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 #define PAGE_ALIGN(buf) \
47 (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
50 int thread_number = 0;
55 unsigned long done_secs = 0;
57 static struct fio_mutex *startup_mutex;
58 static struct fio_mutex *writeout_mutex;
59 static volatile int fio_abort;
60 static int exit_value;
61 static struct itimerval itimer;
62 static pthread_t gtod_thread;
64 struct io_log *agg_io_log[2];
66 #define TERMINATE_ALL (-1)
67 #define JOB_START_TIMEOUT (5 * 1000)
69 void td_set_runstate(struct thread_data *td, int runstate)
71 if (td->runstate == runstate)
74 dprint(FD_PROCESS, "pid=%d: runstate %d -> %d\n", (int) td->pid,
75 td->runstate, runstate);
76 td->runstate = runstate;
79 static void terminate_threads(int group_id)
81 struct thread_data *td;
84 dprint(FD_PROCESS, "terminate group_id=%d\n", group_id);
87 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
88 dprint(FD_PROCESS, "setting terminate on %s/%d\n",
89 td->o.name, (int) td->pid);
91 td->o.start_delay = 0;
94 * if the thread is running, just let it exit
96 if (td->runstate < TD_RUNNING)
97 kill(td->pid, SIGQUIT);
99 struct ioengine_ops *ops = td->io_ops;
101 if (ops && (ops->flags & FIO_SIGQUIT))
102 kill(td->pid, SIGQUIT);
108 static void status_timer_arm(void)
110 itimer.it_value.tv_sec = 0;
111 itimer.it_value.tv_usec = DISK_UTIL_MSEC * 1000;
112 setitimer(ITIMER_REAL, &itimer, NULL);
115 static void sig_alrm(int fio_unused sig)
119 print_thread_status();
125 * Happens on thread runs with ctrl-c, ignore our own SIGQUIT
127 static void sig_quit(int sig)
131 static void sig_int(int sig)
134 printf("\nfio: terminating on signal %d\n", sig);
136 terminate_threads(TERMINATE_ALL);
140 static void sig_ill(int fio_unused sig)
145 log_err("fio: illegal instruction. your cpu does not support "
146 "the sse4.2 instruction for crc32c\n");
147 terminate_threads(TERMINATE_ALL);
151 static void set_sig_handlers(void)
153 struct sigaction act;
155 memset(&act, 0, sizeof(act));
156 act.sa_handler = sig_alrm;
157 act.sa_flags = SA_RESTART;
158 sigaction(SIGALRM, &act, NULL);
160 memset(&act, 0, sizeof(act));
161 act.sa_handler = sig_int;
162 act.sa_flags = SA_RESTART;
163 sigaction(SIGINT, &act, NULL);
165 memset(&act, 0, sizeof(act));
166 act.sa_handler = sig_ill;
167 act.sa_flags = SA_RESTART;
168 sigaction(SIGILL, &act, NULL);
170 memset(&act, 0, sizeof(act));
171 act.sa_handler = sig_quit;
172 act.sa_flags = SA_RESTART;
173 sigaction(SIGQUIT, &act, NULL);
177 * Check if we are above the minimum rate given.
179 static int __check_min_rate(struct thread_data *td, struct timeval *now,
182 unsigned long long bytes = 0;
183 unsigned long iops = 0;
186 unsigned int ratemin = 0;
187 unsigned int rate_iops = 0;
188 unsigned int rate_iops_min = 0;
190 if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
194 * allow a 2 second settle period in the beginning
196 if (mtime_since(&td->start, now) < 2000)
199 iops += td->io_blocks[ddir];
200 bytes += td->this_io_bytes[ddir];
201 ratemin += td->o.ratemin[ddir];
202 rate_iops += td->o.rate_iops[ddir];
203 rate_iops_min += td->o.rate_iops_min[ddir];
206 * if rate blocks is set, sample is running
208 if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
209 spent = mtime_since(&td->lastrate[ddir], now);
210 if (spent < td->o.ratecycle)
213 if (td->o.rate[ddir]) {
215 * check bandwidth specified rate
217 if (bytes < td->rate_bytes[ddir]) {
218 log_err("%s: min rate %u not met\n", td->o.name,
222 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
223 if (rate < ratemin ||
224 bytes < td->rate_bytes[ddir]) {
225 log_err("%s: min rate %u not met, got"
226 " %luKB/sec\n", td->o.name,
233 * checks iops specified rate
235 if (iops < rate_iops) {
236 log_err("%s: min iops rate %u not met\n",
237 td->o.name, rate_iops);
240 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
241 if (rate < rate_iops_min ||
242 iops < td->rate_blocks[ddir]) {
243 log_err("%s: min iops rate %u not met,"
244 " got %lu\n", td->o.name,
245 rate_iops_min, rate);
251 td->rate_bytes[ddir] = bytes;
252 td->rate_blocks[ddir] = iops;
253 memcpy(&td->lastrate[ddir], now, sizeof(*now));
257 static int check_min_rate(struct thread_data *td, struct timeval *now,
258 unsigned long *bytes_done)
263 ret |= __check_min_rate(td, now, 0);
265 ret |= __check_min_rate(td, now, 1);
270 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
274 if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
281 * When job exits, we can cancel the in-flight IO if we are using async
282 * io. Attempt to do so.
284 static void cleanup_pending_aio(struct thread_data *td)
286 struct flist_head *entry, *n;
291 * get immediately available events, if any
293 r = io_u_queued_complete(td, 0, NULL);
298 * now cancel remaining active events
300 if (td->io_ops->cancel) {
301 flist_for_each_safe(entry, n, &td->io_u_busylist) {
302 io_u = flist_entry(entry, struct io_u, list);
305 * if the io_u isn't in flight, then that generally
306 * means someone leaked an io_u. complain but fix
307 * it up, so we don't stall here.
309 if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
310 log_err("fio: non-busy IO on busy list\n");
313 r = td->io_ops->cancel(td, io_u);
321 r = io_u_queued_complete(td, td->cur_depth, NULL);
325 * Helper to handle the final sync of a file. Works just like the normal
326 * io path, just does everything sync.
328 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
330 struct io_u *io_u = __get_io_u(td);
336 io_u->ddir = DDIR_SYNC;
339 if (td_io_prep(td, io_u)) {
345 ret = td_io_queue(td, io_u);
347 td_verror(td, io_u->error, "td_io_queue");
350 } else if (ret == FIO_Q_QUEUED) {
351 if (io_u_queued_complete(td, 1, NULL) < 0)
353 } else if (ret == FIO_Q_COMPLETED) {
355 td_verror(td, io_u->error, "td_io_queue");
359 if (io_u_sync_complete(td, io_u, NULL) < 0)
361 } else if (ret == FIO_Q_BUSY) {
362 if (td_io_commit(td))
370 static inline void update_tv_cache(struct thread_data *td)
372 if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
373 fio_gettime(&td->tv_cache, NULL);
376 static int break_on_this_error(struct thread_data *td, int *retptr)
380 if (ret < 0 || td->error) {
383 if (!td->o.continue_on_error)
391 if (td_non_fatal_error(err)) {
393 * Continue with the I/Os in case of
396 update_error_count(td, err);
400 } else if (td->o.fill_device && err == ENOSPC) {
402 * We expect to hit this error if
403 * fill_device option is set.
410 * Stop the I/O in case of a fatal
413 update_error_count(td, err);
422 * The main verify engine. Runs over the writes we previously submitted,
423 * reads the blocks back in, and checks the crc/md5 of the data.
425 static void do_verify(struct thread_data *td)
433 * sync io first and invalidate cache, to make sure we really
436 for_each_file(td, f, i) {
437 if (!fio_file_open(f))
439 if (fio_io_sync(td, f))
441 if (file_invalidate_cache(td, f))
448 td_set_runstate(td, TD_VERIFYING);
451 while (!td->terminate) {
456 if (runtime_exceeded(td, &td->tv_cache)) {
461 io_u = __get_io_u(td);
465 if (get_next_verify(td, io_u)) {
470 if (td_io_prep(td, io_u)) {
475 if (td->o.verify_async)
476 io_u->end_io = verify_io_u_async;
478 io_u->end_io = verify_io_u;
480 ret = td_io_queue(td, io_u);
482 case FIO_Q_COMPLETED:
485 clear_io_u(td, io_u);
486 } else if (io_u->resid) {
487 int bytes = io_u->xfer_buflen - io_u->resid;
488 struct fio_file *f = io_u->file;
494 td_verror(td, EIO, "full resid");
499 io_u->xfer_buflen = io_u->resid;
500 io_u->xfer_buf += bytes;
501 io_u->offset += bytes;
503 td->ts.short_io_u[io_u->ddir]++;
505 if (io_u->offset == f->real_file_size)
508 requeue_io_u(td, &io_u);
511 ret = io_u_sync_complete(td, io_u, NULL);
519 requeue_io_u(td, &io_u);
520 ret2 = td_io_commit(td);
526 td_verror(td, -ret, "td_io_queue");
530 if (break_on_this_error(td, &ret))
534 * if we can queue more, do so. but check if there are
535 * completed io_u's first.
537 full = queue_full(td) || ret == FIO_Q_BUSY;
538 if (full || !td->o.iodepth_batch_complete) {
539 min_events = min(td->o.iodepth_batch_complete,
541 if (full && !min_events)
546 * Reap required number of io units, if any,
547 * and do the verification on them through
548 * the callback handler
550 if (io_u_queued_complete(td, min_events, NULL) < 0) {
554 } while (full && (td->cur_depth > td->o.iodepth_low));
561 min_events = td->cur_depth;
564 ret = io_u_queued_complete(td, min_events, NULL);
566 cleanup_pending_aio(td);
568 td_set_runstate(td, TD_RUNNING);
572 * Main IO worker function. It retrieves io_u's to process and queues
573 * and reaps them, checking for rate and errors along the way.
575 static void do_io(struct thread_data *td)
580 if (in_ramp_time(td))
581 td_set_runstate(td, TD_RAMP);
583 td_set_runstate(td, TD_RUNNING);
585 while ( (td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
586 ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) ) {
587 struct timeval comp_time;
588 unsigned long bytes_done[2] = { 0, 0 };
598 if (runtime_exceeded(td, &td->tv_cache)) {
608 * Add verification end_io handler, if asked to verify
609 * a previously written file.
611 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
613 if (td->o.verify_async)
614 io_u->end_io = verify_io_u_async;
616 io_u->end_io = verify_io_u;
617 td_set_runstate(td, TD_VERIFYING);
618 } else if (in_ramp_time(td))
619 td_set_runstate(td, TD_RAMP);
621 td_set_runstate(td, TD_RUNNING);
623 ret = td_io_queue(td, io_u);
625 case FIO_Q_COMPLETED:
628 clear_io_u(td, io_u);
629 } else if (io_u->resid) {
630 int bytes = io_u->xfer_buflen - io_u->resid;
631 struct fio_file *f = io_u->file;
637 td_verror(td, EIO, "full resid");
642 io_u->xfer_buflen = io_u->resid;
643 io_u->xfer_buf += bytes;
644 io_u->offset += bytes;
646 td->ts.short_io_u[io_u->ddir]++;
648 if (io_u->offset == f->real_file_size)
651 requeue_io_u(td, &io_u);
654 if (__should_check_rate(td, 0) ||
655 __should_check_rate(td, 1))
656 fio_gettime(&comp_time, NULL);
658 ret = io_u_sync_complete(td, io_u, bytes_done);
665 * if the engine doesn't have a commit hook,
666 * the io_u is really queued. if it does have such
667 * a hook, it has to call io_u_queued() itself.
669 if (td->io_ops->commit == NULL)
670 io_u_queued(td, io_u);
673 requeue_io_u(td, &io_u);
674 ret2 = td_io_commit(td);
684 if (break_on_this_error(td, &ret))
688 * See if we need to complete some commands
690 full = queue_full(td) || ret == FIO_Q_BUSY;
691 if (full || !td->o.iodepth_batch_complete) {
692 min_evts = min(td->o.iodepth_batch_complete,
694 if (full && !min_evts)
697 if (__should_check_rate(td, 0) ||
698 __should_check_rate(td, 1))
699 fio_gettime(&comp_time, NULL);
702 ret = io_u_queued_complete(td, min_evts, bytes_done);
706 } while (full && (td->cur_depth > td->o.iodepth_low));
711 if (!(bytes_done[0] + bytes_done[1]))
714 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
715 if (check_min_rate(td, &comp_time, bytes_done)) {
716 if (exitall_on_terminate)
717 terminate_threads(td->groupid);
718 td_verror(td, EIO, "check_min_rate");
723 if (td->o.thinktime) {
724 unsigned long long b;
726 b = td->io_blocks[0] + td->io_blocks[1];
727 if (!(b % td->o.thinktime_blocks)) {
730 if (td->o.thinktime_spin)
731 usec_spin(td->o.thinktime_spin);
733 left = td->o.thinktime - td->o.thinktime_spin;
735 usec_sleep(td, left);
740 if (td->o.fill_device && td->error == ENOSPC) {
749 ret = io_u_queued_complete(td, i, NULL);
751 if (should_fsync(td) && td->o.end_fsync) {
752 td_set_runstate(td, TD_FSYNCING);
754 for_each_file(td, f, i) {
755 if (!fio_file_open(f))
761 cleanup_pending_aio(td);
764 * stop job if we failed doing any IO
766 if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
770 static void cleanup_io_u(struct thread_data *td)
772 struct flist_head *entry, *n;
775 flist_for_each_safe(entry, n, &td->io_u_freelist) {
776 io_u = flist_entry(entry, struct io_u, list);
778 flist_del(&io_u->list);
785 static int init_io_u(struct thread_data *td)
789 int cl_align, i, max_units;
792 max_units = td->o.iodepth;
793 max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
794 td->orig_buffer_size = (unsigned long long) max_bs
795 * (unsigned long long) max_units;
797 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
800 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
801 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
804 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
805 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
809 if (allocate_io_mem(td))
812 if (td->o.odirect || td->o.mem_align)
813 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
817 cl_align = os_cache_line_size();
819 for (i = 0; i < max_units; i++) {
825 if (posix_memalign(&ptr, cl_align, sizeof(*io_u))) {
826 log_err("fio: posix_memalign=%s\n", strerror(errno));
831 memset(io_u, 0, sizeof(*io_u));
832 INIT_FLIST_HEAD(&io_u->list);
833 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
835 if (!(td->io_ops->flags & FIO_NOIO)) {
836 io_u->buf = p + max_bs * i;
837 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
839 if (td_write(td) && !td->o.refill_buffers)
840 io_u_fill_buffer(td, io_u, max_bs);
844 io_u->flags = IO_U_F_FREE;
845 flist_add(&io_u->list, &td->io_u_freelist);
851 static int switch_ioscheduler(struct thread_data *td)
853 char tmp[256], tmp2[128];
857 if (td->io_ops->flags & FIO_DISKLESSIO)
860 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
862 f = fopen(tmp, "r+");
864 if (errno == ENOENT) {
865 log_err("fio: os or kernel doesn't support IO scheduler"
869 td_verror(td, errno, "fopen iosched");
876 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
877 if (ferror(f) || ret != 1) {
878 td_verror(td, errno, "fwrite");
886 * Read back and check that the selected scheduler is now the default.
888 ret = fread(tmp, 1, sizeof(tmp), f);
889 if (ferror(f) || ret < 0) {
890 td_verror(td, errno, "fread");
895 sprintf(tmp2, "[%s]", td->o.ioscheduler);
896 if (!strstr(tmp, tmp2)) {
897 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
898 td_verror(td, EINVAL, "iosched_switch");
907 static int keep_running(struct thread_data *td)
909 unsigned long long io_done;
913 if (td->o.time_based)
920 io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
922 if (io_done < td->o.size)
928 static void reset_io_counters(struct thread_data *td)
930 td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
931 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
933 td->rate_bytes[0] = td->rate_bytes[1] = 0;
934 td->rate_blocks[0] = td->rate_blocks[1] = 0;
936 td->last_was_sync = 0;
939 * reset file done count if we are to start over
941 if (td->o.time_based || td->o.loops)
942 td->nr_done_files = 0;
945 * Set the same seed to get repeatable runs
947 td_fill_rand_seeds(td);
950 void reset_all_stats(struct thread_data *td)
955 reset_io_counters(td);
957 for (i = 0; i < 2; i++) {
959 td->io_blocks[i] = 0;
960 td->io_issues[i] = 0;
961 td->ts.total_io_u[i] = 0;
964 fio_gettime(&tv, NULL);
965 memcpy(&td->epoch, &tv, sizeof(tv));
966 memcpy(&td->start, &tv, sizeof(tv));
969 static void clear_io_state(struct thread_data *td)
974 reset_io_counters(td);
977 for_each_file(td, f, i)
978 fio_file_clear_done(f);
981 static int exec_string(const char *string)
983 int ret, newlen = strlen(string) + 1 + 8;
986 str = malloc(newlen);
987 sprintf(str, "sh -c %s", string);
991 log_err("fio: exec of cmd <%s> failed\n", str);
998 * Entry point for the thread based jobs. The process based jobs end up
999 * here as well, after a little setup.
1001 static void *thread_main(void *data)
1003 unsigned long long runtime[2], elapsed;
1004 struct thread_data *td = data;
1005 pthread_condattr_t attr;
1008 if (!td->o.use_thread)
1013 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1015 INIT_FLIST_HEAD(&td->io_u_freelist);
1016 INIT_FLIST_HEAD(&td->io_u_busylist);
1017 INIT_FLIST_HEAD(&td->io_u_requeues);
1018 INIT_FLIST_HEAD(&td->io_log_list);
1019 INIT_FLIST_HEAD(&td->io_hist_list);
1020 INIT_FLIST_HEAD(&td->verify_list);
1021 pthread_mutex_init(&td->io_u_lock, NULL);
1022 td->io_hist_tree = RB_ROOT;
1024 pthread_condattr_init(&attr);
1025 pthread_cond_init(&td->verify_cond, &attr);
1026 pthread_cond_init(&td->free_cond, &attr);
1028 td_set_runstate(td, TD_INITIALIZED);
1029 dprint(FD_MUTEX, "up startup_mutex\n");
1030 fio_mutex_up(startup_mutex);
1031 dprint(FD_MUTEX, "wait on td->mutex\n");
1032 fio_mutex_down(td->mutex);
1033 dprint(FD_MUTEX, "done waiting on td->mutex\n");
1036 * the ->mutex mutex is now no longer used, close it to avoid
1037 * eating a file descriptor
1039 fio_mutex_remove(td->mutex);
1042 * May alter parameters that init_io_u() will use, so we need to
1051 if (td->o.verify_async && verify_async_init(td))
1054 if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1055 td_verror(td, errno, "cpu_set_affinity");
1060 * If we have a gettimeofday() thread, make sure we exclude that
1061 * thread from this job
1063 if (td->o.gtod_cpu) {
1064 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1065 if (fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1066 td_verror(td, errno, "cpu_set_affinity");
1071 if (td->ioprio_set) {
1072 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1073 td_verror(td, errno, "ioprio_set");
1078 if (nice(td->o.nice) == -1) {
1079 td_verror(td, errno, "nice");
1083 if (td->o.ioscheduler && switch_ioscheduler(td))
1086 if (!td->o.create_serialize && setup_files(td))
1092 if (init_random_map(td))
1095 if (td->o.exec_prerun) {
1096 if (exec_string(td->o.exec_prerun))
1100 if (td->o.pre_read) {
1101 if (pre_read_files(td) < 0)
1105 fio_gettime(&td->epoch, NULL);
1106 getrusage(RUSAGE_SELF, &td->ts.ru_start);
1108 runtime[0] = runtime[1] = 0;
1110 while (keep_running(td)) {
1111 fio_gettime(&td->start, NULL);
1112 memcpy(&td->ts.stat_sample_time[0], &td->start,
1114 memcpy(&td->ts.stat_sample_time[1], &td->start,
1116 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1118 if (td->o.ratemin[0] || td->o.ratemin[1])
1119 memcpy(&td->lastrate, &td->ts.stat_sample_time,
1120 sizeof(td->lastrate));
1125 prune_io_piece_log(td);
1131 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1132 elapsed = utime_since_now(&td->start);
1133 runtime[DDIR_READ] += elapsed;
1135 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1136 elapsed = utime_since_now(&td->start);
1137 runtime[DDIR_WRITE] += elapsed;
1140 if (td->error || td->terminate)
1143 if (!td->o.do_verify ||
1144 td->o.verify == VERIFY_NONE ||
1145 (td->io_ops->flags & FIO_UNIDIR))
1150 fio_gettime(&td->start, NULL);
1154 runtime[DDIR_READ] += utime_since_now(&td->start);
1156 if (td->error || td->terminate)
1160 update_rusage_stat(td);
1161 td->ts.runtime[0] = (runtime[0] + 999) / 1000;
1162 td->ts.runtime[1] = (runtime[1] + 999) / 1000;
1163 td->ts.total_run_time = mtime_since_now(&td->epoch);
1164 td->ts.io_bytes[0] = td->io_bytes[0];
1165 td->ts.io_bytes[1] = td->io_bytes[1];
1167 fio_mutex_down(writeout_mutex);
1168 if (td->ts.bw_log) {
1169 if (td->o.bw_log_file) {
1170 finish_log_named(td, td->ts.bw_log,
1171 td->o.bw_log_file, "bw");
1173 finish_log(td, td->ts.bw_log, "bw");
1175 if (td->ts.slat_log) {
1176 if (td->o.lat_log_file) {
1177 finish_log_named(td, td->ts.slat_log,
1178 td->o.lat_log_file, "slat");
1180 finish_log(td, td->ts.slat_log, "slat");
1182 if (td->ts.clat_log) {
1183 if (td->o.lat_log_file) {
1184 finish_log_named(td, td->ts.clat_log,
1185 td->o.lat_log_file, "clat");
1187 finish_log(td, td->ts.clat_log, "clat");
1189 fio_mutex_up(writeout_mutex);
1190 if (td->o.exec_postrun)
1191 exec_string(td->o.exec_postrun);
1193 if (exitall_on_terminate)
1194 terminate_threads(td->groupid);
1198 printf("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1201 if (td->o.verify_async)
1202 verify_async_exit(td);
1204 close_and_free_files(td);
1208 if (td->o.cpumask_set) {
1209 int ret = fio_cpuset_exit(&td->o.cpumask);
1211 td_verror(td, ret, "fio_cpuset_exit");
1215 * do this very late, it will log file closing as well
1217 if (td->o.write_iolog_file)
1218 write_iolog_close(td);
1220 options_mem_free(td);
1221 td_set_runstate(td, TD_EXITED);
1222 return (void *) (unsigned long) td->error;
1226 * We cannot pass the td data into a forked process, so attach the td and
1227 * pass it to the thread worker.
1229 static int fork_main(int shmid, int offset)
1231 struct thread_data *td;
1234 data = shmat(shmid, NULL, 0);
1235 if (data == (void *) -1) {
1242 td = data + offset * sizeof(struct thread_data);
1243 ret = thread_main(td);
1245 return (int) (unsigned long) ret;
1249 * Run over the job map and reap the threads that have exited, if any.
1251 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1253 struct thread_data *td;
1254 int i, cputhreads, realthreads, pending, status, ret;
1257 * reap exited threads (TD_EXITED -> TD_REAPED)
1259 realthreads = pending = cputhreads = 0;
1260 for_each_td(td, i) {
1264 * ->io_ops is NULL for a thread that has closed its
1267 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1276 if (td->runstate == TD_REAPED)
1278 if (td->o.use_thread) {
1279 if (td->runstate == TD_EXITED) {
1280 td_set_runstate(td, TD_REAPED);
1287 if (td->runstate == TD_EXITED)
1291 * check if someone quit or got killed in an unusual way
1293 ret = waitpid(td->pid, &status, flags);
1295 if (errno == ECHILD) {
1296 log_err("fio: pid=%d disappeared %d\n",
1297 (int) td->pid, td->runstate);
1298 td_set_runstate(td, TD_REAPED);
1302 } else if (ret == td->pid) {
1303 if (WIFSIGNALED(status)) {
1304 int sig = WTERMSIG(status);
1307 log_err("fio: pid=%d, got signal=%d\n",
1308 (int) td->pid, sig);
1309 td_set_runstate(td, TD_REAPED);
1312 if (WIFEXITED(status)) {
1313 if (WEXITSTATUS(status) && !td->error)
1314 td->error = WEXITSTATUS(status);
1316 td_set_runstate(td, TD_REAPED);
1322 * thread is not dead, continue
1328 (*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
1329 (*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
1336 done_secs += mtime_since_now(&td->epoch) / 1000;
1339 if (*nr_running == cputhreads && !pending && realthreads)
1340 terminate_threads(TERMINATE_ALL);
1343 static void *gtod_thread_main(void *data)
1345 fio_mutex_up(startup_mutex);
1348 * As long as we have jobs around, update the clock. It would be nice
1349 * to have some way of NOT hammering that CPU with gettimeofday(),
1350 * but I'm not sure what to use outside of a simple CPU nop to relax
1351 * it - we don't want to lose precision.
1361 static int fio_start_gtod_thread(void)
1365 ret = pthread_create(>od_thread, NULL, gtod_thread_main, NULL);
1367 log_err("Can't create gtod thread: %s\n", strerror(ret));
1371 ret = pthread_detach(gtod_thread);
1373 log_err("Can't detatch gtod thread: %s\n", strerror(ret));
1377 dprint(FD_MUTEX, "wait on startup_mutex\n");
1378 fio_mutex_down(startup_mutex);
1379 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1384 * Main function for kicking off and reaping jobs, as needed.
1386 static void run_threads(void)
1388 struct thread_data *td;
1389 unsigned long spent;
1390 int i, todo, nr_running, m_rate, t_rate, nr_started;
1392 if (fio_pin_memory())
1395 if (fio_gtod_offload && fio_start_gtod_thread())
1398 if (!terse_output) {
1399 printf("Starting ");
1401 printf("%d thread%s", nr_thread,
1402 nr_thread > 1 ? "s" : "");
1406 printf("%d process%s", nr_process,
1407 nr_process > 1 ? "es" : "");
1415 todo = thread_number;
1418 m_rate = t_rate = 0;
1420 for_each_td(td, i) {
1421 print_status_init(td->thread_number - 1);
1423 if (!td->o.create_serialize) {
1429 * do file setup here so it happens sequentially,
1430 * we don't want X number of threads getting their
1431 * client data interspersed on disk
1433 if (setup_files(td)) {
1436 log_err("fio: pid=%d, err=%d/%s\n",
1437 (int) td->pid, td->error, td->verror);
1438 td_set_runstate(td, TD_REAPED);
1445 * for sharing to work, each job must always open
1446 * its own files. so close them, if we opened them
1449 for_each_file(td, f, i) {
1450 if (fio_file_open(f))
1451 td_io_close_file(td, f);
1461 struct thread_data *map[MAX_JOBS];
1462 struct timeval this_start;
1463 int this_jobs = 0, left;
1466 * create threads (TD_NOT_CREATED -> TD_CREATED)
1468 for_each_td(td, i) {
1469 if (td->runstate != TD_NOT_CREATED)
1473 * never got a chance to start, killed by other
1474 * thread for some reason
1476 if (td->terminate) {
1481 if (td->o.start_delay) {
1482 spent = mtime_since_genesis();
1484 if (td->o.start_delay * 1000 > spent)
1488 if (td->o.stonewall && (nr_started || nr_running)) {
1489 dprint(FD_PROCESS, "%s: stonewall wait\n",
1495 * Set state to created. Thread will transition
1496 * to TD_INITIALIZED when it's done setting up.
1498 td_set_runstate(td, TD_CREATED);
1499 map[this_jobs++] = td;
1502 if (td->o.use_thread) {
1505 dprint(FD_PROCESS, "will pthread_create\n");
1506 ret = pthread_create(&td->thread, NULL,
1509 log_err("pthread_create: %s\n",
1514 ret = pthread_detach(td->thread);
1516 log_err("pthread_detach: %s",
1520 dprint(FD_PROCESS, "will fork\n");
1523 int ret = fork_main(shm_id, i);
1526 } else if (i == fio_debug_jobno)
1527 *fio_debug_jobp = pid;
1529 dprint(FD_MUTEX, "wait on startup_mutex\n");
1530 if (fio_mutex_down_timeout(startup_mutex, 10)) {
1531 log_err("fio: job startup hung? exiting.\n");
1532 terminate_threads(TERMINATE_ALL);
1537 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1541 * Wait for the started threads to transition to
1544 fio_gettime(&this_start, NULL);
1546 while (left && !fio_abort) {
1547 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1552 for (i = 0; i < this_jobs; i++) {
1556 if (td->runstate == TD_INITIALIZED) {
1559 } else if (td->runstate >= TD_EXITED) {
1563 nr_running++; /* work-around... */
1569 log_err("fio: %d jobs failed to start\n", left);
1570 for (i = 0; i < this_jobs; i++) {
1574 kill(td->pid, SIGTERM);
1580 * start created threads (TD_INITIALIZED -> TD_RUNNING).
1582 for_each_td(td, i) {
1583 if (td->runstate != TD_INITIALIZED)
1586 if (in_ramp_time(td))
1587 td_set_runstate(td, TD_RAMP);
1589 td_set_runstate(td, TD_RUNNING);
1592 m_rate += td->o.ratemin[0] + td->o.ratemin[1];
1593 t_rate += td->o.rate[0] + td->o.rate[1];
1595 fio_mutex_up(td->mutex);
1598 reap_threads(&nr_running, &t_rate, &m_rate);
1604 while (nr_running) {
1605 reap_threads(&nr_running, &t_rate, &m_rate);
1613 int main(int argc, char *argv[])
1620 * We need locale for number printing, if it isn't set then just
1621 * go with the US format.
1623 if (!getenv("LC_NUMERIC"))
1624 setlocale(LC_NUMERIC, "en_US");
1626 ps = sysconf(_SC_PAGESIZE);
1628 log_err("Failed to get page size\n");
1635 fio_keywords_init();
1637 if (parse_options(argc, argv))
1644 setup_log(&agg_io_log[DDIR_READ]);
1645 setup_log(&agg_io_log[DDIR_WRITE]);
1648 startup_mutex = fio_mutex_init(0);
1649 writeout_mutex = fio_mutex_init(1);
1660 __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1661 __finish_log(agg_io_log[DDIR_WRITE],
1662 "agg-write_bw.log");
1666 fio_mutex_remove(startup_mutex);
1667 fio_mutex_remove(writeout_mutex);