363 printf("Starting ");
1365 printf("%d thread%s", nr_thread,
1366 nr_thread > 1 ? "s" : "");
1370 printf("%d process%s", nr_process,
1371 nr_process > 1 ? "es" : "");
1379 todo = thread_number;
1382 m_rate = t_rate = 0;
1384 for_each_td(td, i) {
1385 print_status_init(td->thread_number - 1);
1387 if (!td->o.create_serialize) {
1393 * do file setup here so it happens sequentially,
1394 * we don't want X number of threads getting their
1395 * client data interspersed on disk
1397 if (setup_files(td)) {
1400 log_err("fio: pid=%d, err=%d/%s\n",
1401 (int) td->pid, td->error, td->verror);
1402 td_set_runstate(td, TD_REAPED);
1409 * for sharing to work, each job must always open
1410 * its own files. so close them, if we opened them
1413 for_each_file(td, f, i) {
1414 if (fio_file_open(f))
1415 td_io_close_file(td, f);
1425 struct thread_data *map[MAX_JOBS];
1426 struct timeval this_start;
1427 int this_jobs = 0, left;
1430 * create threads (TD_NOT_CREATED -> TD_CREATED)
1432 for_each_td(td, i) {
1433 if (td->runstate != TD_NOT_CREATED)
1437 * never got a chance to start, killed by other
1438 * thread for some reason
1440 if (td->terminate) {
1445 if (td->o.start_delay) {
1446 spent = mtime_since_genesis();
1448 if (td->o.start_delay * 1000 > spent)
1452 if (td->o.stonewall && (nr_started || nr_running)) {
1453 dprint(FD_PROCESS, "%s: stonewall wait\n",
1459 * Set state to created. Thread will transition
1460 * to TD_INITIALIZED when it's done setting up.
1462 td_set_runstate(td, TD_CREATED);
1463 map[this_jobs++] = td;
1466 if (td->o.use_thread) {
1469 dprint(FD_PROCESS, "will pthread_create\n");
1470 ret = pthread_create(&td->thread, NULL,
1473 log_err("pthread_create: %s\n",
1478 ret = pthread_detach(td->thread);
1480 log_err("pthread_detach: %s",
1484 dprint(FD_PROCESS, "will fork\n");
1487 int ret = fork_main(shm_id, i);
1490 } else if (i == fio_debug_jobno)
1491 *fio_debug_jobp = pid;
1493 dprint(FD_MUTEX, "wait on startup_mutex\n");
1494 fio_mutex_down(startup_mutex);
1495 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1499 * Wait for the started threads to transition to
1502 fio_gettime(&this_start, NULL);
1504 while (left && !fio_abort) {
1505 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1510 for (i = 0; i < this_jobs; i++) {
1514 if (td->runstate == TD_INITIALIZED) {
1517 } else if (td->runstate >= TD_EXITED) {
1521 nr_running++; /* work-around... */
1527 log_err("fio: %d jobs failed to start\n", left);
1528 for (i = 0; i < this_jobs; i++) {
1532 kill(td->pid, SIGTERM);
1538 * start created threads (TD_INITIALIZED -> TD_RUNNING).
1540 for_each_td(td, i) {
1541 if (td->runstate != TD_INITIALIZED)
1544 if (in_ramp_time(td))
1545 td_set_runstate(td, TD_RAMP);
1547 td_set_runstate(td, TD_RUNNING);
1550 m_rate += td->o.ratemin[0] + td->o.ratemin[1];
1551 t_rate += td->o.rate[0] + td->o.rate[1];
1553 fio_mutex_up(td->mutex);
1556 reap_threads(&nr_running, &t_rate, &m_rate);
1562 while (nr_running) {
1563 reap_threads(&nr_running, &t_rate, &m_rate);
1571 int main(int argc, char *argv[])
1578 * We need locale for number printing, if it isn't set then just
1579 * go with the US format.
1581 if (!getenv("LC_NUMERIC"))
1582 setlocale(LC_NUMERIC, "en_US");
1584 if (parse_options(argc, argv))
1590 ps = sysconf(_SC_PAGESIZE);
1592 log_err("Failed to get page size\n");
1600 setup_log(&agg_io_log[DDIR_READ]);
1601 setup_log(&agg_io_log[DDIR_WRITE]);
1604 startup_mutex = fio_mutex_init(0);
1605 writeout_mutex = fio_mutex_init(1);
1616 __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1617 __finish_log(agg_io_log[DDIR_WRITE],
1618 "agg-write_bw.log");
1622 fio_mutex_remove(startup_mutex);
1623 fio_mutex_remove(writeout_mutex);