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
48 unsigned long page_mask;
49 unsigned long page_size;
51 #define PAGE_ALIGN(buf) \
52 (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
55 int thread_number = 0;
60 unsigned long done_secs = 0;
62 static struct fio_mutex *startup_mutex;
63 static struct fio_mutex *writeout_mutex;
64 static volatile int fio_abort;
65 static int exit_value;
66 static timer_t ival_timer;
67 static pthread_t gtod_thread;
68 static struct flist_head *cgroup_list;
69 static char *cgroup_mnt;
71 struct io_log *agg_io_log[2];
73 #define TERMINATE_ALL (-1)
74 #define JOB_START_TIMEOUT (5 * 1000)
76 void td_set_runstate(struct thread_data *td, int runstate)
78 if (td->runstate == runstate)
81 dprint(FD_PROCESS, "pid=%d: runstate %d -> %d\n", (int) td->pid,
82 td->runstate, runstate);
83 td->runstate = runstate;
86 static void terminate_threads(int group_id)
88 struct thread_data *td;
91 dprint(FD_PROCESS, "terminate group_id=%d\n", group_id);
94 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
95 dprint(FD_PROCESS, "setting terminate on %s/%d\n",
96 td->o.name, (int) td->pid);
98 td->o.start_delay = 0;
101 * if the thread is running, just let it exit
103 if (td->runstate < TD_RUNNING)
104 kill(td->pid, SIGTERM);
106 struct ioengine_ops *ops = td->io_ops;
108 if (ops && (ops->flags & FIO_SIGTERM))
109 kill(td->pid, SIGTERM);
115 static void status_timer_arm(void)
117 struct itimerspec value;
119 value.it_value.tv_sec = 0;
120 value.it_value.tv_nsec = DISK_UTIL_MSEC * 1000000;
121 value.it_interval.tv_sec = 0;
122 value.it_interval.tv_nsec = DISK_UTIL_MSEC * 1000000;
124 timer_settime(ival_timer, 0, &value, NULL);
127 static void ival_fn(union sigval sig)
131 print_thread_status();
136 * Happens on thread runs with ctrl-c, ignore our own SIGQUIT
138 static void sig_quit(int sig)
142 static void sig_int(int sig)
145 log_info("\nfio: terminating on signal %d\n", sig);
147 terminate_threads(TERMINATE_ALL);
151 static void posix_timer_teardown(void)
153 timer_delete(ival_timer);
156 static void posix_timer_setup(void)
160 memset(&evt, 0, sizeof(evt));
161 evt.sigev_notify = SIGEV_THREAD;
162 evt.sigev_notify_function = ival_fn;
164 if (timer_create(FIO_TIMER_CLOCK, &evt, &ival_timer) < 0)
165 perror("timer_create");
169 static void set_sig_handlers(void)
171 struct sigaction act;
173 memset(&act, 0, sizeof(act));
174 act.sa_handler = sig_int;
175 act.sa_flags = SA_RESTART;
176 sigaction(SIGINT, &act, NULL);
178 memset(&act, 0, sizeof(act));
179 act.sa_handler = sig_quit;
180 act.sa_flags = SA_RESTART;
181 sigaction(SIGTERM, &act, NULL);
185 * Check if we are above the minimum rate given.
187 static int __check_min_rate(struct thread_data *td, struct timeval *now,
190 unsigned long long bytes = 0;
191 unsigned long iops = 0;
194 unsigned int ratemin = 0;
195 unsigned int rate_iops = 0;
196 unsigned int rate_iops_min = 0;
198 assert(ddir_rw(ddir));
200 if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
204 * allow a 2 second settle period in the beginning
206 if (mtime_since(&td->start, now) < 2000)
209 iops += td->io_blocks[ddir];
210 bytes += td->this_io_bytes[ddir];
211 ratemin += td->o.ratemin[ddir];
212 rate_iops += td->o.rate_iops[ddir];
213 rate_iops_min += td->o.rate_iops_min[ddir];
216 * if rate blocks is set, sample is running
218 if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
219 spent = mtime_since(&td->lastrate[ddir], now);
220 if (spent < td->o.ratecycle)
223 if (td->o.rate[ddir]) {
225 * check bandwidth specified rate
227 if (bytes < td->rate_bytes[ddir]) {
228 log_err("%s: min rate %u not met\n", td->o.name,
232 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
233 if (rate < ratemin ||
234 bytes < td->rate_bytes[ddir]) {
235 log_err("%s: min rate %u not met, got"
236 " %luKB/sec\n", td->o.name,
243 * checks iops specified rate
245 if (iops < rate_iops) {
246 log_err("%s: min iops rate %u not met\n",
247 td->o.name, rate_iops);
250 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
251 if (rate < rate_iops_min ||
252 iops < td->rate_blocks[ddir]) {
253 log_err("%s: min iops rate %u not met,"
254 " got %lu\n", td->o.name,
255 rate_iops_min, rate);
261 td->rate_bytes[ddir] = bytes;
262 td->rate_blocks[ddir] = iops;
263 memcpy(&td->lastrate[ddir], now, sizeof(*now));
267 static int check_min_rate(struct thread_data *td, struct timeval *now,
268 unsigned long *bytes_done)
273 ret |= __check_min_rate(td, now, 0);
275 ret |= __check_min_rate(td, now, 1);
280 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
284 if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
291 * When job exits, we can cancel the in-flight IO if we are using async
292 * io. Attempt to do so.
294 static void cleanup_pending_aio(struct thread_data *td)
296 struct flist_head *entry, *n;
301 * get immediately available events, if any
303 r = io_u_queued_complete(td, 0, NULL);
308 * now cancel remaining active events
310 if (td->io_ops->cancel) {
311 flist_for_each_safe(entry, n, &td->io_u_busylist) {
312 io_u = flist_entry(entry, struct io_u, list);
315 * if the io_u isn't in flight, then that generally
316 * means someone leaked an io_u. complain but fix
317 * it up, so we don't stall here.
319 if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
320 log_err("fio: non-busy IO on busy list\n");
323 r = td->io_ops->cancel(td, io_u);
331 r = io_u_queued_complete(td, td->cur_depth, NULL);
335 * Helper to handle the final sync of a file. Works just like the normal
336 * io path, just does everything sync.
338 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
340 struct io_u *io_u = __get_io_u(td);
346 io_u->ddir = DDIR_SYNC;
349 if (td_io_prep(td, io_u)) {
355 ret = td_io_queue(td, io_u);
357 td_verror(td, io_u->error, "td_io_queue");
360 } else if (ret == FIO_Q_QUEUED) {
361 if (io_u_queued_complete(td, 1, NULL) < 0)
363 } else if (ret == FIO_Q_COMPLETED) {
365 td_verror(td, io_u->error, "td_io_queue");
369 if (io_u_sync_complete(td, io_u, NULL) < 0)
371 } else if (ret == FIO_Q_BUSY) {
372 if (td_io_commit(td))
380 static inline void update_tv_cache(struct thread_data *td)
382 if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
383 fio_gettime(&td->tv_cache, NULL);
386 static int break_on_this_error(struct thread_data *td, int *retptr)
390 if (ret < 0 || td->error) {
393 if (!td->o.continue_on_error)
401 if (td_non_fatal_error(err)) {
403 * Continue with the I/Os in case of
406 update_error_count(td, err);
410 } else if (td->o.fill_device && err == ENOSPC) {
412 * We expect to hit this error if
413 * fill_device option is set.
420 * Stop the I/O in case of a fatal
423 update_error_count(td, err);
432 * The main verify engine. Runs over the writes we previously submitted,
433 * reads the blocks back in, and checks the crc/md5 of the data.
435 static void do_verify(struct thread_data *td)
442 dprint(FD_VERIFY, "starting loop\n");
445 * sync io first and invalidate cache, to make sure we really
448 for_each_file(td, f, i) {
449 if (!fio_file_open(f))
451 if (fio_io_sync(td, f))
453 if (file_invalidate_cache(td, f))
460 td_set_runstate(td, TD_VERIFYING);
463 while (!td->terminate) {
468 if (runtime_exceeded(td, &td->tv_cache)) {
473 io_u = __get_io_u(td);
477 if (get_next_verify(td, io_u)) {
482 if (td_io_prep(td, io_u)) {
487 if (td->o.verify_async)
488 io_u->end_io = verify_io_u_async;
490 io_u->end_io = verify_io_u;
492 ret = td_io_queue(td, io_u);
494 case FIO_Q_COMPLETED:
497 clear_io_u(td, io_u);
498 } else if (io_u->resid) {
499 int bytes = io_u->xfer_buflen - io_u->resid;
500 struct fio_file *f = io_u->file;
506 td_verror(td, EIO, "full resid");
511 io_u->xfer_buflen = io_u->resid;
512 io_u->xfer_buf += bytes;
513 io_u->offset += bytes;
515 if (ddir_rw(io_u->ddir))
516 td->ts.short_io_u[io_u->ddir]++;
518 if (io_u->offset == f->real_file_size)
521 requeue_io_u(td, &io_u);
524 ret = io_u_sync_complete(td, io_u, NULL);
532 requeue_io_u(td, &io_u);
533 ret2 = td_io_commit(td);
539 td_verror(td, -ret, "td_io_queue");
543 if (break_on_this_error(td, &ret))
547 * if we can queue more, do so. but check if there are
548 * completed io_u's first.
550 full = queue_full(td) || ret == FIO_Q_BUSY;
551 if (full || !td->o.iodepth_batch_complete) {
552 min_events = min(td->o.iodepth_batch_complete,
554 if (full && !min_events)
559 * Reap required number of io units, if any,
560 * and do the verification on them through
561 * the callback handler
563 if (io_u_queued_complete(td, min_events, NULL) < 0) {
567 } while (full && (td->cur_depth > td->o.iodepth_low));
574 min_events = td->cur_depth;
577 ret = io_u_queued_complete(td, min_events, NULL);
579 cleanup_pending_aio(td);
581 td_set_runstate(td, TD_RUNNING);
583 dprint(FD_VERIFY, "exiting loop\n");
587 * Main IO worker function. It retrieves io_u's to process and queues
588 * and reaps them, checking for rate and errors along the way.
590 static void do_io(struct thread_data *td)
595 if (in_ramp_time(td))
596 td_set_runstate(td, TD_RAMP);
598 td_set_runstate(td, TD_RUNNING);
600 while ( (td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
601 (!flist_empty(&td->trim_list)) ||
602 ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) ) {
603 struct timeval comp_time;
604 unsigned long bytes_done[2] = { 0, 0 };
614 if (runtime_exceeded(td, &td->tv_cache)) {
624 * Add verification end_io handler, if asked to verify
625 * a previously written file.
627 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
629 if (td->o.verify_async)
630 io_u->end_io = verify_io_u_async;
632 io_u->end_io = verify_io_u;
633 td_set_runstate(td, TD_VERIFYING);
634 } else if (in_ramp_time(td))
635 td_set_runstate(td, TD_RAMP);
637 td_set_runstate(td, TD_RUNNING);
639 ret = td_io_queue(td, io_u);
641 case FIO_Q_COMPLETED:
644 clear_io_u(td, io_u);
645 } else if (io_u->resid) {
646 int bytes = io_u->xfer_buflen - io_u->resid;
647 struct fio_file *f = io_u->file;
653 td_verror(td, EIO, "full resid");
658 io_u->xfer_buflen = io_u->resid;
659 io_u->xfer_buf += bytes;
660 io_u->offset += bytes;
662 if (ddir_rw(io_u->ddir))
663 td->ts.short_io_u[io_u->ddir]++;
665 if (io_u->offset == f->real_file_size)
668 requeue_io_u(td, &io_u);
671 if (__should_check_rate(td, 0) ||
672 __should_check_rate(td, 1))
673 fio_gettime(&comp_time, NULL);
675 ret = io_u_sync_complete(td, io_u, bytes_done);
682 * if the engine doesn't have a commit hook,
683 * the io_u is really queued. if it does have such
684 * a hook, it has to call io_u_queued() itself.
686 if (td->io_ops->commit == NULL)
687 io_u_queued(td, io_u);
690 requeue_io_u(td, &io_u);
691 ret2 = td_io_commit(td);
701 if (break_on_this_error(td, &ret))
705 * See if we need to complete some commands
707 full = queue_full(td) || ret == FIO_Q_BUSY;
708 if (full || !td->o.iodepth_batch_complete) {
709 min_evts = min(td->o.iodepth_batch_complete,
711 if (full && !min_evts)
714 if (__should_check_rate(td, 0) ||
715 __should_check_rate(td, 1))
716 fio_gettime(&comp_time, NULL);
719 ret = io_u_queued_complete(td, min_evts, bytes_done);
723 } while (full && (td->cur_depth > td->o.iodepth_low));
728 if (!(bytes_done[0] + bytes_done[1]))
731 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
732 if (check_min_rate(td, &comp_time, bytes_done)) {
733 if (exitall_on_terminate)
734 terminate_threads(td->groupid);
735 td_verror(td, EIO, "check_min_rate");
740 if (td->o.thinktime) {
741 unsigned long long b;
743 b = td->io_blocks[0] + td->io_blocks[1];
744 if (!(b % td->o.thinktime_blocks)) {
747 if (td->o.thinktime_spin)
748 usec_spin(td->o.thinktime_spin);
750 left = td->o.thinktime - td->o.thinktime_spin;
752 usec_sleep(td, left);
757 if (td->trim_entries)
758 printf("trim entries %ld\n", td->trim_entries);
760 if (td->o.fill_device && td->error == ENOSPC) {
769 ret = io_u_queued_complete(td, i, NULL);
771 if (should_fsync(td) && td->o.end_fsync) {
772 td_set_runstate(td, TD_FSYNCING);
774 for_each_file(td, f, i) {
775 if (!fio_file_open(f))
781 cleanup_pending_aio(td);
784 * stop job if we failed doing any IO
786 if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
790 static void cleanup_io_u(struct thread_data *td)
792 struct flist_head *entry, *n;
795 flist_for_each_safe(entry, n, &td->io_u_freelist) {
796 io_u = flist_entry(entry, struct io_u, list);
798 flist_del(&io_u->list);
805 static int init_io_u(struct thread_data *td)
809 int cl_align, i, max_units;
812 max_units = td->o.iodepth;
813 max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
814 td->orig_buffer_size = (unsigned long long) max_bs
815 * (unsigned long long) max_units;
817 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
820 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
821 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
824 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
825 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
829 if (allocate_io_mem(td))
832 if (td->o.odirect || td->o.mem_align ||
833 (td->io_ops->flags & FIO_RAWIO))
834 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
838 cl_align = os_cache_line_size();
840 for (i = 0; i < max_units; i++) {
846 if (posix_memalign(&ptr, cl_align, sizeof(*io_u))) {
847 log_err("fio: posix_memalign=%s\n", strerror(errno));
852 memset(io_u, 0, sizeof(*io_u));
853 INIT_FLIST_HEAD(&io_u->list);
854 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
856 if (!(td->io_ops->flags & FIO_NOIO)) {
857 io_u->buf = p + max_bs * i;
858 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
860 if (td_write(td) && !td->o.refill_buffers)
861 io_u_fill_buffer(td, io_u, max_bs);
862 else if (td_write(td) && td->o.verify_pattern_bytes) {
864 * Fill the buffer with the pattern if we are
865 * going to be doing writes.
867 fill_pattern(td, io_u->buf, max_bs, io_u);
872 io_u->flags = IO_U_F_FREE;
873 flist_add(&io_u->list, &td->io_u_freelist);
879 static int switch_ioscheduler(struct thread_data *td)
881 char tmp[256], tmp2[128];
885 if (td->io_ops->flags & FIO_DISKLESSIO)
888 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
890 f = fopen(tmp, "r+");
892 if (errno == ENOENT) {
893 log_err("fio: os or kernel doesn't support IO scheduler"
897 td_verror(td, errno, "fopen iosched");
904 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
905 if (ferror(f) || ret != 1) {
906 td_verror(td, errno, "fwrite");
914 * Read back and check that the selected scheduler is now the default.
916 ret = fread(tmp, 1, sizeof(tmp), f);
917 if (ferror(f) || ret < 0) {
918 td_verror(td, errno, "fread");
923 sprintf(tmp2, "[%s]", td->o.ioscheduler);
924 if (!strstr(tmp, tmp2)) {
925 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
926 td_verror(td, EINVAL, "iosched_switch");
935 static int keep_running(struct thread_data *td)
937 unsigned long long io_done;
941 if (td->o.time_based)
948 io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
950 if (io_done < td->o.size)
956 static void reset_io_counters(struct thread_data *td)
958 td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
959 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
961 td->rate_bytes[0] = td->rate_bytes[1] = 0;
962 td->rate_blocks[0] = td->rate_blocks[1] = 0;
964 td->last_was_sync = 0;
967 * reset file done count if we are to start over
969 if (td->o.time_based || td->o.loops)
970 td->nr_done_files = 0;
973 * Set the same seed to get repeatable runs
975 td_fill_rand_seeds(td);
978 void reset_all_stats(struct thread_data *td)
983 reset_io_counters(td);
985 for (i = 0; i < 2; i++) {
987 td->io_blocks[i] = 0;
988 td->io_issues[i] = 0;
989 td->ts.total_io_u[i] = 0;
992 fio_gettime(&tv, NULL);
993 td->ts.runtime[0] = 0;
994 td->ts.runtime[1] = 0;
995 memcpy(&td->epoch, &tv, sizeof(tv));
996 memcpy(&td->start, &tv, sizeof(tv));
999 static void clear_io_state(struct thread_data *td)
1004 reset_io_counters(td);
1007 for_each_file(td, f, i)
1008 fio_file_clear_done(f);
1011 static int exec_string(const char *string)
1013 int ret, newlen = strlen(string) + 1 + 8;
1016 str = malloc(newlen);
1017 sprintf(str, "sh -c %s", string);
1021 log_err("fio: exec of cmd <%s> failed\n", str);
1028 * Entry point for the thread based jobs. The process based jobs end up
1029 * here as well, after a little setup.
1031 static void *thread_main(void *data)
1033 unsigned long long elapsed;
1034 struct thread_data *td = data;
1035 pthread_condattr_t attr;
1038 if (!td->o.use_thread)
1043 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1045 INIT_FLIST_HEAD(&td->io_u_freelist);
1046 INIT_FLIST_HEAD(&td->io_u_busylist);
1047 INIT_FLIST_HEAD(&td->io_u_requeues);
1048 INIT_FLIST_HEAD(&td->io_log_list);
1049 INIT_FLIST_HEAD(&td->io_hist_list);
1050 INIT_FLIST_HEAD(&td->verify_list);
1051 INIT_FLIST_HEAD(&td->trim_list);
1052 pthread_mutex_init(&td->io_u_lock, NULL);
1053 td->io_hist_tree = RB_ROOT;
1055 pthread_condattr_init(&attr);
1056 pthread_cond_init(&td->verify_cond, &attr);
1057 pthread_cond_init(&td->free_cond, &attr);
1059 td_set_runstate(td, TD_INITIALIZED);
1060 dprint(FD_MUTEX, "up startup_mutex\n");
1061 fio_mutex_up(startup_mutex);
1062 dprint(FD_MUTEX, "wait on td->mutex\n");
1063 fio_mutex_down(td->mutex);
1064 dprint(FD_MUTEX, "done waiting on td->mutex\n");
1067 * the ->mutex mutex is now no longer used, close it to avoid
1068 * eating a file descriptor
1070 fio_mutex_remove(td->mutex);
1072 if (td->o.uid != -1U && setuid(td->o.uid)) {
1073 td_verror(td, errno, "setuid");
1076 if (td->o.gid != -1U && setgid(td->o.gid)) {
1077 td_verror(td, errno, "setgid");
1082 * May alter parameters that init_io_u() will use, so we need to
1091 if (td->o.verify_async && verify_async_init(td))
1094 if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1095 td_verror(td, errno, "cpu_set_affinity");
1100 * If we have a gettimeofday() thread, make sure we exclude that
1101 * thread from this job
1103 if (td->o.gtod_cpu) {
1104 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1105 if (fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1106 td_verror(td, errno, "cpu_set_affinity");
1111 if (td->ioprio_set) {
1112 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1113 td_verror(td, errno, "ioprio_set");
1118 if (td->o.cgroup_weight && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1121 if (nice(td->o.nice) == -1) {
1122 td_verror(td, errno, "nice");
1126 if (td->o.ioscheduler && switch_ioscheduler(td))
1129 if (!td->o.create_serialize && setup_files(td))
1135 if (init_random_map(td))
1138 if (td->o.exec_prerun) {
1139 if (exec_string(td->o.exec_prerun))
1143 if (td->o.pre_read) {
1144 if (pre_read_files(td) < 0)
1148 fio_gettime(&td->epoch, NULL);
1149 getrusage(RUSAGE_SELF, &td->ts.ru_start);
1152 while (keep_running(td)) {
1153 fio_gettime(&td->start, NULL);
1154 memcpy(&td->ts.stat_sample_time[0], &td->start,
1156 memcpy(&td->ts.stat_sample_time[1], &td->start,
1158 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1160 if (td->o.ratemin[0] || td->o.ratemin[1])
1161 memcpy(&td->lastrate, &td->ts.stat_sample_time,
1162 sizeof(td->lastrate));
1167 prune_io_piece_log(td);
1173 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1174 elapsed = utime_since_now(&td->start);
1175 td->ts.runtime[DDIR_READ] += elapsed;
1177 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1178 elapsed = utime_since_now(&td->start);
1179 td->ts.runtime[DDIR_WRITE] += elapsed;
1182 if (td->error || td->terminate)
1185 if (!td->o.do_verify ||
1186 td->o.verify == VERIFY_NONE ||
1187 (td->io_ops->flags & FIO_UNIDIR))
1192 fio_gettime(&td->start, NULL);
1196 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1198 if (td->error || td->terminate)
1202 update_rusage_stat(td);
1203 td->ts.runtime[0] = (td->ts.runtime[0] + 999) / 1000;
1204 td->ts.runtime[1] = (td->ts.runtime[1] + 999) / 1000;
1205 td->ts.total_run_time = mtime_since_now(&td->epoch);
1206 td->ts.io_bytes[0] = td->io_bytes[0];
1207 td->ts.io_bytes[1] = td->io_bytes[1];
1209 fio_mutex_down(writeout_mutex);
1210 if (td->ts.bw_log) {
1211 if (td->o.bw_log_file) {
1212 finish_log_named(td, td->ts.bw_log,
1213 td->o.bw_log_file, "bw");
1215 finish_log(td, td->ts.bw_log, "bw");
1217 if (td->ts.lat_log) {
1218 if (td->o.lat_log_file) {
1219 finish_log_named(td, td->ts.lat_log,
1220 td->o.lat_log_file, "lat");
1222 finish_log(td, td->ts.lat_log, "lat");
1224 if (td->ts.slat_log) {
1225 if (td->o.lat_log_file) {
1226 finish_log_named(td, td->ts.slat_log,
1227 td->o.lat_log_file, "slat");
1229 finish_log(td, td->ts.slat_log, "slat");
1231 if (td->ts.clat_log) {
1232 if (td->o.lat_log_file) {
1233 finish_log_named(td, td->ts.clat_log,
1234 td->o.lat_log_file, "clat");
1236 finish_log(td, td->ts.clat_log, "clat");
1238 fio_mutex_up(writeout_mutex);
1239 if (td->o.exec_postrun)
1240 exec_string(td->o.exec_postrun);
1242 if (exitall_on_terminate)
1243 terminate_threads(td->groupid);
1247 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1250 if (td->o.verify_async)
1251 verify_async_exit(td);
1253 close_and_free_files(td);
1256 cgroup_shutdown(td, &cgroup_mnt);
1258 if (td->o.cpumask_set) {
1259 int ret = fio_cpuset_exit(&td->o.cpumask);
1261 td_verror(td, ret, "fio_cpuset_exit");
1265 * do this very late, it will log file closing as well
1267 if (td->o.write_iolog_file)
1268 write_iolog_close(td);
1270 options_mem_free(td);
1271 td_set_runstate(td, TD_EXITED);
1272 return (void *) (unsigned long) td->error;
1276 * We cannot pass the td data into a forked process, so attach the td and
1277 * pass it to the thread worker.
1279 static int fork_main(int shmid, int offset)
1281 struct thread_data *td;
1284 data = shmat(shmid, NULL, 0);
1285 if (data == (void *) -1) {
1292 td = data + offset * sizeof(struct thread_data);
1293 ret = thread_main(td);
1295 return (int) (unsigned long) ret;
1299 * Run over the job map and reap the threads that have exited, if any.
1301 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1303 struct thread_data *td;
1304 int i, cputhreads, realthreads, pending, status, ret;
1307 * reap exited threads (TD_EXITED -> TD_REAPED)
1309 realthreads = pending = cputhreads = 0;
1310 for_each_td(td, i) {
1314 * ->io_ops is NULL for a thread that has closed its
1317 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1326 if (td->runstate == TD_REAPED)
1328 if (td->o.use_thread) {
1329 if (td->runstate == TD_EXITED) {
1330 td_set_runstate(td, TD_REAPED);
1337 if (td->runstate == TD_EXITED)
1341 * check if someone quit or got killed in an unusual way
1343 ret = waitpid(td->pid, &status, flags);
1345 if (errno == ECHILD) {
1346 log_err("fio: pid=%d disappeared %d\n",
1347 (int) td->pid, td->runstate);
1348 td_set_runstate(td, TD_REAPED);
1352 } else if (ret == td->pid) {
1353 if (WIFSIGNALED(status)) {
1354 int sig = WTERMSIG(status);
1357 log_err("fio: pid=%d, got signal=%d\n",
1358 (int) td->pid, sig);
1359 td_set_runstate(td, TD_REAPED);
1362 if (WIFEXITED(status)) {
1363 if (WEXITSTATUS(status) && !td->error)
1364 td->error = WEXITSTATUS(status);
1366 td_set_runstate(td, TD_REAPED);
1372 * thread is not dead, continue
1378 (*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
1379 (*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
1386 done_secs += mtime_since_now(&td->epoch) / 1000;
1389 if (*nr_running == cputhreads && !pending && realthreads)
1390 terminate_threads(TERMINATE_ALL);
1393 static void *gtod_thread_main(void *data)
1395 fio_mutex_up(startup_mutex);
1398 * As long as we have jobs around, update the clock. It would be nice
1399 * to have some way of NOT hammering that CPU with gettimeofday(),
1400 * but I'm not sure what to use outside of a simple CPU nop to relax
1401 * it - we don't want to lose precision.
1411 static int fio_start_gtod_thread(void)
1413 pthread_attr_t attr;
1416 pthread_attr_init(&attr);
1417 pthread_attr_setstacksize(&attr, PTHREAD_STACK_MIN);
1418 ret = pthread_create(>od_thread, &attr, gtod_thread_main, NULL);
1419 pthread_attr_destroy(&attr);
1421 log_err("Can't create gtod thread: %s\n", strerror(ret));
1425 ret = pthread_detach(gtod_thread);
1427 log_err("Can't detatch gtod thread: %s\n", strerror(ret));
1431 dprint(FD_MUTEX, "wait on startup_mutex\n");
1432 fio_mutex_down(startup_mutex);
1433 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1438 * Main function for kicking off and reaping jobs, as needed.
1440 static void run_threads(void)
1442 struct thread_data *td;
1443 unsigned long spent;
1444 int i, todo, nr_running, m_rate, t_rate, nr_started;
1446 if (fio_pin_memory())
1449 if (fio_gtod_offload && fio_start_gtod_thread())
1452 if (!terse_output) {
1453 log_info("Starting ");
1455 log_info("%d thread%s", nr_thread,
1456 nr_thread > 1 ? "s" : "");
1460 log_info("%d process%s", nr_process,
1461 nr_process > 1 ? "es" : "");
1469 todo = thread_number;
1472 m_rate = t_rate = 0;
1474 for_each_td(td, i) {
1475 print_status_init(td->thread_number - 1);
1477 if (!td->o.create_serialize) {
1483 * do file setup here so it happens sequentially,
1484 * we don't want X number of threads getting their
1485 * client data interspersed on disk
1487 if (setup_files(td)) {
1490 log_err("fio: pid=%d, err=%d/%s\n",
1491 (int) td->pid, td->error, td->verror);
1492 td_set_runstate(td, TD_REAPED);
1499 * for sharing to work, each job must always open
1500 * its own files. so close them, if we opened them
1503 for_each_file(td, f, i) {
1504 if (fio_file_open(f))
1505 td_io_close_file(td, f);
1515 struct thread_data *map[MAX_JOBS];
1516 struct timeval this_start;
1517 int this_jobs = 0, left;
1520 * create threads (TD_NOT_CREATED -> TD_CREATED)
1522 for_each_td(td, i) {
1523 if (td->runstate != TD_NOT_CREATED)
1527 * never got a chance to start, killed by other
1528 * thread for some reason
1530 if (td->terminate) {
1535 if (td->o.start_delay) {
1536 spent = mtime_since_genesis();
1538 if (td->o.start_delay * 1000 > spent)
1542 if (td->o.stonewall && (nr_started || nr_running)) {
1543 dprint(FD_PROCESS, "%s: stonewall wait\n",
1549 * Set state to created. Thread will transition
1550 * to TD_INITIALIZED when it's done setting up.
1552 td_set_runstate(td, TD_CREATED);
1553 map[this_jobs++] = td;
1556 if (td->o.use_thread) {
1559 dprint(FD_PROCESS, "will pthread_create\n");
1560 ret = pthread_create(&td->thread, NULL,
1563 log_err("pthread_create: %s\n",
1568 ret = pthread_detach(td->thread);
1570 log_err("pthread_detach: %s",
1574 dprint(FD_PROCESS, "will fork\n");
1577 int ret = fork_main(shm_id, i);
1580 } else if (i == fio_debug_jobno)
1581 *fio_debug_jobp = pid;
1583 dprint(FD_MUTEX, "wait on startup_mutex\n");
1584 if (fio_mutex_down_timeout(startup_mutex, 10)) {
1585 log_err("fio: job startup hung? exiting.\n");
1586 terminate_threads(TERMINATE_ALL);
1591 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1595 * Wait for the started threads to transition to
1598 fio_gettime(&this_start, NULL);
1600 while (left && !fio_abort) {
1601 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1606 for (i = 0; i < this_jobs; i++) {
1610 if (td->runstate == TD_INITIALIZED) {
1613 } else if (td->runstate >= TD_EXITED) {
1617 nr_running++; /* work-around... */
1623 log_err("fio: %d jobs failed to start\n", left);
1624 for (i = 0; i < this_jobs; i++) {
1628 kill(td->pid, SIGTERM);
1634 * start created threads (TD_INITIALIZED -> TD_RUNNING).
1636 for_each_td(td, i) {
1637 if (td->runstate != TD_INITIALIZED)
1640 if (in_ramp_time(td))
1641 td_set_runstate(td, TD_RAMP);
1643 td_set_runstate(td, TD_RUNNING);
1646 m_rate += td->o.ratemin[0] + td->o.ratemin[1];
1647 t_rate += td->o.rate[0] + td->o.rate[1];
1649 fio_mutex_up(td->mutex);
1652 reap_threads(&nr_running, &t_rate, &m_rate);
1658 while (nr_running) {
1659 reap_threads(&nr_running, &t_rate, &m_rate);
1667 int main(int argc, char *argv[])
1672 init_rand(&__fio_rand_state);
1675 * We need locale for number printing, if it isn't set then just
1676 * go with the US format.
1678 if (!getenv("LC_NUMERIC"))
1679 setlocale(LC_NUMERIC, "en_US");
1681 ps = sysconf(_SC_PAGESIZE);
1683 log_err("Failed to get page size\n");
1690 fio_keywords_init();
1692 if (parse_options(argc, argv))
1695 if (exec_profile && load_profile(exec_profile))
1702 setup_log(&agg_io_log[DDIR_READ]);
1703 setup_log(&agg_io_log[DDIR_WRITE]);
1706 startup_mutex = fio_mutex_init(0);
1707 if (startup_mutex == NULL)
1709 writeout_mutex = fio_mutex_init(1);
1710 if (writeout_mutex == NULL)
1715 posix_timer_setup();
1718 cgroup_list = smalloc(sizeof(*cgroup_list));
1719 INIT_FLIST_HEAD(cgroup_list);
1726 __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1727 __finish_log(agg_io_log[DDIR_WRITE],
1728 "agg-write_bw.log");
1732 cgroup_kill(cgroup_list);
1736 posix_timer_teardown();
1737 fio_mutex_remove(startup_mutex);
1738 fio_mutex_remove(writeout_mutex);