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);
175 static inline int should_check_rate(struct thread_data *td)
177 struct thread_options *o = &td->o;
180 * If some rate setting was given, we need to check it
182 if (o->rate || o->ratemin || o->rate_iops || o->rate_iops_min)
189 * Check if we are above the minimum rate given.
191 static int check_min_rate(struct thread_data *td, struct timeval *now)
193 unsigned long long bytes = 0;
194 unsigned long iops = 0;
199 * allow a 2 second settle period in the beginning
201 if (mtime_since(&td->start, now) < 2000)
205 iops += td->io_blocks[DDIR_READ];
206 bytes += td->this_io_bytes[DDIR_READ];
209 iops += td->io_blocks[DDIR_WRITE];
210 bytes += td->this_io_bytes[DDIR_WRITE];
214 * if rate blocks is set, sample is running
216 if (td->rate_bytes || td->rate_blocks) {
217 spent = mtime_since(&td->lastrate, now);
218 if (spent < td->o.ratecycle)
223 * check bandwidth specified rate
225 if (bytes < td->rate_bytes) {
226 log_err("%s: min rate %u not met\n", td->o.name,
230 rate = (bytes - td->rate_bytes) / spent;
231 if (rate < td->o.ratemin ||
232 bytes < td->rate_bytes) {
233 log_err("%s: min rate %u not met, got"
234 " %luKiB/sec\n", td->o.name,
235 td->o.ratemin, rate);
241 * checks iops specified rate
243 if (iops < td->o.rate_iops) {
244 log_err("%s: min iops rate %u not met\n",
245 td->o.name, td->o.rate_iops);
248 rate = (iops - td->rate_blocks) / spent;
249 if (rate < td->o.rate_iops_min ||
250 iops < td->rate_blocks) {
251 log_err("%s: min iops rate %u not met,"
252 " got %lu\n", td->o.name,
260 td->rate_bytes = bytes;
261 td->rate_blocks = iops;
262 memcpy(&td->lastrate, now, sizeof(*now));
266 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
270 if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
277 * When job exits, we can cancel the in-flight IO if we are using async
278 * io. Attempt to do so.
280 static void cleanup_pending_aio(struct thread_data *td)
282 struct flist_head *entry, *n;
287 * get immediately available events, if any
289 r = io_u_queued_complete(td, 0);
294 * now cancel remaining active events
296 if (td->io_ops->cancel) {
297 flist_for_each_safe(entry, n, &td->io_u_busylist) {
298 io_u = flist_entry(entry, struct io_u, list);
301 * if the io_u isn't in flight, then that generally
302 * means someone leaked an io_u. complain but fix
303 * it up, so we don't stall here.
305 if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
306 log_err("fio: non-busy IO on busy list\n");
309 r = td->io_ops->cancel(td, io_u);
317 r = io_u_queued_complete(td, td->cur_depth);
321 * Helper to handle the final sync of a file. Works just like the normal
322 * io path, just does everything sync.
324 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
326 struct io_u *io_u = __get_io_u(td);
332 io_u->ddir = DDIR_SYNC;
335 if (td_io_prep(td, io_u)) {
341 ret = td_io_queue(td, io_u);
343 td_verror(td, io_u->error, "td_io_queue");
346 } else if (ret == FIO_Q_QUEUED) {
347 if (io_u_queued_complete(td, 1) < 0)
349 } else if (ret == FIO_Q_COMPLETED) {
351 td_verror(td, io_u->error, "td_io_queue");
355 if (io_u_sync_complete(td, io_u) < 0)
357 } else if (ret == FIO_Q_BUSY) {
358 if (td_io_commit(td))
366 static inline void update_tv_cache(struct thread_data *td)
368 if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
369 fio_gettime(&td->tv_cache, NULL);
373 * The main verify engine. Runs over the writes we previously submitted,
374 * reads the blocks back in, and checks the crc/md5 of the data.
376 static void do_verify(struct thread_data *td)
384 * sync io first and invalidate cache, to make sure we really
387 for_each_file(td, f, i) {
388 if (!fio_file_open(f))
390 if (fio_io_sync(td, f))
392 if (file_invalidate_cache(td, f))
399 td_set_runstate(td, TD_VERIFYING);
402 while (!td->terminate) {
405 io_u = __get_io_u(td);
411 if (runtime_exceeded(td, &td->tv_cache)) {
417 if (get_next_verify(td, io_u)) {
422 if (td_io_prep(td, io_u)) {
427 io_u->end_io = verify_io_u;
429 ret = td_io_queue(td, io_u);
431 case FIO_Q_COMPLETED:
434 else if (io_u->resid) {
435 int bytes = io_u->xfer_buflen - io_u->resid;
436 struct fio_file *f = io_u->file;
442 td_verror(td, EIO, "full resid");
447 io_u->xfer_buflen = io_u->resid;
448 io_u->xfer_buf += bytes;
449 io_u->offset += bytes;
451 td->ts.short_io_u[io_u->ddir]++;
453 if (io_u->offset == f->real_file_size)
456 requeue_io_u(td, &io_u);
459 ret = io_u_sync_complete(td, io_u);
467 requeue_io_u(td, &io_u);
468 ret2 = td_io_commit(td);
474 td_verror(td, -ret, "td_io_queue");
478 if (ret < 0 || td->error)
482 * if we can queue more, do so. but check if there are
483 * completed io_u's first.
485 full = queue_full(td) || ret == FIO_Q_BUSY;
486 if (full || !td->o.iodepth_batch_complete) {
487 min_events = td->o.iodepth_batch_complete;
488 if (full && !min_events)
493 * Reap required number of io units, if any,
494 * and do the verification on them through
495 * the callback handler
497 if (io_u_queued_complete(td, min_events) < 0) {
501 } while (full && (td->cur_depth > td->o.iodepth_low));
508 min_events = td->cur_depth;
511 ret = io_u_queued_complete(td, min_events);
513 cleanup_pending_aio(td);
515 td_set_runstate(td, TD_RUNNING);
519 * Main IO worker function. It retrieves io_u's to process and queues
520 * and reaps them, checking for rate and errors along the way.
522 static void do_io(struct thread_data *td)
528 if (in_ramp_time(td))
529 td_set_runstate(td, TD_RAMP);
531 td_set_runstate(td, TD_RUNNING);
533 while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) {
534 struct timeval comp_time;
549 if (runtime_exceeded(td, &td->tv_cache)) {
556 * Add verification end_io handler, if asked to verify
557 * a previously written file.
559 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ) {
560 io_u->end_io = verify_io_u;
561 td_set_runstate(td, TD_VERIFYING);
562 } else if (in_ramp_time(td))
563 td_set_runstate(td, TD_RAMP);
565 td_set_runstate(td, TD_RUNNING);
567 ret = td_io_queue(td, io_u);
569 case FIO_Q_COMPLETED:
572 else if (io_u->resid) {
573 int bytes = io_u->xfer_buflen - io_u->resid;
574 struct fio_file *f = io_u->file;
580 td_verror(td, EIO, "full resid");
585 io_u->xfer_buflen = io_u->resid;
586 io_u->xfer_buf += bytes;
587 io_u->offset += bytes;
589 td->ts.short_io_u[io_u->ddir]++;
591 if (io_u->offset == f->real_file_size)
594 requeue_io_u(td, &io_u);
597 if (should_check_rate(td))
598 fio_gettime(&comp_time, NULL);
600 bytes_done = io_u_sync_complete(td, io_u);
607 * if the engine doesn't have a commit hook,
608 * the io_u is really queued. if it does have such
609 * a hook, it has to call io_u_queued() itself.
611 if (td->io_ops->commit == NULL)
612 io_u_queued(td, io_u);
615 requeue_io_u(td, &io_u);
616 ret2 = td_io_commit(td);
626 if (ret < 0 || td->error)
630 * See if we need to complete some commands
632 full = queue_full(td) || ret == FIO_Q_BUSY;
633 if (full || !td->o.iodepth_batch_complete) {
634 min_evts = td->o.iodepth_batch_complete;
635 if (full && !min_evts)
638 if (should_check_rate(td))
639 fio_gettime(&comp_time, NULL);
642 ret = io_u_queued_complete(td, min_evts);
647 } while (full && (td->cur_depth > td->o.iodepth_low));
656 * the rate is batched for now, it should work for batches
657 * of completions except the very first one which may look
660 if (!in_ramp_time(td) && should_check_rate(td)) {
661 usec = utime_since(&td->tv_cache, &comp_time);
663 rate_throttle(td, usec, bytes_done);
665 if (check_min_rate(td, &comp_time)) {
666 if (exitall_on_terminate)
667 terminate_threads(td->groupid);
668 td_verror(td, EIO, "check_min_rate");
673 if (td->o.thinktime) {
674 unsigned long long b;
676 b = td->io_blocks[0] + td->io_blocks[1];
677 if (!(b % td->o.thinktime_blocks)) {
680 if (td->o.thinktime_spin)
681 usec_spin(td->o.thinktime_spin);
683 left = td->o.thinktime - td->o.thinktime_spin;
685 usec_sleep(td, left);
690 if (td->o.fill_device && td->error == ENOSPC) {
699 ret = io_u_queued_complete(td, i);
701 if (should_fsync(td) && td->o.end_fsync) {
702 td_set_runstate(td, TD_FSYNCING);
704 for_each_file(td, f, i) {
705 if (!fio_file_open(f))
711 cleanup_pending_aio(td);
714 * stop job if we failed doing any IO
716 if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
720 static void cleanup_io_u(struct thread_data *td)
722 struct flist_head *entry, *n;
725 flist_for_each_safe(entry, n, &td->io_u_freelist) {
726 io_u = flist_entry(entry, struct io_u, list);
728 flist_del(&io_u->list);
735 static int init_io_u(struct thread_data *td)
739 int cl_align, i, max_units;
742 max_units = td->o.iodepth;
743 max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
744 td->orig_buffer_size = (unsigned long long) max_bs
745 * (unsigned long long) max_units;
747 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
750 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
751 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
754 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
755 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
759 if (allocate_io_mem(td))
763 p = ALIGN(td->orig_buffer);
767 cl_align = os_cache_line_size();
769 for (i = 0; i < max_units; i++) {
775 if (posix_memalign(&ptr, cl_align, sizeof(*io_u))) {
776 log_err("fio: posix_memalign=%s\n", strerror(errno));
781 memset(io_u, 0, sizeof(*io_u));
782 INIT_FLIST_HEAD(&io_u->list);
784 if (!(td->io_ops->flags & FIO_NOIO)) {
785 io_u->buf = p + max_bs * i;
787 if (td_write(td) && !td->o.refill_buffers)
788 io_u_fill_buffer(td, io_u, max_bs);
792 io_u->flags = IO_U_F_FREE;
793 flist_add(&io_u->list, &td->io_u_freelist);
799 static int switch_ioscheduler(struct thread_data *td)
801 char tmp[256], tmp2[128];
805 if (td->io_ops->flags & FIO_DISKLESSIO)
808 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
810 f = fopen(tmp, "r+");
812 if (errno == ENOENT) {
813 log_err("fio: os or kernel doesn't support IO scheduler"
817 td_verror(td, errno, "fopen iosched");
824 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
825 if (ferror(f) || ret != 1) {
826 td_verror(td, errno, "fwrite");
834 * Read back and check that the selected scheduler is now the default.
836 ret = fread(tmp, 1, sizeof(tmp), f);
837 if (ferror(f) || ret < 0) {
838 td_verror(td, errno, "fread");
843 sprintf(tmp2, "[%s]", td->o.ioscheduler);
844 if (!strstr(tmp, tmp2)) {
845 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
846 td_verror(td, EINVAL, "iosched_switch");
855 static int keep_running(struct thread_data *td)
857 unsigned long long io_done;
861 if (td->o.time_based)
868 io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
870 if (io_done < td->o.size)
876 static void reset_io_counters(struct thread_data *td)
878 td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
879 td->this_io_bytes[0] = td->this_io_bytes[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 elapsed = utime_since_now(&td->start);
1068 runtime[DDIR_READ] += elapsed;
1070 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1071 elapsed = utime_since_now(&td->start);
1072 runtime[DDIR_WRITE] += elapsed;
1075 if (td->error || td->terminate)
1078 if (!td->o.do_verify ||
1079 td->o.verify == VERIFY_NONE ||
1080 (td->io_ops->flags & FIO_UNIDIR))
1085 fio_gettime(&td->start, NULL);
1089 runtime[DDIR_READ] += utime_since_now(&td->start);
1091 if (td->error || td->terminate)
1095 update_rusage_stat(td);
1096 td->ts.runtime[0] = (runtime[0] + 999) / 1000;
1097 td->ts.runtime[1] = (runtime[1] + 999) / 1000;
1098 td->ts.total_run_time = mtime_since_now(&td->epoch);
1099 td->ts.io_bytes[0] = td->io_bytes[0];
1100 td->ts.io_bytes[1] = td->io_bytes[1];
1102 fio_mutex_down(writeout_mutex);
1103 if (td->ts.bw_log) {
1104 if (td->o.bw_log_file) {
1105 finish_log_named(td, td->ts.bw_log,
1106 td->o.bw_log_file, "bw");
1108 finish_log(td, td->ts.bw_log, "bw");
1110 if (td->ts.slat_log) {
1111 if (td->o.lat_log_file) {
1112 finish_log_named(td, td->ts.slat_log,
1113 td->o.lat_log_file, "slat");
1115 finish_log(td, td->ts.slat_log, "slat");
1117 if (td->ts.clat_log) {
1118 if (td->o.lat_log_file) {
1119 finish_log_named(td, td->ts.clat_log,
1120 td->o.lat_log_file, "clat");
1122 finish_log(td, td->ts.clat_log, "clat");
1124 fio_mutex_up(writeout_mutex);
1125 if (td->o.exec_postrun)
1126 exec_string(td->o.exec_postrun);
1128 if (exitall_on_terminate)
1129 terminate_threads(td->groupid);
1133 printf("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1135 close_and_free_files(td);
1139 if (td->o.cpumask_set) {
1140 int ret = fio_cpuset_exit(&td->o.cpumask);
1142 td_verror(td, ret, "fio_cpuset_exit");
1146 * do this very late, it will log file closing as well
1148 if (td->o.write_iolog_file)
1149 write_iolog_close(td);
1151 options_mem_free(td);
1152 td_set_runstate(td, TD_EXITED);
1153 return (void *) (unsigned long) td->error;
1157 * We cannot pass the td data into a forked process, so attach the td and
1158 * pass it to the thread worker.
1160 static int fork_main(int shmid, int offset)
1162 struct thread_data *td;
1165 data = shmat(shmid, NULL, 0);
1166 if (data == (void *) -1) {
1173 td = data + offset * sizeof(struct thread_data);
1174 ret = thread_main(td);
1176 return (int) (unsigned long) ret;
1180 * Run over the job map and reap the threads that have exited, if any.
1182 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1184 struct thread_data *td;
1185 int i, cputhreads, realthreads, pending, status, ret;
1188 * reap exited threads (TD_EXITED -> TD_REAPED)
1190 realthreads = pending = cputhreads = 0;
1191 for_each_td(td, i) {
1195 * ->io_ops is NULL for a thread that has closed its
1198 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1207 if (td->runstate == TD_REAPED)
1209 if (td->o.use_thread) {
1210 if (td->runstate == TD_EXITED) {
1211 td_set_runstate(td, TD_REAPED);
1218 if (td->runstate == TD_EXITED)
1222 * check if someone quit or got killed in an unusual way
1224 ret = waitpid(td->pid, &status, flags);
1226 if (errno == ECHILD) {
1227 log_err("fio: pid=%d disappeared %d\n",
1228 (int) td->pid, td->runstate);
1229 td_set_runstate(td, TD_REAPED);
1233 } else if (ret == td->pid) {
1234 if (WIFSIGNALED(status)) {
1235 int sig = WTERMSIG(status);
1238 log_err("fio: pid=%d, got signal=%d\n",
1239 (int) td->pid, sig);
1240 td_set_runstate(td, TD_REAPED);
1243 if (WIFEXITED(status)) {
1244 if (WEXITSTATUS(status) && !td->error)
1245 td->error = WEXITSTATUS(status);
1247 td_set_runstate(td, TD_REAPED);
1253 * thread is not dead, continue
1259 (*m_rate) -= td->o.ratemin;
1260 (*t_rate) -= td->o.rate;
1267 done_secs += mtime_since_now(&td->epoch) / 1000;
1270 if (*nr_running == cputhreads && !pending && realthreads)
1271 terminate_threads(TERMINATE_ALL);
1274 static void *gtod_thread_main(void *data)
1276 fio_mutex_up(startup_mutex);
1279 * As long as we have jobs around, update the clock. It would be nice
1280 * to have some way of NOT hammering that CPU with gettimeofday(),
1281 * but I'm not sure what to use outside of a simple CPU nop to relax
1282 * it - we don't want to lose precision.
1292 static int fio_start_gtod_thread(void)
1296 ret = pthread_create(>od_thread, NULL, gtod_thread_main, NULL);
1298 log_err("Can't create gtod thread: %s\n", strerror(ret));
1302 ret = pthread_detach(gtod_thread);
1304 log_err("Can't detatch gtod thread: %s\n", strerror(ret));
1308 dprint(FD_MUTEX, "wait on startup_mutex\n");
1309 fio_mutex_down(startup_mutex);
1310 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1315 * Main function for kicking off and reaping jobs, as needed.
1317 static void run_threads(void)
1319 struct thread_data *td;
1320 unsigned long spent;
1321 int i, todo, nr_running, m_rate, t_rate, nr_started;
1323 if (fio_pin_memory())
1326 if (fio_gtod_offload && fio_start_gtod_thread())
1329 if (!terse_output) {
1330 printf("Starting ");
1332 printf("%d thread%s", nr_thread,
1333 nr_thread > 1 ? "s" : "");
1337 printf("%d process%s", nr_process,
1338 nr_process > 1 ? "es" : "");
1346 todo = thread_number;
1349 m_rate = t_rate = 0;
1351 for_each_td(td, i) {
1352 print_status_init(td->thread_number - 1);
1354 if (!td->o.create_serialize) {
1360 * do file setup here so it happens sequentially,
1361 * we don't want X number of threads getting their
1362 * client data interspersed on disk
1364 if (setup_files(td)) {
1367 log_err("fio: pid=%d, err=%d/%s\n",
1368 (int) td->pid, td->error, td->verror);
1369 td_set_runstate(td, TD_REAPED);
1376 * for sharing to work, each job must always open
1377 * its own files. so close them, if we opened them
1380 for_each_file(td, f, i)
1381 td_io_close_file(td, f);
1390 struct thread_data *map[MAX_JOBS];
1391 struct timeval this_start;
1392 int this_jobs = 0, left;
1395 * create threads (TD_NOT_CREATED -> TD_CREATED)
1397 for_each_td(td, i) {
1398 if (td->runstate != TD_NOT_CREATED)
1402 * never got a chance to start, killed by other
1403 * thread for some reason
1405 if (td->terminate) {
1410 if (td->o.start_delay) {
1411 spent = mtime_since_genesis();
1413 if (td->o.start_delay * 1000 > spent)
1417 if (td->o.stonewall && (nr_started || nr_running)) {
1418 dprint(FD_PROCESS, "%s: stonewall wait\n",
1424 * Set state to created. Thread will transition
1425 * to TD_INITIALIZED when it's done setting up.
1427 td_set_runstate(td, TD_CREATED);
1428 map[this_jobs++] = td;
1431 if (td->o.use_thread) {
1434 dprint(FD_PROCESS, "will pthread_create\n");
1435 ret = pthread_create(&td->thread, NULL,
1438 log_err("pthread_create: %s\n",
1443 ret = pthread_detach(td->thread);
1445 log_err("pthread_detach: %s",
1449 dprint(FD_PROCESS, "will fork\n");
1452 int ret = fork_main(shm_id, i);
1455 } else if (i == fio_debug_jobno)
1456 *fio_debug_jobp = pid;
1458 dprint(FD_MUTEX, "wait on startup_mutex\n");
1459 fio_mutex_down(startup_mutex);
1460 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1464 * Wait for the started threads to transition to
1467 fio_gettime(&this_start, NULL);
1469 while (left && !fio_abort) {
1470 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1475 for (i = 0; i < this_jobs; i++) {
1479 if (td->runstate == TD_INITIALIZED) {
1482 } else if (td->runstate >= TD_EXITED) {
1486 nr_running++; /* work-around... */
1492 log_err("fio: %d jobs failed to start\n", left);
1493 for (i = 0; i < this_jobs; i++) {
1497 kill(td->pid, SIGTERM);
1503 * start created threads (TD_INITIALIZED -> TD_RUNNING).
1505 for_each_td(td, i) {
1506 if (td->runstate != TD_INITIALIZED)
1509 if (in_ramp_time(td))
1510 td_set_runstate(td, TD_RAMP);
1512 td_set_runstate(td, TD_RUNNING);
1515 m_rate += td->o.ratemin;
1516 t_rate += td->o.rate;
1518 fio_mutex_up(td->mutex);
1521 reap_threads(&nr_running, &t_rate, &m_rate);
1527 while (nr_running) {
1528 reap_threads(&nr_running, &t_rate, &m_rate);
1536 int main(int argc, char *argv[])
1543 * We need locale for number printing, if it isn't set then just
1544 * go with the US format.
1546 if (!getenv("LC_NUMERIC"))
1547 setlocale(LC_NUMERIC, "en_US");
1549 if (parse_options(argc, argv))
1555 ps = sysconf(_SC_PAGESIZE);
1557 log_err("Failed to get page size\n");
1565 setup_log(&agg_io_log[DDIR_READ]);
1566 setup_log(&agg_io_log[DDIR_WRITE]);
1569 startup_mutex = fio_mutex_init(0);
1570 writeout_mutex = fio_mutex_init(1);
1581 __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1582 __finish_log(agg_io_log[DDIR_WRITE],
1583 "agg-write_bw.log");
1587 fio_mutex_remove(startup_mutex);
1588 fio_mutex_remove(writeout_mutex);