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
46 unsigned long page_mask;
47 unsigned long page_size;
49 #define PAGE_ALIGN(buf) \
50 (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
53 int thread_number = 0;
58 unsigned long done_secs = 0;
60 static struct fio_mutex *startup_mutex;
61 static struct fio_mutex *writeout_mutex;
62 static volatile int fio_abort;
63 static int exit_value;
64 static struct itimerval itimer;
65 static pthread_t gtod_thread;
66 static struct flist_head *cgroup_list;
67 static char *cgroup_mnt;
69 struct io_log *agg_io_log[2];
71 #define TERMINATE_ALL (-1)
72 #define JOB_START_TIMEOUT (5 * 1000)
74 void td_set_runstate(struct thread_data *td, int runstate)
76 if (td->runstate == runstate)
79 dprint(FD_PROCESS, "pid=%d: runstate %d -> %d\n", (int) td->pid,
80 td->runstate, runstate);
81 td->runstate = runstate;
84 static void terminate_threads(int group_id)
86 struct thread_data *td;
89 dprint(FD_PROCESS, "terminate group_id=%d\n", group_id);
92 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
93 dprint(FD_PROCESS, "setting terminate on %s/%d\n",
94 td->o.name, (int) td->pid);
96 td->o.start_delay = 0;
99 * if the thread is running, just let it exit
101 if (td->runstate < TD_RUNNING)
102 kill(td->pid, SIGQUIT);
104 struct ioengine_ops *ops = td->io_ops;
106 if (ops && (ops->flags & FIO_SIGQUIT))
107 kill(td->pid, SIGQUIT);
113 static void status_timer_arm(void)
115 itimer.it_value.tv_sec = 0;
116 itimer.it_value.tv_usec = DISK_UTIL_MSEC * 1000;
117 setitimer(ITIMER_REAL, &itimer, NULL);
120 static void sig_alrm(int fio_unused sig)
124 print_thread_status();
130 * Happens on thread runs with ctrl-c, ignore our own SIGQUIT
132 static void sig_quit(int sig)
136 static void sig_int(int sig)
139 log_info("\nfio: terminating on signal %d\n", sig);
141 terminate_threads(TERMINATE_ALL);
145 static void set_sig_handlers(void)
147 struct sigaction act;
149 memset(&act, 0, sizeof(act));
150 act.sa_handler = sig_alrm;
151 act.sa_flags = SA_RESTART;
152 sigaction(SIGALRM, &act, NULL);
154 memset(&act, 0, sizeof(act));
155 act.sa_handler = sig_int;
156 act.sa_flags = SA_RESTART;
157 sigaction(SIGINT, &act, NULL);
159 memset(&act, 0, sizeof(act));
160 act.sa_handler = sig_quit;
161 act.sa_flags = SA_RESTART;
162 sigaction(SIGQUIT, &act, NULL);
166 * Check if we are above the minimum rate given.
168 static int __check_min_rate(struct thread_data *td, struct timeval *now,
171 unsigned long long bytes = 0;
172 unsigned long iops = 0;
175 unsigned int ratemin = 0;
176 unsigned int rate_iops = 0;
177 unsigned int rate_iops_min = 0;
179 if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
183 * allow a 2 second settle period in the beginning
185 if (mtime_since(&td->start, now) < 2000)
188 iops += td->io_blocks[ddir];
189 bytes += td->this_io_bytes[ddir];
190 ratemin += td->o.ratemin[ddir];
191 rate_iops += td->o.rate_iops[ddir];
192 rate_iops_min += td->o.rate_iops_min[ddir];
195 * if rate blocks is set, sample is running
197 if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
198 spent = mtime_since(&td->lastrate[ddir], now);
199 if (spent < td->o.ratecycle)
202 if (td->o.rate[ddir]) {
204 * check bandwidth specified rate
206 if (bytes < td->rate_bytes[ddir]) {
207 log_err("%s: min rate %u not met\n", td->o.name,
211 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
212 if (rate < ratemin ||
213 bytes < td->rate_bytes[ddir]) {
214 log_err("%s: min rate %u not met, got"
215 " %luKB/sec\n", td->o.name,
222 * checks iops specified rate
224 if (iops < rate_iops) {
225 log_err("%s: min iops rate %u not met\n",
226 td->o.name, rate_iops);
229 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
230 if (rate < rate_iops_min ||
231 iops < td->rate_blocks[ddir]) {
232 log_err("%s: min iops rate %u not met,"
233 " got %lu\n", td->o.name,
234 rate_iops_min, rate);
240 td->rate_bytes[ddir] = bytes;
241 td->rate_blocks[ddir] = iops;
242 memcpy(&td->lastrate[ddir], now, sizeof(*now));
246 static int check_min_rate(struct thread_data *td, struct timeval *now,
247 unsigned long *bytes_done)
252 ret |= __check_min_rate(td, now, 0);
254 ret |= __check_min_rate(td, now, 1);
259 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
263 if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
270 * When job exits, we can cancel the in-flight IO if we are using async
271 * io. Attempt to do so.
273 static void cleanup_pending_aio(struct thread_data *td)
275 struct flist_head *entry, *n;
280 * get immediately available events, if any
282 r = io_u_queued_complete(td, 0, NULL);
287 * now cancel remaining active events
289 if (td->io_ops->cancel) {
290 flist_for_each_safe(entry, n, &td->io_u_busylist) {
291 io_u = flist_entry(entry, struct io_u, list);
294 * if the io_u isn't in flight, then that generally
295 * means someone leaked an io_u. complain but fix
296 * it up, so we don't stall here.
298 if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
299 log_err("fio: non-busy IO on busy list\n");
302 r = td->io_ops->cancel(td, io_u);
310 r = io_u_queued_complete(td, td->cur_depth, NULL);
314 * Helper to handle the final sync of a file. Works just like the normal
315 * io path, just does everything sync.
317 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
319 struct io_u *io_u = __get_io_u(td);
325 io_u->ddir = DDIR_SYNC;
328 if (td_io_prep(td, io_u)) {
334 ret = td_io_queue(td, io_u);
336 td_verror(td, io_u->error, "td_io_queue");
339 } else if (ret == FIO_Q_QUEUED) {
340 if (io_u_queued_complete(td, 1, NULL) < 0)
342 } else if (ret == FIO_Q_COMPLETED) {
344 td_verror(td, io_u->error, "td_io_queue");
348 if (io_u_sync_complete(td, io_u, NULL) < 0)
350 } else if (ret == FIO_Q_BUSY) {
351 if (td_io_commit(td))
359 static inline void update_tv_cache(struct thread_data *td)
361 if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
362 fio_gettime(&td->tv_cache, NULL);
365 static int break_on_this_error(struct thread_data *td, int *retptr)
369 if (ret < 0 || td->error) {
372 if (!td->o.continue_on_error)
380 if (td_non_fatal_error(err)) {
382 * Continue with the I/Os in case of
385 update_error_count(td, err);
389 } else if (td->o.fill_device && err == ENOSPC) {
391 * We expect to hit this error if
392 * fill_device option is set.
399 * Stop the I/O in case of a fatal
402 update_error_count(td, err);
411 * The main verify engine. Runs over the writes we previously submitted,
412 * reads the blocks back in, and checks the crc/md5 of the data.
414 static void do_verify(struct thread_data *td)
421 dprint(FD_VERIFY, "starting loop\n");
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) {
447 if (runtime_exceeded(td, &td->tv_cache)) {
452 io_u = __get_io_u(td);
456 if (get_next_verify(td, io_u)) {
461 if (td_io_prep(td, io_u)) {
466 if (td->o.verify_async)
467 io_u->end_io = verify_io_u_async;
469 io_u->end_io = verify_io_u;
471 ret = td_io_queue(td, io_u);
473 case FIO_Q_COMPLETED:
476 clear_io_u(td, io_u);
477 } else if (io_u->resid) {
478 int bytes = io_u->xfer_buflen - io_u->resid;
479 struct fio_file *f = io_u->file;
485 td_verror(td, EIO, "full resid");
490 io_u->xfer_buflen = io_u->resid;
491 io_u->xfer_buf += bytes;
492 io_u->offset += bytes;
494 td->ts.short_io_u[io_u->ddir]++;
496 if (io_u->offset == f->real_file_size)
499 requeue_io_u(td, &io_u);
502 ret = io_u_sync_complete(td, io_u, NULL);
510 requeue_io_u(td, &io_u);
511 ret2 = td_io_commit(td);
517 td_verror(td, -ret, "td_io_queue");
521 if (break_on_this_error(td, &ret))
525 * if we can queue more, do so. but check if there are
526 * completed io_u's first.
528 full = queue_full(td) || ret == FIO_Q_BUSY;
529 if (full || !td->o.iodepth_batch_complete) {
530 min_events = min(td->o.iodepth_batch_complete,
532 if (full && !min_events)
537 * Reap required number of io units, if any,
538 * and do the verification on them through
539 * the callback handler
541 if (io_u_queued_complete(td, min_events, NULL) < 0) {
545 } while (full && (td->cur_depth > td->o.iodepth_low));
552 min_events = td->cur_depth;
555 ret = io_u_queued_complete(td, min_events, NULL);
557 cleanup_pending_aio(td);
559 td_set_runstate(td, TD_RUNNING);
561 dprint(FD_VERIFY, "exiting loop\n");
565 * Main IO worker function. It retrieves io_u's to process and queues
566 * and reaps them, checking for rate and errors along the way.
568 static void do_io(struct thread_data *td)
573 if (in_ramp_time(td))
574 td_set_runstate(td, TD_RAMP);
576 td_set_runstate(td, TD_RUNNING);
578 while ( (td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
579 ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) ) {
580 struct timeval comp_time;
581 unsigned long bytes_done[2] = { 0, 0 };
591 if (runtime_exceeded(td, &td->tv_cache)) {
601 * Add verification end_io handler, if asked to verify
602 * a previously written file.
604 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
606 if (td->o.verify_async)
607 io_u->end_io = verify_io_u_async;
609 io_u->end_io = verify_io_u;
610 td_set_runstate(td, TD_VERIFYING);
611 } else if (in_ramp_time(td))
612 td_set_runstate(td, TD_RAMP);
614 td_set_runstate(td, TD_RUNNING);
616 ret = td_io_queue(td, io_u);
618 case FIO_Q_COMPLETED:
621 clear_io_u(td, io_u);
622 } else if (io_u->resid) {
623 int bytes = io_u->xfer_buflen - io_u->resid;
624 struct fio_file *f = io_u->file;
630 td_verror(td, EIO, "full resid");
635 io_u->xfer_buflen = io_u->resid;
636 io_u->xfer_buf += bytes;
637 io_u->offset += bytes;
639 td->ts.short_io_u[io_u->ddir]++;
641 if (io_u->offset == f->real_file_size)
644 requeue_io_u(td, &io_u);
647 if (__should_check_rate(td, 0) ||
648 __should_check_rate(td, 1))
649 fio_gettime(&comp_time, NULL);
651 ret = io_u_sync_complete(td, io_u, bytes_done);
658 * if the engine doesn't have a commit hook,
659 * the io_u is really queued. if it does have such
660 * a hook, it has to call io_u_queued() itself.
662 if (td->io_ops->commit == NULL)
663 io_u_queued(td, io_u);
666 requeue_io_u(td, &io_u);
667 ret2 = td_io_commit(td);
677 if (break_on_this_error(td, &ret))
681 * See if we need to complete some commands
683 full = queue_full(td) || ret == FIO_Q_BUSY;
684 if (full || !td->o.iodepth_batch_complete) {
685 min_evts = min(td->o.iodepth_batch_complete,
687 if (full && !min_evts)
690 if (__should_check_rate(td, 0) ||
691 __should_check_rate(td, 1))
692 fio_gettime(&comp_time, NULL);
695 ret = io_u_queued_complete(td, min_evts, bytes_done);
699 } while (full && (td->cur_depth > td->o.iodepth_low));
704 if (!(bytes_done[0] + bytes_done[1]))
707 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
708 if (check_min_rate(td, &comp_time, bytes_done)) {
709 if (exitall_on_terminate)
710 terminate_threads(td->groupid);
711 td_verror(td, EIO, "check_min_rate");
716 if (td->o.thinktime) {
717 unsigned long long b;
719 b = td->io_blocks[0] + td->io_blocks[1];
720 if (!(b % td->o.thinktime_blocks)) {
723 if (td->o.thinktime_spin)
724 usec_spin(td->o.thinktime_spin);
726 left = td->o.thinktime - td->o.thinktime_spin;
728 usec_sleep(td, left);
733 if (td->o.fill_device && td->error == ENOSPC) {
742 ret = io_u_queued_complete(td, i, NULL);
744 if (should_fsync(td) && td->o.end_fsync) {
745 td_set_runstate(td, TD_FSYNCING);
747 for_each_file(td, f, i) {
748 if (!fio_file_open(f))
754 cleanup_pending_aio(td);
757 * stop job if we failed doing any IO
759 if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
763 static void cleanup_io_u(struct thread_data *td)
765 struct flist_head *entry, *n;
768 flist_for_each_safe(entry, n, &td->io_u_freelist) {
769 io_u = flist_entry(entry, struct io_u, list);
771 flist_del(&io_u->list);
778 static int init_io_u(struct thread_data *td)
782 int cl_align, i, max_units;
785 max_units = td->o.iodepth;
786 max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
787 td->orig_buffer_size = (unsigned long long) max_bs
788 * (unsigned long long) max_units;
790 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
793 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
794 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
797 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
798 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
802 if (allocate_io_mem(td))
805 if (td->o.odirect || td->o.mem_align)
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 if (posix_memalign(&ptr, cl_align, sizeof(*io_u))) {
819 log_err("fio: posix_memalign=%s\n", strerror(errno));
824 memset(io_u, 0, sizeof(*io_u));
825 INIT_FLIST_HEAD(&io_u->list);
826 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
828 if (!(td->io_ops->flags & FIO_NOIO)) {
829 io_u->buf = p + max_bs * i;
830 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
832 if (td_write(td) && !td->o.refill_buffers)
833 io_u_fill_buffer(td, io_u, max_bs);
837 io_u->flags = IO_U_F_FREE;
838 flist_add(&io_u->list, &td->io_u_freelist);
844 static int switch_ioscheduler(struct thread_data *td)
846 char tmp[256], tmp2[128];
850 if (td->io_ops->flags & FIO_DISKLESSIO)
853 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
855 f = fopen(tmp, "r+");
857 if (errno == ENOENT) {
858 log_err("fio: os or kernel doesn't support IO scheduler"
862 td_verror(td, errno, "fopen iosched");
869 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
870 if (ferror(f) || ret != 1) {
871 td_verror(td, errno, "fwrite");
879 * Read back and check that the selected scheduler is now the default.
881 ret = fread(tmp, 1, sizeof(tmp), f);
882 if (ferror(f) || ret < 0) {
883 td_verror(td, errno, "fread");
888 sprintf(tmp2, "[%s]", td->o.ioscheduler);
889 if (!strstr(tmp, tmp2)) {
890 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
891 td_verror(td, EINVAL, "iosched_switch");
900 static int keep_running(struct thread_data *td)
902 unsigned long long io_done;
906 if (td->o.time_based)
913 io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
915 if (io_done < td->o.size)
921 static void reset_io_counters(struct thread_data *td)
923 td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
924 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
926 td->rate_bytes[0] = td->rate_bytes[1] = 0;
927 td->rate_blocks[0] = td->rate_blocks[1] = 0;
929 td->last_was_sync = 0;
932 * reset file done count if we are to start over
934 if (td->o.time_based || td->o.loops)
935 td->nr_done_files = 0;
938 * Set the same seed to get repeatable runs
940 td_fill_rand_seeds(td);
943 void reset_all_stats(struct thread_data *td)
948 reset_io_counters(td);
950 for (i = 0; i < 2; i++) {
952 td->io_blocks[i] = 0;
953 td->io_issues[i] = 0;
954 td->ts.total_io_u[i] = 0;
957 fio_gettime(&tv, NULL);
958 memcpy(&td->epoch, &tv, sizeof(tv));
959 memcpy(&td->start, &tv, sizeof(tv));
962 static void clear_io_state(struct thread_data *td)
967 reset_io_counters(td);
970 for_each_file(td, f, i)
971 fio_file_clear_done(f);
974 static int exec_string(const char *string)
976 int ret, newlen = strlen(string) + 1 + 8;
979 str = malloc(newlen);
980 sprintf(str, "sh -c %s", string);
984 log_err("fio: exec of cmd <%s> failed\n", str);
991 * Entry point for the thread based jobs. The process based jobs end up
992 * here as well, after a little setup.
994 static void *thread_main(void *data)
996 unsigned long long runtime[2], elapsed;
997 struct thread_data *td = data;
998 pthread_condattr_t attr;
1001 if (!td->o.use_thread)
1006 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1008 INIT_FLIST_HEAD(&td->io_u_freelist);
1009 INIT_FLIST_HEAD(&td->io_u_busylist);
1010 INIT_FLIST_HEAD(&td->io_u_requeues);
1011 INIT_FLIST_HEAD(&td->io_log_list);
1012 INIT_FLIST_HEAD(&td->io_hist_list);
1013 INIT_FLIST_HEAD(&td->verify_list);
1014 pthread_mutex_init(&td->io_u_lock, NULL);
1015 td->io_hist_tree = RB_ROOT;
1017 pthread_condattr_init(&attr);
1018 pthread_cond_init(&td->verify_cond, &attr);
1019 pthread_cond_init(&td->free_cond, &attr);
1021 td_set_runstate(td, TD_INITIALIZED);
1022 dprint(FD_MUTEX, "up startup_mutex\n");
1023 fio_mutex_up(startup_mutex);
1024 dprint(FD_MUTEX, "wait on td->mutex\n");
1025 fio_mutex_down(td->mutex);
1026 dprint(FD_MUTEX, "done waiting on td->mutex\n");
1029 * the ->mutex mutex is now no longer used, close it to avoid
1030 * eating a file descriptor
1032 fio_mutex_remove(td->mutex);
1034 if (td->o.uid != -1U && setuid(td->o.uid)) {
1035 td_verror(td, errno, "setuid");
1038 if (td->o.gid != -1U && setgid(td->o.gid)) {
1039 td_verror(td, errno, "setgid");
1044 * May alter parameters that init_io_u() will use, so we need to
1053 if (td->o.verify_async && verify_async_init(td))
1056 if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1057 td_verror(td, errno, "cpu_set_affinity");
1062 * If we have a gettimeofday() thread, make sure we exclude that
1063 * thread from this job
1065 if (td->o.gtod_cpu) {
1066 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1067 if (fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1068 td_verror(td, errno, "cpu_set_affinity");
1073 if (td->ioprio_set) {
1074 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1075 td_verror(td, errno, "ioprio_set");
1080 if (td->o.cgroup_weight && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1083 if (nice(td->o.nice) == -1) {
1084 td_verror(td, errno, "nice");
1088 if (td->o.ioscheduler && switch_ioscheduler(td))
1091 if (!td->o.create_serialize && setup_files(td))
1097 if (init_random_map(td))
1100 if (td->o.exec_prerun) {
1101 if (exec_string(td->o.exec_prerun))
1105 if (td->o.pre_read) {
1106 if (pre_read_files(td) < 0)
1110 fio_gettime(&td->epoch, NULL);
1111 getrusage(RUSAGE_SELF, &td->ts.ru_start);
1113 runtime[0] = runtime[1] = 0;
1115 while (keep_running(td)) {
1116 fio_gettime(&td->start, NULL);
1117 memcpy(&td->ts.stat_sample_time[0], &td->start,
1119 memcpy(&td->ts.stat_sample_time[1], &td->start,
1121 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1123 if (td->o.ratemin[0] || td->o.ratemin[1])
1124 memcpy(&td->lastrate, &td->ts.stat_sample_time,
1125 sizeof(td->lastrate));
1130 prune_io_piece_log(td);
1136 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1137 elapsed = utime_since_now(&td->start);
1138 runtime[DDIR_READ] += elapsed;
1140 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1141 elapsed = utime_since_now(&td->start);
1142 runtime[DDIR_WRITE] += elapsed;
1145 if (td->error || td->terminate)
1148 if (!td->o.do_verify ||
1149 td->o.verify == VERIFY_NONE ||
1150 (td->io_ops->flags & FIO_UNIDIR))
1155 fio_gettime(&td->start, NULL);
1159 runtime[DDIR_READ] += utime_since_now(&td->start);
1161 if (td->error || td->terminate)
1165 update_rusage_stat(td);
1166 td->ts.runtime[0] = (runtime[0] + 999) / 1000;
1167 td->ts.runtime[1] = (runtime[1] + 999) / 1000;
1168 td->ts.total_run_time = mtime_since_now(&td->epoch);
1169 td->ts.io_bytes[0] = td->io_bytes[0];
1170 td->ts.io_bytes[1] = td->io_bytes[1];
1172 fio_mutex_down(writeout_mutex);
1173 if (td->ts.bw_log) {
1174 if (td->o.bw_log_file) {
1175 finish_log_named(td, td->ts.bw_log,
1176 td->o.bw_log_file, "bw");
1178 finish_log(td, td->ts.bw_log, "bw");
1180 if (td->ts.slat_log) {
1181 if (td->o.lat_log_file) {
1182 finish_log_named(td, td->ts.slat_log,
1183 td->o.lat_log_file, "slat");
1185 finish_log(td, td->ts.slat_log, "slat");
1187 if (td->ts.clat_log) {
1188 if (td->o.lat_log_file) {
1189 finish_log_named(td, td->ts.clat_log,
1190 td->o.lat_log_file, "clat");
1192 finish_log(td, td->ts.clat_log, "clat");
1194 fio_mutex_up(writeout_mutex);
1195 if (td->o.exec_postrun)
1196 exec_string(td->o.exec_postrun);
1198 if (exitall_on_terminate)
1199 terminate_threads(td->groupid);
1203 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1206 if (td->o.verify_async)
1207 verify_async_exit(td);
1209 close_and_free_files(td);
1212 cgroup_shutdown(td, &cgroup_mnt);
1214 if (td->o.cpumask_set) {
1215 int ret = fio_cpuset_exit(&td->o.cpumask);
1217 td_verror(td, ret, "fio_cpuset_exit");
1221 * do this very late, it will log file closing as well
1223 if (td->o.write_iolog_file)
1224 write_iolog_close(td);
1226 options_mem_free(td);
1227 td_set_runstate(td, TD_EXITED);
1228 return (void *) (unsigned long) td->error;
1232 * We cannot pass the td data into a forked process, so attach the td and
1233 * pass it to the thread worker.
1235 static int fork_main(int shmid, int offset)
1237 struct thread_data *td;
1240 data = shmat(shmid, NULL, 0);
1241 if (data == (void *) -1) {
1248 td = data + offset * sizeof(struct thread_data);
1249 ret = thread_main(td);
1251 return (int) (unsigned long) ret;
1255 * Run over the job map and reap the threads that have exited, if any.
1257 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1259 struct thread_data *td;
1260 int i, cputhreads, realthreads, pending, status, ret;
1263 * reap exited threads (TD_EXITED -> TD_REAPED)
1265 realthreads = pending = cputhreads = 0;
1266 for_each_td(td, i) {
1270 * ->io_ops is NULL for a thread that has closed its
1273 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1282 if (td->runstate == TD_REAPED)
1284 if (td->o.use_thread) {
1285 if (td->runstate == TD_EXITED) {
1286 td_set_runstate(td, TD_REAPED);
1293 if (td->runstate == TD_EXITED)
1297 * check if someone quit or got killed in an unusual way
1299 ret = waitpid(td->pid, &status, flags);
1301 if (errno == ECHILD) {
1302 log_err("fio: pid=%d disappeared %d\n",
1303 (int) td->pid, td->runstate);
1304 td_set_runstate(td, TD_REAPED);
1308 } else if (ret == td->pid) {
1309 if (WIFSIGNALED(status)) {
1310 int sig = WTERMSIG(status);
1313 log_err("fio: pid=%d, got signal=%d\n",
1314 (int) td->pid, sig);
1315 td_set_runstate(td, TD_REAPED);
1318 if (WIFEXITED(status)) {
1319 if (WEXITSTATUS(status) && !td->error)
1320 td->error = WEXITSTATUS(status);
1322 td_set_runstate(td, TD_REAPED);
1328 * thread is not dead, continue
1334 (*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
1335 (*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
1342 done_secs += mtime_since_now(&td->epoch) / 1000;
1345 if (*nr_running == cputhreads && !pending && realthreads)
1346 terminate_threads(TERMINATE_ALL);
1349 static void *gtod_thread_main(void *data)
1351 fio_mutex_up(startup_mutex);
1354 * As long as we have jobs around, update the clock. It would be nice
1355 * to have some way of NOT hammering that CPU with gettimeofday(),
1356 * but I'm not sure what to use outside of a simple CPU nop to relax
1357 * it - we don't want to lose precision.
1367 static int fio_start_gtod_thread(void)
1371 ret = pthread_create(>od_thread, NULL, gtod_thread_main, NULL);
1373 log_err("Can't create gtod thread: %s\n", strerror(ret));
1377 ret = pthread_detach(gtod_thread);
1379 log_err("Can't detatch gtod thread: %s\n", strerror(ret));
1383 dprint(FD_MUTEX, "wait on startup_mutex\n");
1384 fio_mutex_down(startup_mutex);
1385 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1390 * Main function for kicking off and reaping jobs, as needed.
1392 static void run_threads(void)
1394 struct thread_data *td;
1395 unsigned long spent;
1396 int i, todo, nr_running, m_rate, t_rate, nr_started;
1398 if (fio_pin_memory())
1401 if (fio_gtod_offload && fio_start_gtod_thread())
1404 if (!terse_output) {
1405 log_info("Starting ");
1407 log_info("%d thread%s", nr_thread,
1408 nr_thread > 1 ? "s" : "");
1412 log_info("%d process%s", nr_process,
1413 nr_process > 1 ? "es" : "");
1421 todo = thread_number;
1424 m_rate = t_rate = 0;
1426 for_each_td(td, i) {
1427 print_status_init(td->thread_number - 1);
1429 if (!td->o.create_serialize) {
1435 * do file setup here so it happens sequentially,
1436 * we don't want X number of threads getting their
1437 * client data interspersed on disk
1439 if (setup_files(td)) {
1442 log_err("fio: pid=%d, err=%d/%s\n",
1443 (int) td->pid, td->error, td->verror);
1444 td_set_runstate(td, TD_REAPED);
1451 * for sharing to work, each job must always open
1452 * its own files. so close them, if we opened them
1455 for_each_file(td, f, i) {
1456 if (fio_file_open(f))
1457 td_io_close_file(td, f);
1467 struct thread_data *map[MAX_JOBS];
1468 struct timeval this_start;
1469 int this_jobs = 0, left;
1472 * create threads (TD_NOT_CREATED -> TD_CREATED)
1474 for_each_td(td, i) {
1475 if (td->runstate != TD_NOT_CREATED)
1479 * never got a chance to start, killed by other
1480 * thread for some reason
1482 if (td->terminate) {
1487 if (td->o.start_delay) {
1488 spent = mtime_since_genesis();
1490 if (td->o.start_delay * 1000 > spent)
1494 if (td->o.stonewall && (nr_started || nr_running)) {
1495 dprint(FD_PROCESS, "%s: stonewall wait\n",
1501 * Set state to created. Thread will transition
1502 * to TD_INITIALIZED when it's done setting up.
1504 td_set_runstate(td, TD_CREATED);
1505 map[this_jobs++] = td;
1508 if (td->o.use_thread) {
1511 dprint(FD_PROCESS, "will pthread_create\n");
1512 ret = pthread_create(&td->thread, NULL,
1515 log_err("pthread_create: %s\n",
1520 ret = pthread_detach(td->thread);
1522 log_err("pthread_detach: %s",
1526 dprint(FD_PROCESS, "will fork\n");
1529 int ret = fork_main(shm_id, i);
1532 } else if (i == fio_debug_jobno)
1533 *fio_debug_jobp = pid;
1535 dprint(FD_MUTEX, "wait on startup_mutex\n");
1536 if (fio_mutex_down_timeout(startup_mutex, 10)) {
1537 log_err("fio: job startup hung? exiting.\n");
1538 terminate_threads(TERMINATE_ALL);
1543 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1547 * Wait for the started threads to transition to
1550 fio_gettime(&this_start, NULL);
1552 while (left && !fio_abort) {
1553 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1558 for (i = 0; i < this_jobs; i++) {
1562 if (td->runstate == TD_INITIALIZED) {
1565 } else if (td->runstate >= TD_EXITED) {
1569 nr_running++; /* work-around... */
1575 log_err("fio: %d jobs failed to start\n", left);
1576 for (i = 0; i < this_jobs; i++) {
1580 kill(td->pid, SIGTERM);
1586 * start created threads (TD_INITIALIZED -> TD_RUNNING).
1588 for_each_td(td, i) {
1589 if (td->runstate != TD_INITIALIZED)
1592 if (in_ramp_time(td))
1593 td_set_runstate(td, TD_RAMP);
1595 td_set_runstate(td, TD_RUNNING);
1598 m_rate += td->o.ratemin[0] + td->o.ratemin[1];
1599 t_rate += td->o.rate[0] + td->o.rate[1];
1601 fio_mutex_up(td->mutex);
1604 reap_threads(&nr_running, &t_rate, &m_rate);
1610 while (nr_running) {
1611 reap_threads(&nr_running, &t_rate, &m_rate);
1619 int main(int argc, char *argv[])
1624 init_rand(&__fio_rand_state);
1627 * We need locale for number printing, if it isn't set then just
1628 * go with the US format.
1630 if (!getenv("LC_NUMERIC"))
1631 setlocale(LC_NUMERIC, "en_US");
1633 ps = sysconf(_SC_PAGESIZE);
1635 log_err("Failed to get page size\n");
1642 fio_keywords_init();
1644 if (parse_options(argc, argv))
1647 if (exec_profile && load_profile(exec_profile))
1654 setup_log(&agg_io_log[DDIR_READ]);
1655 setup_log(&agg_io_log[DDIR_WRITE]);
1658 startup_mutex = fio_mutex_init(0);
1659 writeout_mutex = fio_mutex_init(1);
1665 cgroup_list = smalloc(sizeof(*cgroup_list));
1666 INIT_FLIST_HEAD(cgroup_list);
1673 __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1674 __finish_log(agg_io_log[DDIR_WRITE],
1675 "agg-write_bw.log");
1679 cgroup_kill(cgroup_list);
1683 fio_mutex_remove(startup_mutex);
1684 fio_mutex_remove(writeout_mutex);