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
50 unsigned long page_mask;
51 unsigned long page_size;
53 #define PAGE_ALIGN(buf) \
54 (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
57 int thread_number = 0;
62 unsigned long done_secs = 0;
64 static struct fio_mutex *startup_mutex;
65 static struct fio_mutex *writeout_mutex;
66 static volatile int fio_abort;
67 static int exit_value;
68 static pthread_t gtod_thread;
69 static pthread_t disk_util_thread;
70 static struct flist_head *cgroup_list;
71 static char *cgroup_mnt;
73 struct io_log *agg_io_log[2];
75 #define TERMINATE_ALL (-1)
76 #define JOB_START_TIMEOUT (5 * 1000)
78 void td_set_runstate(struct thread_data *td, int runstate)
80 if (td->runstate == runstate)
83 dprint(FD_PROCESS, "pid=%d: runstate %d -> %d\n", (int) td->pid,
84 td->runstate, runstate);
85 td->runstate = runstate;
88 static void terminate_threads(int group_id)
90 struct thread_data *td;
93 dprint(FD_PROCESS, "terminate group_id=%d\n", group_id);
96 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
97 dprint(FD_PROCESS, "setting terminate on %s/%d\n",
98 td->o.name, (int) td->pid);
100 td->o.start_delay = 0;
103 * if the thread is running, just let it exit
107 else if (td->runstate < TD_RAMP)
108 kill(td->pid, SIGTERM);
110 struct ioengine_ops *ops = td->io_ops;
112 if (ops && (ops->flags & FIO_SIGTERM))
113 kill(td->pid, SIGTERM);
119 static void sig_int(int sig)
122 log_info("\nfio: terminating on signal %d\n", sig);
125 terminate_threads(TERMINATE_ALL);
129 static void *disk_thread_main(void *data)
131 fio_mutex_up(startup_mutex);
134 usleep(DISK_UTIL_MSEC * 1000);
138 print_thread_status();
144 static int create_disk_util_thread(void)
148 ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
150 log_err("Can't create disk util thread: %s\n", strerror(ret));
154 ret = pthread_detach(disk_util_thread);
156 log_err("Can't detatch disk util thread: %s\n", strerror(ret));
160 dprint(FD_MUTEX, "wait on startup_mutex\n");
161 fio_mutex_down(startup_mutex);
162 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
166 static void set_sig_handlers(void)
168 struct sigaction act;
170 memset(&act, 0, sizeof(act));
171 act.sa_handler = sig_int;
172 act.sa_flags = SA_RESTART;
173 sigaction(SIGINT, &act, NULL);
175 memset(&act, 0, sizeof(act));
176 act.sa_handler = sig_int;
177 act.sa_flags = SA_RESTART;
178 sigaction(SIGTERM, &act, NULL);
182 * Check if we are above the minimum rate given.
184 static int __check_min_rate(struct thread_data *td, struct timeval *now,
187 unsigned long long bytes = 0;
188 unsigned long iops = 0;
191 unsigned int ratemin = 0;
192 unsigned int rate_iops = 0;
193 unsigned int rate_iops_min = 0;
195 assert(ddir_rw(ddir));
197 if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
201 * allow a 2 second settle period in the beginning
203 if (mtime_since(&td->start, now) < 2000)
206 iops += td->io_blocks[ddir];
207 bytes += td->this_io_bytes[ddir];
208 ratemin += td->o.ratemin[ddir];
209 rate_iops += td->o.rate_iops[ddir];
210 rate_iops_min += td->o.rate_iops_min[ddir];
213 * if rate blocks is set, sample is running
215 if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
216 spent = mtime_since(&td->lastrate[ddir], now);
217 if (spent < td->o.ratecycle)
220 if (td->o.rate[ddir]) {
222 * check bandwidth specified rate
224 if (bytes < td->rate_bytes[ddir]) {
225 log_err("%s: min rate %u not met\n", td->o.name,
229 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
230 if (rate < ratemin ||
231 bytes < td->rate_bytes[ddir]) {
232 log_err("%s: min rate %u not met, got"
233 " %luKB/sec\n", td->o.name,
240 * checks iops specified rate
242 if (iops < rate_iops) {
243 log_err("%s: min iops rate %u not met\n",
244 td->o.name, rate_iops);
247 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
248 if (rate < rate_iops_min ||
249 iops < td->rate_blocks[ddir]) {
250 log_err("%s: min iops rate %u not met,"
251 " got %lu\n", td->o.name,
252 rate_iops_min, rate);
258 td->rate_bytes[ddir] = bytes;
259 td->rate_blocks[ddir] = iops;
260 memcpy(&td->lastrate[ddir], now, sizeof(*now));
264 static int check_min_rate(struct thread_data *td, struct timeval *now,
265 unsigned long *bytes_done)
270 ret |= __check_min_rate(td, now, 0);
272 ret |= __check_min_rate(td, now, 1);
277 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
281 if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
288 * When job exits, we can cancel the in-flight IO if we are using async
289 * io. Attempt to do so.
291 static void cleanup_pending_aio(struct thread_data *td)
293 struct flist_head *entry, *n;
298 * get immediately available events, if any
300 r = io_u_queued_complete(td, 0, NULL);
305 * now cancel remaining active events
307 if (td->io_ops->cancel) {
308 flist_for_each_safe(entry, n, &td->io_u_busylist) {
309 io_u = flist_entry(entry, struct io_u, list);
312 * if the io_u isn't in flight, then that generally
313 * means someone leaked an io_u. complain but fix
314 * it up, so we don't stall here.
316 if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
317 log_err("fio: non-busy IO on busy list\n");
320 r = td->io_ops->cancel(td, io_u);
328 r = io_u_queued_complete(td, td->cur_depth, NULL);
332 * Helper to handle the final sync of a file. Works just like the normal
333 * io path, just does everything sync.
335 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
337 struct io_u *io_u = __get_io_u(td);
343 io_u->ddir = DDIR_SYNC;
346 if (td_io_prep(td, io_u)) {
352 ret = td_io_queue(td, io_u);
354 td_verror(td, io_u->error, "td_io_queue");
357 } else if (ret == FIO_Q_QUEUED) {
358 if (io_u_queued_complete(td, 1, NULL) < 0)
360 } else if (ret == FIO_Q_COMPLETED) {
362 td_verror(td, io_u->error, "td_io_queue");
366 if (io_u_sync_complete(td, io_u, NULL) < 0)
368 } else if (ret == FIO_Q_BUSY) {
369 if (td_io_commit(td))
377 static inline void __update_tv_cache(struct thread_data *td)
379 fio_gettime(&td->tv_cache, NULL);
382 static inline void update_tv_cache(struct thread_data *td)
384 if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
385 __update_tv_cache(td);
388 static int break_on_this_error(struct thread_data *td, int *retptr)
392 if (ret < 0 || td->error) {
395 if (!td->o.continue_on_error)
403 if (td_non_fatal_error(err)) {
405 * Continue with the I/Os in case of
408 update_error_count(td, err);
412 } else if (td->o.fill_device && err == ENOSPC) {
414 * We expect to hit this error if
415 * fill_device option is set.
422 * Stop the I/O in case of a fatal
425 update_error_count(td, err);
434 * The main verify engine. Runs over the writes we previously submitted,
435 * reads the blocks back in, and checks the crc/md5 of the data.
437 static void do_verify(struct thread_data *td)
444 dprint(FD_VERIFY, "starting loop\n");
447 * sync io first and invalidate cache, to make sure we really
450 for_each_file(td, f, i) {
451 if (!fio_file_open(f))
453 if (fio_io_sync(td, f))
455 if (file_invalidate_cache(td, f))
462 td_set_runstate(td, TD_VERIFYING);
465 while (!td->terminate) {
470 if (runtime_exceeded(td, &td->tv_cache)) {
471 __update_tv_cache(td);
472 if (runtime_exceeded(td, &td->tv_cache)) {
478 io_u = __get_io_u(td);
482 if (get_next_verify(td, io_u)) {
487 if (td_io_prep(td, io_u)) {
492 if (td->o.verify_async)
493 io_u->end_io = verify_io_u_async;
495 io_u->end_io = verify_io_u;
497 ret = td_io_queue(td, io_u);
499 case FIO_Q_COMPLETED:
502 clear_io_u(td, io_u);
503 } else if (io_u->resid) {
504 int bytes = io_u->xfer_buflen - io_u->resid;
510 td_verror(td, EIO, "full resid");
515 io_u->xfer_buflen = io_u->resid;
516 io_u->xfer_buf += bytes;
517 io_u->offset += bytes;
519 if (ddir_rw(io_u->ddir))
520 td->ts.short_io_u[io_u->ddir]++;
523 if (io_u->offset == f->real_file_size)
526 requeue_io_u(td, &io_u);
529 ret = io_u_sync_complete(td, io_u, NULL);
537 requeue_io_u(td, &io_u);
538 ret2 = td_io_commit(td);
544 td_verror(td, -ret, "td_io_queue");
548 if (break_on_this_error(td, &ret))
552 * if we can queue more, do so. but check if there are
553 * completed io_u's first.
555 full = queue_full(td) || ret == FIO_Q_BUSY;
556 if (full || !td->o.iodepth_batch_complete) {
557 min_events = min(td->o.iodepth_batch_complete,
559 if (full && !min_events)
564 * Reap required number of io units, if any,
565 * and do the verification on them through
566 * the callback handler
568 if (io_u_queued_complete(td, min_events, NULL) < 0) {
572 } while (full && (td->cur_depth > td->o.iodepth_low));
579 min_events = td->cur_depth;
582 ret = io_u_queued_complete(td, min_events, NULL);
584 cleanup_pending_aio(td);
586 td_set_runstate(td, TD_RUNNING);
588 dprint(FD_VERIFY, "exiting loop\n");
592 * Main IO worker function. It retrieves io_u's to process and queues
593 * and reaps them, checking for rate and errors along the way.
595 static void do_io(struct thread_data *td)
600 if (in_ramp_time(td))
601 td_set_runstate(td, TD_RAMP);
603 td_set_runstate(td, TD_RUNNING);
605 while ( (td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
606 (!flist_empty(&td->trim_list)) ||
607 ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) ) {
608 struct timeval comp_time;
609 unsigned long bytes_done[2] = { 0, 0 };
619 if (runtime_exceeded(td, &td->tv_cache)) {
620 __update_tv_cache(td);
621 if (runtime_exceeded(td, &td->tv_cache)) {
632 * Add verification end_io handler, if asked to verify
633 * a previously written file.
635 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
637 if (td->o.verify_async)
638 io_u->end_io = verify_io_u_async;
640 io_u->end_io = verify_io_u;
641 td_set_runstate(td, TD_VERIFYING);
642 } else if (in_ramp_time(td))
643 td_set_runstate(td, TD_RAMP);
645 td_set_runstate(td, TD_RUNNING);
647 ret = td_io_queue(td, io_u);
649 case FIO_Q_COMPLETED:
652 clear_io_u(td, io_u);
653 } else if (io_u->resid) {
654 int bytes = io_u->xfer_buflen - io_u->resid;
655 struct fio_file *f = io_u->file;
661 td_verror(td, EIO, "full resid");
666 io_u->xfer_buflen = io_u->resid;
667 io_u->xfer_buf += bytes;
668 io_u->offset += bytes;
670 if (ddir_rw(io_u->ddir))
671 td->ts.short_io_u[io_u->ddir]++;
673 if (io_u->offset == f->real_file_size)
676 requeue_io_u(td, &io_u);
679 if (__should_check_rate(td, 0) ||
680 __should_check_rate(td, 1))
681 fio_gettime(&comp_time, NULL);
683 ret = io_u_sync_complete(td, io_u, bytes_done);
690 * if the engine doesn't have a commit hook,
691 * the io_u is really queued. if it does have such
692 * a hook, it has to call io_u_queued() itself.
694 if (td->io_ops->commit == NULL)
695 io_u_queued(td, io_u);
698 requeue_io_u(td, &io_u);
699 ret2 = td_io_commit(td);
709 if (break_on_this_error(td, &ret))
713 * See if we need to complete some commands
715 full = queue_full(td) || ret == FIO_Q_BUSY;
716 if (full || !td->o.iodepth_batch_complete) {
717 min_evts = min(td->o.iodepth_batch_complete,
719 if (full && !min_evts)
722 if (__should_check_rate(td, 0) ||
723 __should_check_rate(td, 1))
724 fio_gettime(&comp_time, NULL);
727 ret = io_u_queued_complete(td, min_evts, bytes_done);
731 } while (full && (td->cur_depth > td->o.iodepth_low));
736 if (!(bytes_done[0] + bytes_done[1]))
739 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
740 if (check_min_rate(td, &comp_time, bytes_done)) {
741 if (exitall_on_terminate)
742 terminate_threads(td->groupid);
743 td_verror(td, EIO, "check_min_rate");
748 if (td->o.thinktime) {
749 unsigned long long b;
751 b = td->io_blocks[0] + td->io_blocks[1];
752 if (!(b % td->o.thinktime_blocks)) {
755 if (td->o.thinktime_spin)
756 usec_spin(td->o.thinktime_spin);
758 left = td->o.thinktime - td->o.thinktime_spin;
760 usec_sleep(td, left);
765 if (td->trim_entries)
766 printf("trim entries %ld\n", td->trim_entries);
768 if (td->o.fill_device && td->error == ENOSPC) {
777 ret = io_u_queued_complete(td, i, NULL);
778 if (td->o.fill_device && td->error == ENOSPC)
782 if (should_fsync(td) && td->o.end_fsync) {
783 td_set_runstate(td, TD_FSYNCING);
785 for_each_file(td, f, i) {
786 if (!fio_file_open(f))
792 cleanup_pending_aio(td);
795 * stop job if we failed doing any IO
797 if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
801 static void cleanup_io_u(struct thread_data *td)
803 struct flist_head *entry, *n;
806 flist_for_each_safe(entry, n, &td->io_u_freelist) {
807 io_u = flist_entry(entry, struct io_u, list);
809 flist_del(&io_u->list);
810 fio_memfree(io_u, sizeof(*io_u));
816 static int init_io_u(struct thread_data *td)
820 int cl_align, i, max_units;
823 max_units = td->o.iodepth;
824 max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
825 td->orig_buffer_size = (unsigned long long) max_bs
826 * (unsigned long long) max_units;
828 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
831 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
832 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
835 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
836 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
840 if (allocate_io_mem(td))
843 if (td->o.odirect || td->o.mem_align ||
844 (td->io_ops->flags & FIO_RAWIO))
845 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
849 cl_align = os_cache_line_size();
851 for (i = 0; i < max_units; i++) {
857 ptr = fio_memalign(cl_align, sizeof(*io_u));
859 log_err("fio: unable to allocate aligned memory\n");
864 memset(io_u, 0, sizeof(*io_u));
865 INIT_FLIST_HEAD(&io_u->list);
866 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
868 if (!(td->io_ops->flags & FIO_NOIO)) {
869 io_u->buf = p + max_bs * i;
870 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
872 if (td_write(td) && !td->o.refill_buffers)
873 io_u_fill_buffer(td, io_u, max_bs);
874 else if (td_write(td) && td->o.verify_pattern_bytes) {
876 * Fill the buffer with the pattern if we are
877 * going to be doing writes.
879 fill_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
884 io_u->flags = IO_U_F_FREE;
885 flist_add(&io_u->list, &td->io_u_freelist);
891 static int switch_ioscheduler(struct thread_data *td)
893 char tmp[256], tmp2[128];
897 if (td->io_ops->flags & FIO_DISKLESSIO)
900 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
902 f = fopen(tmp, "r+");
904 if (errno == ENOENT) {
905 log_err("fio: os or kernel doesn't support IO scheduler"
909 td_verror(td, errno, "fopen iosched");
916 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
917 if (ferror(f) || ret != 1) {
918 td_verror(td, errno, "fwrite");
926 * Read back and check that the selected scheduler is now the default.
928 ret = fread(tmp, 1, sizeof(tmp), f);
929 if (ferror(f) || ret < 0) {
930 td_verror(td, errno, "fread");
935 sprintf(tmp2, "[%s]", td->o.ioscheduler);
936 if (!strstr(tmp, tmp2)) {
937 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
938 td_verror(td, EINVAL, "iosched_switch");
947 static int keep_running(struct thread_data *td)
949 unsigned long long io_done;
953 if (td->o.time_based)
960 io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
962 if (io_done < td->o.size)
968 static void reset_io_counters(struct thread_data *td)
970 td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
971 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
973 td->rate_bytes[0] = td->rate_bytes[1] = 0;
974 td->rate_blocks[0] = td->rate_blocks[1] = 0;
976 td->last_was_sync = 0;
979 * reset file done count if we are to start over
981 if (td->o.time_based || td->o.loops)
982 td->nr_done_files = 0;
985 void reset_all_stats(struct thread_data *td)
990 reset_io_counters(td);
992 for (i = 0; i < 2; i++) {
994 td->io_blocks[i] = 0;
995 td->io_issues[i] = 0;
996 td->ts.total_io_u[i] = 0;
999 fio_gettime(&tv, NULL);
1000 td->ts.runtime[0] = 0;
1001 td->ts.runtime[1] = 0;
1002 memcpy(&td->epoch, &tv, sizeof(tv));
1003 memcpy(&td->start, &tv, sizeof(tv));
1006 static void clear_io_state(struct thread_data *td)
1011 reset_io_counters(td);
1014 for_each_file(td, f, i)
1015 fio_file_clear_done(f);
1018 * Set the same seed to get repeatable runs
1020 td_fill_rand_seeds(td);
1023 static int exec_string(const char *string)
1025 int ret, newlen = strlen(string) + 1 + 8;
1028 str = malloc(newlen);
1029 sprintf(str, "sh -c %s", string);
1033 log_err("fio: exec of cmd <%s> failed\n", str);
1040 * Entry point for the thread based jobs. The process based jobs end up
1041 * here as well, after a little setup.
1043 static void *thread_main(void *data)
1045 unsigned long long elapsed;
1046 struct thread_data *td = data;
1047 pthread_condattr_t attr;
1050 if (!td->o.use_thread) {
1056 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1058 INIT_FLIST_HEAD(&td->io_u_freelist);
1059 INIT_FLIST_HEAD(&td->io_u_busylist);
1060 INIT_FLIST_HEAD(&td->io_u_requeues);
1061 INIT_FLIST_HEAD(&td->io_log_list);
1062 INIT_FLIST_HEAD(&td->io_hist_list);
1063 INIT_FLIST_HEAD(&td->verify_list);
1064 INIT_FLIST_HEAD(&td->trim_list);
1065 pthread_mutex_init(&td->io_u_lock, NULL);
1066 td->io_hist_tree = RB_ROOT;
1068 pthread_condattr_init(&attr);
1069 pthread_cond_init(&td->verify_cond, &attr);
1070 pthread_cond_init(&td->free_cond, &attr);
1072 td_set_runstate(td, TD_INITIALIZED);
1073 dprint(FD_MUTEX, "up startup_mutex\n");
1074 fio_mutex_up(startup_mutex);
1075 dprint(FD_MUTEX, "wait on td->mutex\n");
1076 fio_mutex_down(td->mutex);
1077 dprint(FD_MUTEX, "done waiting on td->mutex\n");
1080 * the ->mutex mutex is now no longer used, close it to avoid
1081 * eating a file descriptor
1083 fio_mutex_remove(td->mutex);
1086 * A new gid requires privilege, so we need to do this before setting
1089 if (td->o.gid != -1U && setgid(td->o.gid)) {
1090 td_verror(td, errno, "setgid");
1093 if (td->o.uid != -1U && setuid(td->o.uid)) {
1094 td_verror(td, errno, "setuid");
1099 * If we have a gettimeofday() thread, make sure we exclude that
1100 * thread from this job
1103 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1106 * Set affinity first, in case it has an impact on the memory
1109 if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1110 td_verror(td, errno, "cpu_set_affinity");
1115 * May alter parameters that init_io_u() will use, so we need to
1124 if (td->o.verify_async && verify_async_init(td))
1127 if (td->ioprio_set) {
1128 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1129 td_verror(td, errno, "ioprio_set");
1134 if (td->o.cgroup_weight && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1137 if (nice(td->o.nice) == -1) {
1138 td_verror(td, errno, "nice");
1142 if (td->o.ioscheduler && switch_ioscheduler(td))
1145 if (!td->o.create_serialize && setup_files(td))
1151 if (init_random_map(td))
1154 if (td->o.exec_prerun) {
1155 if (exec_string(td->o.exec_prerun))
1159 if (td->o.pre_read) {
1160 if (pre_read_files(td) < 0)
1164 fio_gettime(&td->epoch, NULL);
1165 getrusage(RUSAGE_SELF, &td->ts.ru_start);
1168 while (keep_running(td)) {
1169 fio_gettime(&td->start, NULL);
1170 memcpy(&td->ts.stat_sample_time[0], &td->start,
1172 memcpy(&td->ts.stat_sample_time[1], &td->start,
1174 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1176 if (td->o.ratemin[0] || td->o.ratemin[1])
1177 memcpy(&td->lastrate, &td->ts.stat_sample_time,
1178 sizeof(td->lastrate));
1183 prune_io_piece_log(td);
1189 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1190 elapsed = utime_since_now(&td->start);
1191 td->ts.runtime[DDIR_READ] += elapsed;
1193 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1194 elapsed = utime_since_now(&td->start);
1195 td->ts.runtime[DDIR_WRITE] += elapsed;
1198 if (td->error || td->terminate)
1201 if (!td->o.do_verify ||
1202 td->o.verify == VERIFY_NONE ||
1203 (td->io_ops->flags & FIO_UNIDIR))
1208 fio_gettime(&td->start, NULL);
1212 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1214 if (td->error || td->terminate)
1218 update_rusage_stat(td);
1219 td->ts.runtime[0] = (td->ts.runtime[0] + 999) / 1000;
1220 td->ts.runtime[1] = (td->ts.runtime[1] + 999) / 1000;
1221 td->ts.total_run_time = mtime_since_now(&td->epoch);
1222 td->ts.io_bytes[0] = td->io_bytes[0];
1223 td->ts.io_bytes[1] = td->io_bytes[1];
1225 fio_mutex_down(writeout_mutex);
1226 if (td->ts.bw_log) {
1227 if (td->o.bw_log_file) {
1228 finish_log_named(td, td->ts.bw_log,
1229 td->o.bw_log_file, "bw");
1231 finish_log(td, td->ts.bw_log, "bw");
1233 if (td->ts.lat_log) {
1234 if (td->o.lat_log_file) {
1235 finish_log_named(td, td->ts.lat_log,
1236 td->o.lat_log_file, "lat");
1238 finish_log(td, td->ts.lat_log, "lat");
1240 if (td->ts.slat_log) {
1241 if (td->o.lat_log_file) {
1242 finish_log_named(td, td->ts.slat_log,
1243 td->o.lat_log_file, "slat");
1245 finish_log(td, td->ts.slat_log, "slat");
1247 if (td->ts.clat_log) {
1248 if (td->o.lat_log_file) {
1249 finish_log_named(td, td->ts.clat_log,
1250 td->o.lat_log_file, "clat");
1252 finish_log(td, td->ts.clat_log, "clat");
1254 fio_mutex_up(writeout_mutex);
1255 if (td->o.exec_postrun)
1256 exec_string(td->o.exec_postrun);
1258 if (exitall_on_terminate)
1259 terminate_threads(td->groupid);
1263 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1266 if (td->o.verify_async)
1267 verify_async_exit(td);
1269 close_and_free_files(td);
1272 cgroup_shutdown(td, &cgroup_mnt);
1274 if (td->o.cpumask_set) {
1275 int ret = fio_cpuset_exit(&td->o.cpumask);
1277 td_verror(td, ret, "fio_cpuset_exit");
1281 * do this very late, it will log file closing as well
1283 if (td->o.write_iolog_file)
1284 write_iolog_close(td);
1286 options_mem_free(td);
1287 td_set_runstate(td, TD_EXITED);
1288 return (void *) (unsigned long) td->error;
1292 * We cannot pass the td data into a forked process, so attach the td and
1293 * pass it to the thread worker.
1295 static int fork_main(int shmid, int offset)
1297 struct thread_data *td;
1301 data = shmat(shmid, NULL, 0);
1302 if (data == (void *) -1) {
1310 * HP-UX inherits shm mappings?
1315 td = data + offset * sizeof(struct thread_data);
1316 ret = thread_main(td);
1318 return (int) (unsigned long) ret;
1322 * Run over the job map and reap the threads that have exited, if any.
1324 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1326 struct thread_data *td;
1327 int i, cputhreads, realthreads, pending, status, ret;
1330 * reap exited threads (TD_EXITED -> TD_REAPED)
1332 realthreads = pending = cputhreads = 0;
1333 for_each_td(td, i) {
1337 * ->io_ops is NULL for a thread that has closed its
1340 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1349 if (td->runstate == TD_REAPED)
1351 if (td->o.use_thread) {
1352 if (td->runstate == TD_EXITED) {
1353 td_set_runstate(td, TD_REAPED);
1360 if (td->runstate == TD_EXITED)
1364 * check if someone quit or got killed in an unusual way
1366 ret = waitpid(td->pid, &status, flags);
1368 if (errno == ECHILD) {
1369 log_err("fio: pid=%d disappeared %d\n",
1370 (int) td->pid, td->runstate);
1371 td_set_runstate(td, TD_REAPED);
1375 } else if (ret == td->pid) {
1376 if (WIFSIGNALED(status)) {
1377 int sig = WTERMSIG(status);
1380 log_err("fio: pid=%d, got signal=%d\n",
1381 (int) td->pid, sig);
1382 td_set_runstate(td, TD_REAPED);
1385 if (WIFEXITED(status)) {
1386 if (WEXITSTATUS(status) && !td->error)
1387 td->error = WEXITSTATUS(status);
1389 td_set_runstate(td, TD_REAPED);
1395 * thread is not dead, continue
1401 (*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
1402 (*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
1409 done_secs += mtime_since_now(&td->epoch) / 1000;
1412 if (*nr_running == cputhreads && !pending && realthreads)
1413 terminate_threads(TERMINATE_ALL);
1416 static void *gtod_thread_main(void *data)
1418 fio_mutex_up(startup_mutex);
1421 * As long as we have jobs around, update the clock. It would be nice
1422 * to have some way of NOT hammering that CPU with gettimeofday(),
1423 * but I'm not sure what to use outside of a simple CPU nop to relax
1424 * it - we don't want to lose precision.
1434 static int fio_start_gtod_thread(void)
1436 pthread_attr_t attr;
1439 pthread_attr_init(&attr);
1440 pthread_attr_setstacksize(&attr, PTHREAD_STACK_MIN);
1441 ret = pthread_create(>od_thread, &attr, gtod_thread_main, NULL);
1442 pthread_attr_destroy(&attr);
1444 log_err("Can't create gtod thread: %s\n", strerror(ret));
1448 ret = pthread_detach(gtod_thread);
1450 log_err("Can't detatch gtod thread: %s\n", strerror(ret));
1454 dprint(FD_MUTEX, "wait on startup_mutex\n");
1455 fio_mutex_down(startup_mutex);
1456 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1461 * Main function for kicking off and reaping jobs, as needed.
1463 static void run_threads(void)
1465 struct thread_data *td;
1466 unsigned long spent;
1467 int i, todo, nr_running, m_rate, t_rate, nr_started;
1469 if (fio_pin_memory())
1472 if (fio_gtod_offload && fio_start_gtod_thread())
1475 if (!terse_output) {
1476 log_info("Starting ");
1478 log_info("%d thread%s", nr_thread,
1479 nr_thread > 1 ? "s" : "");
1483 log_info("%d process%s", nr_process,
1484 nr_process > 1 ? "es" : "");
1492 todo = thread_number;
1495 m_rate = t_rate = 0;
1497 for_each_td(td, i) {
1498 print_status_init(td->thread_number - 1);
1500 if (!td->o.create_serialize) {
1506 * do file setup here so it happens sequentially,
1507 * we don't want X number of threads getting their
1508 * client data interspersed on disk
1510 if (setup_files(td)) {
1513 log_err("fio: pid=%d, err=%d/%s\n",
1514 (int) td->pid, td->error, td->verror);
1515 td_set_runstate(td, TD_REAPED);
1522 * for sharing to work, each job must always open
1523 * its own files. so close them, if we opened them
1526 for_each_file(td, f, j) {
1527 if (fio_file_open(f))
1528 td_io_close_file(td, f);
1538 struct thread_data *map[REAL_MAX_JOBS];
1539 struct timeval this_start;
1540 int this_jobs = 0, left;
1543 * create threads (TD_NOT_CREATED -> TD_CREATED)
1545 for_each_td(td, i) {
1546 if (td->runstate != TD_NOT_CREATED)
1550 * never got a chance to start, killed by other
1551 * thread for some reason
1553 if (td->terminate) {
1558 if (td->o.start_delay) {
1559 spent = mtime_since_genesis();
1561 if (td->o.start_delay * 1000 > spent)
1565 if (td->o.stonewall && (nr_started || nr_running)) {
1566 dprint(FD_PROCESS, "%s: stonewall wait\n",
1572 * Set state to created. Thread will transition
1573 * to TD_INITIALIZED when it's done setting up.
1575 td_set_runstate(td, TD_CREATED);
1576 map[this_jobs++] = td;
1579 if (td->o.use_thread) {
1582 dprint(FD_PROCESS, "will pthread_create\n");
1583 ret = pthread_create(&td->thread, NULL,
1586 log_err("pthread_create: %s\n",
1591 ret = pthread_detach(td->thread);
1593 log_err("pthread_detach: %s",
1597 dprint(FD_PROCESS, "will fork\n");
1600 int ret = fork_main(shm_id, i);
1603 } else if (i == fio_debug_jobno)
1604 *fio_debug_jobp = pid;
1606 dprint(FD_MUTEX, "wait on startup_mutex\n");
1607 if (fio_mutex_down_timeout(startup_mutex, 10)) {
1608 log_err("fio: job startup hung? exiting.\n");
1609 terminate_threads(TERMINATE_ALL);
1614 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1618 * Wait for the started threads to transition to
1621 fio_gettime(&this_start, NULL);
1623 while (left && !fio_abort) {
1624 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1629 for (i = 0; i < this_jobs; i++) {
1633 if (td->runstate == TD_INITIALIZED) {
1636 } else if (td->runstate >= TD_EXITED) {
1640 nr_running++; /* work-around... */
1646 log_err("fio: %d jobs failed to start\n", left);
1647 for (i = 0; i < this_jobs; i++) {
1651 kill(td->pid, SIGTERM);
1657 * start created threads (TD_INITIALIZED -> TD_RUNNING).
1659 for_each_td(td, i) {
1660 if (td->runstate != TD_INITIALIZED)
1663 if (in_ramp_time(td))
1664 td_set_runstate(td, TD_RAMP);
1666 td_set_runstate(td, TD_RUNNING);
1669 m_rate += td->o.ratemin[0] + td->o.ratemin[1];
1670 t_rate += td->o.rate[0] + td->o.rate[1];
1672 fio_mutex_up(td->mutex);
1675 reap_threads(&nr_running, &t_rate, &m_rate);
1681 while (nr_running) {
1682 reap_threads(&nr_running, &t_rate, &m_rate);
1690 int main(int argc, char *argv[])
1695 init_rand(&__fio_rand_state);
1698 * We need locale for number printing, if it isn't set then just
1699 * go with the US format.
1701 if (!getenv("LC_NUMERIC"))
1702 setlocale(LC_NUMERIC, "en_US");
1704 ps = sysconf(_SC_PAGESIZE);
1706 log_err("Failed to get page size\n");
1713 fio_keywords_init();
1715 if (parse_options(argc, argv))
1718 if (exec_profile && load_profile(exec_profile))
1725 setup_log(&agg_io_log[DDIR_READ]);
1726 setup_log(&agg_io_log[DDIR_WRITE]);
1729 startup_mutex = fio_mutex_init(0);
1730 if (startup_mutex == NULL)
1732 writeout_mutex = fio_mutex_init(1);
1733 if (writeout_mutex == NULL)
1737 create_disk_util_thread();
1739 cgroup_list = smalloc(sizeof(*cgroup_list));
1740 INIT_FLIST_HEAD(cgroup_list);
1747 __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1748 __finish_log(agg_io_log[DDIR_WRITE],
1749 "agg-write_bw.log");
1753 cgroup_kill(cgroup_list);
1757 fio_mutex_remove(startup_mutex);
1758 fio_mutex_remove(writeout_mutex);