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
42 unsigned long page_mask;
43 unsigned long page_size;
45 (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
48 int thread_number = 0;
53 unsigned long done_secs = 0;
55 static struct fio_mutex *startup_mutex;
56 static struct fio_mutex *writeout_mutex;
57 static volatile int fio_abort;
58 static int exit_value;
59 static struct itimerval itimer;
60 static pthread_t gtod_thread;
62 struct io_log *agg_io_log[2];
64 #define TERMINATE_ALL (-1)
65 #define JOB_START_TIMEOUT (5 * 1000)
67 void td_set_runstate(struct thread_data *td, int runstate)
69 if (td->runstate == runstate)
72 dprint(FD_PROCESS, "pid=%d: runstate %d -> %d\n", (int) td->pid,
73 td->runstate, runstate);
74 td->runstate = runstate;
77 static void terminate_threads(int group_id)
79 struct thread_data *td;
82 dprint(FD_PROCESS, "terminate group_id=%d\n", group_id);
85 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
86 dprint(FD_PROCESS, "setting terminate on %s/%d\n",
87 td->o.name, (int) td->pid);
89 td->o.start_delay = 0;
92 * if the thread is running, just let it exit
94 if (td->runstate < TD_RUNNING)
95 kill(td->pid, SIGQUIT);
97 struct ioengine_ops *ops = td->io_ops;
99 if (ops && (ops->flags & FIO_SIGQUIT))
100 kill(td->pid, SIGQUIT);
106 static void status_timer_arm(void)
108 itimer.it_value.tv_sec = 0;
109 itimer.it_value.tv_usec = DISK_UTIL_MSEC * 1000;
110 setitimer(ITIMER_REAL, &itimer, NULL);
113 static void sig_alrm(int fio_unused sig)
117 print_thread_status();
123 * Happens on thread runs with ctrl-c, ignore our own SIGQUIT
125 static void sig_quit(int sig)
129 static void sig_int(int sig)
132 printf("\nfio: terminating on signal %d\n", sig);
134 terminate_threads(TERMINATE_ALL);
138 static void sig_ill(int fio_unused sig)
143 log_err("fio: illegal instruction. your cpu does not support "
144 "the sse4.2 instruction for crc32c\n");
145 terminate_threads(TERMINATE_ALL);
149 static void set_sig_handlers(void)
151 struct sigaction act;
153 memset(&act, 0, sizeof(act));
154 act.sa_handler = sig_alrm;
155 act.sa_flags = SA_RESTART;
156 sigaction(SIGALRM, &act, NULL);
158 memset(&act, 0, sizeof(act));
159 act.sa_handler = sig_int;
160 act.sa_flags = SA_RESTART;
161 sigaction(SIGINT, &act, NULL);
163 memset(&act, 0, sizeof(act));
164 act.sa_handler = sig_ill;
165 act.sa_flags = SA_RESTART;
166 sigaction(SIGILL, &act, NULL);
168 memset(&act, 0, sizeof(act));
169 act.sa_handler = sig_quit;
170 act.sa_flags = SA_RESTART;
171 sigaction(SIGQUIT, &act, NULL);
174 static inline int should_check_rate(struct thread_data *td)
176 struct thread_options *o = &td->o;
179 * If some rate setting was given, we need to check it
181 if (o->rate || o->ratemin || o->rate_iops || o->rate_iops_min)
188 * Check if we are above the minimum rate given.
190 static int check_min_rate(struct thread_data *td, struct timeval *now)
192 unsigned long long bytes = 0;
193 unsigned long iops = 0;
198 * allow a 2 second settle period in the beginning
200 if (mtime_since(&td->start, now) < 2000)
204 iops += td->io_blocks[DDIR_READ];
205 bytes += td->this_io_bytes[DDIR_READ];
208 iops += td->io_blocks[DDIR_WRITE];
209 bytes += td->this_io_bytes[DDIR_WRITE];
213 * if rate blocks is set, sample is running
215 if (td->rate_bytes || td->rate_blocks) {
216 spent = mtime_since(&td->lastrate, now);
217 if (spent < td->o.ratecycle)
222 * check bandwidth specified rate
224 if (bytes < td->rate_bytes) {
225 log_err("%s: min rate %u not met\n", td->o.name,
229 rate = (bytes - td->rate_bytes) / spent;
230 if (rate < td->o.ratemin ||
231 bytes < td->rate_bytes) {
232 log_err("%s: min rate %u not met, got"
233 " %luKiB/sec\n", td->o.name,
234 td->o.ratemin, rate);
240 * checks iops specified rate
242 if (iops < td->o.rate_iops) {
243 log_err("%s: min iops rate %u not met\n",
244 td->o.name, td->o.rate_iops);
247 rate = (iops - td->rate_blocks) / spent;
248 if (rate < td->o.rate_iops_min ||
249 iops < td->rate_blocks) {
250 log_err("%s: min iops rate %u not met,"
251 " got %lu\n", td->o.name,
259 td->rate_bytes = bytes;
260 td->rate_blocks = iops;
261 memcpy(&td->lastrate, now, sizeof(*now));
265 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
269 if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
276 * When job exits, we can cancel the in-flight IO if we are using async
277 * io. Attempt to do so.
279 static void cleanup_pending_aio(struct thread_data *td)
281 struct flist_head *entry, *n;
286 * get immediately available events, if any
288 r = io_u_queued_complete(td, 0);
293 * now cancel remaining active events
295 if (td->io_ops->cancel) {
296 flist_for_each_safe(entry, n, &td->io_u_busylist) {
297 io_u = flist_entry(entry, struct io_u, list);
300 * if the io_u isn't in flight, then that generally
301 * means someone leaked an io_u. complain but fix
302 * it up, so we don't stall here.
304 if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
305 log_err("fio: non-busy IO on busy list\n");
308 r = td->io_ops->cancel(td, io_u);
316 r = io_u_queued_complete(td, td->cur_depth);
320 * Helper to handle the final sync of a file. Works just like the normal
321 * io path, just does everything sync.
323 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
325 struct io_u *io_u = __get_io_u(td);
331 io_u->ddir = DDIR_SYNC;
334 if (td_io_prep(td, io_u)) {
340 ret = td_io_queue(td, io_u);
342 td_verror(td, io_u->error, "td_io_queue");
345 } else if (ret == FIO_Q_QUEUED) {
346 if (io_u_queued_complete(td, 1) < 0)
348 } else if (ret == FIO_Q_COMPLETED) {
350 td_verror(td, io_u->error, "td_io_queue");
354 if (io_u_sync_complete(td, io_u) < 0)
356 } else if (ret == FIO_Q_BUSY) {
357 if (td_io_commit(td))
365 static inline void update_tv_cache(struct thread_data *td)
367 if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
368 fio_gettime(&td->tv_cache, NULL);
372 * The main verify engine. Runs over the writes we previously submitted,
373 * reads the blocks back in, and checks the crc/md5 of the data.
375 static void do_verify(struct thread_data *td)
383 * sync io first and invalidate cache, to make sure we really
386 for_each_file(td, f, i) {
387 if (!fio_file_open(f))
389 if (fio_io_sync(td, f))
391 if (file_invalidate_cache(td, f))
398 td_set_runstate(td, TD_VERIFYING);
401 while (!td->terminate) {
404 io_u = __get_io_u(td);
410 if (runtime_exceeded(td, &td->tv_cache)) {
416 if (get_next_verify(td, io_u)) {
421 if (td_io_prep(td, io_u)) {
426 io_u->end_io = verify_io_u;
428 ret = td_io_queue(td, io_u);
430 case FIO_Q_COMPLETED:
433 else if (io_u->resid) {
434 int bytes = io_u->xfer_buflen - io_u->resid;
435 struct fio_file *f = io_u->file;
441 td_verror(td, EIO, "full resid");
446 io_u->xfer_buflen = io_u->resid;
447 io_u->xfer_buf += bytes;
448 io_u->offset += bytes;
450 td->ts.short_io_u[io_u->ddir]++;
452 if (io_u->offset == f->real_file_size)
455 requeue_io_u(td, &io_u);
458 ret = io_u_sync_complete(td, io_u);
466 requeue_io_u(td, &io_u);
467 ret2 = td_io_commit(td);
473 td_verror(td, -ret, "td_io_queue");
477 if (ret < 0 || td->error)
481 * if we can queue more, do so. but check if there are
482 * completed io_u's first.
484 full = queue_full(td) || ret == FIO_Q_BUSY;
485 if (full || !td->o.iodepth_batch_complete) {
486 min_events = td->o.iodepth_batch_complete;
487 if (full && !min_events)
492 * Reap required number of io units, if any,
493 * and do the verification on them through
494 * the callback handler
496 if (io_u_queued_complete(td, min_events) < 0) {
500 } while (full && (td->cur_depth > td->o.iodepth_low));
507 min_events = td->cur_depth;
510 ret = io_u_queued_complete(td, min_events);
512 cleanup_pending_aio(td);
514 td_set_runstate(td, TD_RUNNING);
518 * Main IO worker function. It retrieves io_u's to process and queues
519 * and reaps them, checking for rate and errors along the way.
521 static void do_io(struct thread_data *td)
527 if (in_ramp_time(td))
528 td_set_runstate(td, TD_RAMP);
530 td_set_runstate(td, TD_RUNNING);
532 while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) {
533 struct timeval comp_time;
548 if (runtime_exceeded(td, &td->tv_cache)) {
555 * Add verification end_io handler, if asked to verify
556 * a previously written file.
558 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ) {
559 io_u->end_io = verify_io_u;
560 td_set_runstate(td, TD_VERIFYING);
561 } else if (in_ramp_time(td))
562 td_set_runstate(td, TD_RAMP);
564 td_set_runstate(td, TD_RUNNING);
566 ret = td_io_queue(td, io_u);
568 case FIO_Q_COMPLETED:
571 else if (io_u->resid) {
572 int bytes = io_u->xfer_buflen - io_u->resid;
573 struct fio_file *f = io_u->file;
579 td_verror(td, EIO, "full resid");
584 io_u->xfer_buflen = io_u->resid;
585 io_u->xfer_buf += bytes;
586 io_u->offset += bytes;
588 td->ts.short_io_u[io_u->ddir]++;
590 if (io_u->offset == f->real_file_size)
593 requeue_io_u(td, &io_u);
596 if (should_check_rate(td))
597 fio_gettime(&comp_time, NULL);
599 bytes_done = io_u_sync_complete(td, io_u);
606 * if the engine doesn't have a commit hook,
607 * the io_u is really queued. if it does have such
608 * a hook, it has to call io_u_queued() itself.
610 if (td->io_ops->commit == NULL)
611 io_u_queued(td, io_u);
614 requeue_io_u(td, &io_u);
615 ret2 = td_io_commit(td);
625 if (ret < 0 || td->error)
629 * See if we need to complete some commands
631 full = queue_full(td) || ret == FIO_Q_BUSY;
632 if (full || !td->o.iodepth_batch_complete) {
633 min_evts = td->o.iodepth_batch_complete;
634 if (full && !min_evts)
637 if (should_check_rate(td))
638 fio_gettime(&comp_time, NULL);
641 ret = io_u_queued_complete(td, min_evts);
646 } while (full && (td->cur_depth > td->o.iodepth_low));
655 * the rate is batched for now, it should work for batches
656 * of completions except the very first one which may look
659 if (!in_ramp_time(td) && should_check_rate(td)) {
660 usec = utime_since(&td->tv_cache, &comp_time);
662 rate_throttle(td, usec, bytes_done);
664 if (check_min_rate(td, &comp_time)) {
665 if (exitall_on_terminate)
666 terminate_threads(td->groupid);
667 td_verror(td, EIO, "check_min_rate");
672 if (td->o.thinktime) {
673 unsigned long long b;
675 b = td->io_blocks[0] + td->io_blocks[1];
676 if (!(b % td->o.thinktime_blocks)) {
679 if (td->o.thinktime_spin)
680 usec_spin(td->o.thinktime_spin);
682 left = td->o.thinktime - td->o.thinktime_spin;
684 usec_sleep(td, left);
689 if (td->o.fill_device && td->error == ENOSPC) {
698 ret = io_u_queued_complete(td, i);
700 if (should_fsync(td) && td->o.end_fsync) {
701 td_set_runstate(td, TD_FSYNCING);
703 for_each_file(td, f, i) {
704 if (!fio_file_open(f))
710 cleanup_pending_aio(td);
713 * stop job if we failed doing any IO
715 if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
719 static void cleanup_io_u(struct thread_data *td)
721 struct flist_head *entry, *n;
724 flist_for_each_safe(entry, n, &td->io_u_freelist) {
725 io_u = flist_entry(entry, struct io_u, list);
727 flist_del(&io_u->list);
734 static int init_io_u(struct thread_data *td)
738 int cl_align, i, max_units;
741 max_units = td->o.iodepth;
742 max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
743 td->orig_buffer_size = (unsigned long long) max_bs
744 * (unsigned long long) max_units;
746 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
749 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
750 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
753 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
754 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
758 if (allocate_io_mem(td))
762 p = ALIGN(td->orig_buffer);
766 cl_align = os_cache_line_size();
768 for (i = 0; i < max_units; i++) {
774 if (posix_memalign(&ptr, cl_align, sizeof(*io_u))) {
775 log_err("fio: posix_memalign=%s\n", strerror(errno));
780 memset(io_u, 0, sizeof(*io_u));
781 INIT_FLIST_HEAD(&io_u->list);
783 if (!(td->io_ops->flags & FIO_NOIO)) {
784 io_u->buf = p + max_bs * i;
786 if (td_write(td) && !td->o.refill_buffers)
787 io_u_fill_buffer(td, io_u, max_bs);
791 io_u->flags = IO_U_F_FREE;
792 flist_add(&io_u->list, &td->io_u_freelist);
798 static int switch_ioscheduler(struct thread_data *td)
800 char tmp[256], tmp2[128];
804 if (td->io_ops->flags & FIO_DISKLESSIO)
807 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
809 f = fopen(tmp, "r+");
811 if (errno == ENOENT) {
812 log_err("fio: os or kernel doesn't support IO scheduler"
816 td_verror(td, errno, "fopen iosched");
823 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
824 if (ferror(f) || ret != 1) {
825 td_verror(td, errno, "fwrite");
833 * Read back and check that the selected scheduler is now the default.
835 ret = fread(tmp, 1, sizeof(tmp), f);
836 if (ferror(f) || ret < 0) {
837 td_verror(td, errno, "fread");
842 sprintf(tmp2, "[%s]", td->o.ioscheduler);
843 if (!strstr(tmp, tmp2)) {
844 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
845 td_verror(td, EINVAL, "iosched_switch");
854 static int keep_running(struct thread_data *td)
856 unsigned long long io_done;
860 if (td->o.time_based)
867 io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
869 if (io_done < td->o.size)
875 static void reset_io_counters(struct thread_data *td)
877 td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
878 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
882 td->rw_end_set[0] = td->rw_end_set[1] = 0;
884 td->last_was_sync = 0;
887 * reset file done count if we are to start over
889 if (td->o.time_based || td->o.loops)
890 td->nr_done_files = 0;
893 * Set the same seed to get repeatable runs
895 td_fill_rand_seeds(td);
898 void reset_all_stats(struct thread_data *td)
903 reset_io_counters(td);
905 for (i = 0; i < 2; i++) {
907 td->io_blocks[i] = 0;
908 td->io_issues[i] = 0;
909 td->ts.total_io_u[i] = 0;
912 fio_gettime(&tv, NULL);
913 memcpy(&td->epoch, &tv, sizeof(tv));
914 memcpy(&td->start, &tv, sizeof(tv));
917 static void clear_io_state(struct thread_data *td)
922 reset_io_counters(td);
925 for_each_file(td, f, i)
926 fio_file_clear_done(f);
929 static int exec_string(const char *string)
931 int ret, newlen = strlen(string) + 1 + 8;
934 str = malloc(newlen);
935 sprintf(str, "sh -c %s", string);
939 log_err("fio: exec of cmd <%s> failed\n", str);
946 * Entry point for the thread based jobs. The process based jobs end up
947 * here as well, after a little setup.
949 static void *thread_main(void *data)
951 unsigned long long runtime[2], elapsed;
952 struct thread_data *td = data;
955 if (!td->o.use_thread)
960 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
962 INIT_FLIST_HEAD(&td->io_u_freelist);
963 INIT_FLIST_HEAD(&td->io_u_busylist);
964 INIT_FLIST_HEAD(&td->io_u_requeues);
965 INIT_FLIST_HEAD(&td->io_log_list);
966 INIT_FLIST_HEAD(&td->io_hist_list);
967 td->io_hist_tree = RB_ROOT;
969 td_set_runstate(td, TD_INITIALIZED);
970 dprint(FD_MUTEX, "up startup_mutex\n");
971 fio_mutex_up(startup_mutex);
972 dprint(FD_MUTEX, "wait on td->mutex\n");
973 fio_mutex_down(td->mutex);
974 dprint(FD_MUTEX, "done waiting on td->mutex\n");
977 * the ->mutex mutex is now no longer used, close it to avoid
978 * eating a file descriptor
980 fio_mutex_remove(td->mutex);
983 * May alter parameters that init_io_u() will use, so we need to
992 if (td->o.cpumask_set && fio_setaffinity(td) == -1) {
993 td_verror(td, errno, "cpu_set_affinity");
998 * If we have a gettimeofday() thread, make sure we exclude that
999 * thread from this job
1001 if (td->o.gtod_cpu) {
1002 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1003 if (fio_setaffinity(td) == -1) {
1004 td_verror(td, errno, "cpu_set_affinity");
1009 if (td->ioprio_set) {
1010 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1011 td_verror(td, errno, "ioprio_set");
1016 if (nice(td->o.nice) == -1) {
1017 td_verror(td, errno, "nice");
1021 if (td->o.ioscheduler && switch_ioscheduler(td))
1024 if (!td->o.create_serialize && setup_files(td))
1030 if (init_random_map(td))
1033 if (td->o.exec_prerun) {
1034 if (exec_string(td->o.exec_prerun))
1038 if (td->o.pre_read) {
1039 if (pre_read_files(td) < 0)
1043 fio_gettime(&td->epoch, NULL);
1044 getrusage(RUSAGE_SELF, &td->ts.ru_start);
1046 runtime[0] = runtime[1] = 0;
1048 while (keep_running(td)) {
1049 fio_gettime(&td->start, NULL);
1050 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
1051 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1054 memcpy(&td->lastrate, &td->ts.stat_sample_time,
1055 sizeof(td->lastrate));
1060 prune_io_piece_log(td);
1066 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1067 if (td->rw_end_set[DDIR_READ])
1068 elapsed = utime_since(&td->start,
1069 &td->rw_end[DDIR_READ]);
1071 elapsed = utime_since_now(&td->start);
1073 runtime[DDIR_READ] += elapsed;
1075 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1076 if (td->rw_end_set[DDIR_WRITE])
1077 elapsed = utime_since(&td->start,
1078 &td->rw_end[DDIR_WRITE]);
1080 elapsed = utime_since_now(&td->start);
1082 runtime[DDIR_WRITE] += elapsed;
1085 if (td->error || td->terminate)
1088 if (!td->o.do_verify ||
1089 td->o.verify == VERIFY_NONE ||
1090 (td->io_ops->flags & FIO_UNIDIR))
1095 fio_gettime(&td->start, NULL);
1099 runtime[DDIR_READ] += utime_since_now(&td->start);
1101 if (td->error || td->terminate)
1105 update_rusage_stat(td);
1106 td->ts.runtime[0] = (runtime[0] + 999) / 1000;
1107 td->ts.runtime[1] = (runtime[1] + 999) / 1000;
1108 td->ts.total_run_time = mtime_since_now(&td->epoch);
1109 td->ts.io_bytes[0] = td->io_bytes[0];
1110 td->ts.io_bytes[1] = td->io_bytes[1];
1112 fio_mutex_down(writeout_mutex);
1113 if (td->ts.bw_log) {
1114 if (td->o.bw_log_file) {
1115 finish_log_named(td, td->ts.bw_log,
1116 td->o.bw_log_file, "bw");
1118 finish_log(td, td->ts.bw_log, "bw");
1120 if (td->ts.slat_log) {
1121 if (td->o.lat_log_file) {
1122 finish_log_named(td, td->ts.slat_log,
1123 td->o.lat_log_file, "slat");
1125 finish_log(td, td->ts.slat_log, "slat");
1127 if (td->ts.clat_log) {
1128 if (td->o.lat_log_file) {
1129 finish_log_named(td, td->ts.clat_log,
1130 td->o.lat_log_file, "clat");
1132 finish_log(td, td->ts.clat_log, "clat");
1134 fio_mutex_up(writeout_mutex);
1135 if (td->o.exec_postrun)
1136 exec_string(td->o.exec_postrun);
1138 if (exitall_on_terminate)
1139 terminate_threads(td->groupid);
1143 printf("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1145 close_and_free_files(td);
1149 if (td->o.cpumask_set) {
1150 int ret = fio_cpuset_exit(&td->o.cpumask);
1152 td_verror(td, ret, "fio_cpuset_exit");
1156 * do this very late, it will log file closing as well
1158 if (td->o.write_iolog_file)
1159 write_iolog_close(td);
1161 options_mem_free(td);
1162 td_set_runstate(td, TD_EXITED);
1163 return (void *) (unsigned long) td->error;
1167 * We cannot pass the td data into a forked process, so attach the td and
1168 * pass it to the thread worker.
1170 static int fork_main(int shmid, int offset)
1172 struct thread_data *td;
1175 data = shmat(shmid, NULL, 0);
1176 if (data == (void *) -1) {
1183 td = data + offset * sizeof(struct thread_data);
1184 ret = thread_main(td);
1186 return (int) (unsigned long) ret;
1190 * Run over the job map and reap the threads that have exited, if any.
1192 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1194 struct thread_data *td;
1195 int i, cputhreads, realthreads, pending, status, ret;
1198 * reap exited threads (TD_EXITED -> TD_REAPED)
1200 realthreads = pending = cputhreads = 0;
1201 for_each_td(td, i) {
1205 * ->io_ops is NULL for a thread that has closed its
1208 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1217 if (td->runstate == TD_REAPED)
1219 if (td->o.use_thread) {
1220 if (td->runstate == TD_EXITED) {
1221 td_set_runstate(td, TD_REAPED);
1228 if (td->runstate == TD_EXITED)
1232 * check if someone quit or got killed in an unusual way
1234 ret = waitpid(td->pid, &status, flags);
1236 if (errno == ECHILD) {
1237 log_err("fio: pid=%d disappeared %d\n",
1238 (int) td->pid, td->runstate);
1239 td_set_runstate(td, TD_REAPED);
1243 } else if (ret == td->pid) {
1244 if (WIFSIGNALED(status)) {
1245 int sig = WTERMSIG(status);
1248 log_err("fio: pid=%d, got signal=%d\n",
1249 (int) td->pid, sig);
1250 td_set_runstate(td, TD_REAPED);
1253 if (WIFEXITED(status)) {
1254 if (WEXITSTATUS(status) && !td->error)
1255 td->error = WEXITSTATUS(status);
1257 td_set_runstate(td, TD_REAPED);
1263 * thread is not dead, continue
1269 (*m_rate) -= td->o.ratemin;
1270 (*t_rate) -= td->o.rate;
1277 done_secs += mtime_since_now(&td->epoch) / 1000;
1280 if (*nr_running == cputhreads && !pending && realthreads)
1281 terminate_threads(TERMINATE_ALL);
1284 static void *gtod_thread_main(void *data)
1286 fio_mutex_up(startup_mutex);
1289 * As long as we have jobs around, update the clock. It would be nice
1290 * to have some way of NOT hammering that CPU with gettimeofday(),
1291 * but I'm not sure what to use outside of a simple CPU nop to relax
1292 * it - we don't want to lose precision.
1302 static int fio_start_gtod_thread(void)
1306 ret = pthread_create(>od_thread, NULL, gtod_thread_main, NULL);
1308 log_err("Can't create gtod thread: %s\n", strerror(ret));
1312 ret = pthread_detach(gtod_thread);
1314 log_err("Can't detatch gtod thread: %s\n", strerror(ret));
1318 dprint(FD_MUTEX, "wait on startup_mutex\n");
1319 fio_mutex_down(startup_mutex);
1320 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1325 * Main function for kicking off and reaping jobs, as needed.
1327 static void run_threads(void)
1329 struct thread_data *td;
1330 unsigned long spent;
1331 int i, todo, nr_running, m_rate, t_rate, nr_started;
1333 if (fio_pin_memory())
1336 if (fio_gtod_offload && fio_start_gtod_thread())
1339 if (!terse_output) {
1340 printf("Starting ");
1342 printf("%d thread%s", nr_thread,
1343 nr_thread > 1 ? "s" : "");
1347 printf("%d process%s", nr_process,
1348 nr_process > 1 ? "es" : "");
1356 todo = thread_number;
1359 m_rate = t_rate = 0;
1361 for_each_td(td, i) {
1362 print_status_init(td->thread_number - 1);
1364 if (!td->o.create_serialize) {
1370 * do file setup here so it happens sequentially,
1371 * we don't want X number of threads getting their
1372 * client data interspersed on disk
1374 if (setup_files(td)) {
1377 log_err("fio: pid=%d, err=%d/%s\n",
1378 (int) td->pid, td->error, td->verror);
1379 td_set_runstate(td, TD_REAPED);
1386 * for sharing to work, each job must always open
1387 * its own files. so close them, if we opened them
1390 for_each_file(td, f, i)
1391 td_io_close_file(td, f);
1400 struct thread_data *map[MAX_JOBS];
1401 struct timeval this_start;
1402 int this_jobs = 0, left;
1405 * create threads (TD_NOT_CREATED -> TD_CREATED)
1407 for_each_td(td, i) {
1408 if (td->runstate != TD_NOT_CREATED)
1412 * never got a chance to start, killed by other
1413 * thread for some reason
1415 if (td->terminate) {
1420 if (td->o.start_delay) {
1421 spent = mtime_since_genesis();
1423 if (td->o.start_delay * 1000 > spent)
1427 if (td->o.stonewall && (nr_started || nr_running)) {
1428 dprint(FD_PROCESS, "%s: stonewall wait\n",
1434 * Set state to created. Thread will transition
1435 * to TD_INITIALIZED when it's done setting up.
1437 td_set_runstate(td, TD_CREATED);
1438 map[this_jobs++] = td;
1441 if (td->o.use_thread) {
1444 dprint(FD_PROCESS, "will pthread_create\n");
1445 ret = pthread_create(&td->thread, NULL,
1448 log_err("pthread_create: %s\n",
1453 ret = pthread_detach(td->thread);
1455 log_err("pthread_detach: %s",
1459 dprint(FD_PROCESS, "will fork\n");
1462 int ret = fork_main(shm_id, i);
1465 } else if (i == fio_debug_jobno)
1466 *fio_debug_jobp = pid;
1468 dprint(FD_MUTEX, "wait on startup_mutex\n");
1469 fio_mutex_down(startup_mutex);
1470 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1474 * Wait for the started threads to transition to
1477 fio_gettime(&this_start, NULL);
1479 while (left && !fio_abort) {
1480 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1485 for (i = 0; i < this_jobs; i++) {
1489 if (td->runstate == TD_INITIALIZED) {
1492 } else if (td->runstate >= TD_EXITED) {
1496 nr_running++; /* work-around... */
1502 log_err("fio: %d jobs failed to start\n", left);
1503 for (i = 0; i < this_jobs; i++) {
1507 kill(td->pid, SIGTERM);
1513 * start created threads (TD_INITIALIZED -> TD_RUNNING).
1515 for_each_td(td, i) {
1516 if (td->runstate != TD_INITIALIZED)
1519 if (in_ramp_time(td))
1520 td_set_runstate(td, TD_RAMP);
1522 td_set_runstate(td, TD_RUNNING);
1525 m_rate += td->o.ratemin;
1526 t_rate += td->o.rate;
1528 fio_mutex_up(td->mutex);
1531 reap_threads(&nr_running, &t_rate, &m_rate);
1537 while (nr_running) {
1538 reap_threads(&nr_running, &t_rate, &m_rate);
1546 int main(int argc, char *argv[])
1553 * We need locale for number printing, if it isn't set then just
1554 * go with the US format.
1556 if (!getenv("LC_NUMERIC"))
1557 setlocale(LC_NUMERIC, "en_US");
1559 if (parse_options(argc, argv))
1565 ps = sysconf(_SC_PAGESIZE);
1567 log_err("Failed to get page size\n");
1575 setup_log(&agg_io_log[DDIR_READ]);
1576 setup_log(&agg_io_log[DDIR_WRITE]);
1579 startup_mutex = fio_mutex_init(0);
1580 writeout_mutex = fio_mutex_init(1);
1591 __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1592 __finish_log(agg_io_log[DDIR_WRITE],
1593 "agg-write_bw.log");
1597 fio_mutex_remove(startup_mutex);
1598 fio_mutex_remove(writeout_mutex);