2 * fio - the flexible io tester
4 * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
5 * Copyright (C) 2006-2012 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
52 static pthread_t disk_util_thread;
53 static struct fio_mutex *disk_thread_mutex;
54 static struct fio_mutex *startup_mutex;
55 static struct fio_mutex *writeout_mutex;
56 static struct flist_head *cgroup_list;
57 static char *cgroup_mnt;
58 static int exit_value;
59 static volatile int fio_abort;
60 static unsigned int nr_process = 0;
61 static unsigned int nr_thread = 0;
63 struct io_log *agg_io_log[2];
66 unsigned int thread_number = 0;
69 unsigned long done_secs = 0;
71 #define PAGE_ALIGN(buf) \
72 (char *) (((uintptr_t) (buf) + page_mask) & ~page_mask)
74 #define JOB_START_TIMEOUT (5 * 1000)
76 static void sig_int(int sig)
80 fio_server_got_signal(sig);
82 log_info("\nfio: terminating on signal %d\n", sig);
87 fio_terminate_threads(TERMINATE_ALL);
91 static void sig_show_status(int sig)
93 show_running_run_stats();
96 static void set_sig_handlers(void)
100 memset(&act, 0, sizeof(act));
101 act.sa_handler = sig_int;
102 act.sa_flags = SA_RESTART;
103 sigaction(SIGINT, &act, NULL);
105 memset(&act, 0, sizeof(act));
106 act.sa_handler = sig_int;
107 act.sa_flags = SA_RESTART;
108 sigaction(SIGTERM, &act, NULL);
110 memset(&act, 0, sizeof(act));
111 act.sa_handler = sig_show_status;
112 act.sa_flags = SA_RESTART;
113 sigaction(SIGUSR1, &act, NULL);
116 memset(&act, 0, sizeof(act));
117 act.sa_handler = sig_int;
118 act.sa_flags = SA_RESTART;
119 sigaction(SIGPIPE, &act, NULL);
124 * Check if we are above the minimum rate given.
126 static int __check_min_rate(struct thread_data *td, struct timeval *now,
129 unsigned long long bytes = 0;
130 unsigned long iops = 0;
133 unsigned int ratemin = 0;
134 unsigned int rate_iops = 0;
135 unsigned int rate_iops_min = 0;
137 assert(ddir_rw(ddir));
139 if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
143 * allow a 2 second settle period in the beginning
145 if (mtime_since(&td->start, now) < 2000)
148 iops += td->this_io_blocks[ddir];
149 bytes += td->this_io_bytes[ddir];
150 ratemin += td->o.ratemin[ddir];
151 rate_iops += td->o.rate_iops[ddir];
152 rate_iops_min += td->o.rate_iops_min[ddir];
155 * if rate blocks is set, sample is running
157 if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
158 spent = mtime_since(&td->lastrate[ddir], now);
159 if (spent < td->o.ratecycle)
162 if (td->o.rate[ddir]) {
164 * check bandwidth specified rate
166 if (bytes < td->rate_bytes[ddir]) {
167 log_err("%s: min rate %u not met\n", td->o.name,
171 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
172 if (rate < ratemin ||
173 bytes < td->rate_bytes[ddir]) {
174 log_err("%s: min rate %u not met, got"
175 " %luKB/sec\n", td->o.name,
182 * checks iops specified rate
184 if (iops < rate_iops) {
185 log_err("%s: min iops rate %u not met\n",
186 td->o.name, rate_iops);
189 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
190 if (rate < rate_iops_min ||
191 iops < td->rate_blocks[ddir]) {
192 log_err("%s: min iops rate %u not met,"
193 " got %lu\n", td->o.name,
194 rate_iops_min, rate);
200 td->rate_bytes[ddir] = bytes;
201 td->rate_blocks[ddir] = iops;
202 memcpy(&td->lastrate[ddir], now, sizeof(*now));
206 static int check_min_rate(struct thread_data *td, struct timeval *now,
207 unsigned long *bytes_done)
212 ret |= __check_min_rate(td, now, 0);
214 ret |= __check_min_rate(td, now, 1);
220 * When job exits, we can cancel the in-flight IO if we are using async
221 * io. Attempt to do so.
223 static void cleanup_pending_aio(struct thread_data *td)
225 struct flist_head *entry, *n;
230 * get immediately available events, if any
232 r = io_u_queued_complete(td, 0, NULL);
237 * now cancel remaining active events
239 if (td->io_ops->cancel) {
240 flist_for_each_safe(entry, n, &td->io_u_busylist) {
241 io_u = flist_entry(entry, struct io_u, list);
244 * if the io_u isn't in flight, then that generally
245 * means someone leaked an io_u. complain but fix
246 * it up, so we don't stall here.
248 if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
249 log_err("fio: non-busy IO on busy list\n");
252 r = td->io_ops->cancel(td, io_u);
260 r = io_u_queued_complete(td, td->cur_depth, NULL);
264 * Helper to handle the final sync of a file. Works just like the normal
265 * io path, just does everything sync.
267 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
269 struct io_u *io_u = __get_io_u(td);
275 io_u->ddir = DDIR_SYNC;
278 if (td_io_prep(td, io_u)) {
284 ret = td_io_queue(td, io_u);
286 td_verror(td, io_u->error, "td_io_queue");
289 } else if (ret == FIO_Q_QUEUED) {
290 if (io_u_queued_complete(td, 1, NULL) < 0)
292 } else if (ret == FIO_Q_COMPLETED) {
294 td_verror(td, io_u->error, "td_io_queue");
298 if (io_u_sync_complete(td, io_u, NULL) < 0)
300 } else if (ret == FIO_Q_BUSY) {
301 if (td_io_commit(td))
309 static inline void __update_tv_cache(struct thread_data *td)
311 fio_gettime(&td->tv_cache, NULL);
314 static inline void update_tv_cache(struct thread_data *td)
316 if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
317 __update_tv_cache(td);
320 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
322 if (in_ramp_time(td))
326 if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
332 static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
337 if (ret < 0 || td->error) {
345 if (!(td->o.continue_on_error & td_error_type(ddir, err)))
348 if (td_non_fatal_error(err)) {
350 * Continue with the I/Os in case of
353 update_error_count(td, err);
357 } else if (td->o.fill_device && err == ENOSPC) {
359 * We expect to hit this error if
360 * fill_device option is set.
367 * Stop the I/O in case of a fatal
370 update_error_count(td, err);
379 * The main verify engine. Runs over the writes we previously submitted,
380 * reads the blocks back in, and checks the crc/md5 of the data.
382 static void do_verify(struct thread_data *td)
389 dprint(FD_VERIFY, "starting loop\n");
392 * sync io first and invalidate cache, to make sure we really
395 for_each_file(td, f, i) {
396 if (!fio_file_open(f))
398 if (fio_io_sync(td, f))
400 if (file_invalidate_cache(td, f))
407 td_set_runstate(td, TD_VERIFYING);
410 while (!td->terminate) {
415 if (runtime_exceeded(td, &td->tv_cache)) {
416 __update_tv_cache(td);
417 if (runtime_exceeded(td, &td->tv_cache)) {
423 if (flow_threshold_exceeded(td))
426 io_u = __get_io_u(td);
430 if (get_next_verify(td, io_u)) {
435 if (td_io_prep(td, io_u)) {
440 if (td->o.verify_async)
441 io_u->end_io = verify_io_u_async;
443 io_u->end_io = verify_io_u;
445 ret = td_io_queue(td, io_u);
447 case FIO_Q_COMPLETED:
450 clear_io_u(td, io_u);
451 } else if (io_u->resid) {
452 int bytes = io_u->xfer_buflen - io_u->resid;
458 td_verror(td, EIO, "full resid");
463 io_u->xfer_buflen = io_u->resid;
464 io_u->xfer_buf += bytes;
465 io_u->offset += bytes;
467 if (ddir_rw(io_u->ddir))
468 td->ts.short_io_u[io_u->ddir]++;
471 if (io_u->offset == f->real_file_size)
474 requeue_io_u(td, &io_u);
477 ret = io_u_sync_complete(td, io_u, NULL);
485 requeue_io_u(td, &io_u);
486 ret2 = td_io_commit(td);
492 td_verror(td, -ret, "td_io_queue");
496 if (break_on_this_error(td, io_u->ddir, &ret))
500 * if we can queue more, do so. but check if there are
501 * completed io_u's first. Note that we can get BUSY even
502 * without IO queued, if the system is resource starved.
504 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
505 if (full || !td->o.iodepth_batch_complete) {
506 min_events = min(td->o.iodepth_batch_complete,
509 * if the queue is full, we MUST reap at least 1 event
511 if (full && !min_events)
516 * Reap required number of io units, if any,
517 * and do the verification on them through
518 * the callback handler
520 if (io_u_queued_complete(td, min_events, NULL) < 0) {
524 } while (full && (td->cur_depth > td->o.iodepth_low));
531 min_events = td->cur_depth;
534 ret = io_u_queued_complete(td, min_events, NULL);
536 cleanup_pending_aio(td);
538 td_set_runstate(td, TD_RUNNING);
540 dprint(FD_VERIFY, "exiting loop\n");
543 static int io_bytes_exceeded(struct thread_data *td)
545 unsigned long long bytes;
548 bytes = td->this_io_bytes[0] + td->this_io_bytes[1];
549 else if (td_write(td))
550 bytes = td->this_io_bytes[1];
552 bytes = td->this_io_bytes[0];
554 return bytes >= td->o.size;
558 * Main IO worker function. It retrieves io_u's to process and queues
559 * and reaps them, checking for rate and errors along the way.
561 static void do_io(struct thread_data *td)
566 if (in_ramp_time(td))
567 td_set_runstate(td, TD_RAMP);
569 td_set_runstate(td, TD_RUNNING);
571 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
572 (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td) ||
574 struct timeval comp_time;
575 unsigned long bytes_done[2] = { 0, 0 };
586 if (runtime_exceeded(td, &td->tv_cache)) {
587 __update_tv_cache(td);
588 if (runtime_exceeded(td, &td->tv_cache)) {
594 if (flow_threshold_exceeded(td))
604 * Add verification end_io handler if:
605 * - Asked to verify (!td_rw(td))
606 * - Or the io_u is from our verify list (mixed write/ver)
608 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
609 ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
610 if (td->o.verify_async)
611 io_u->end_io = verify_io_u_async;
613 io_u->end_io = verify_io_u;
614 td_set_runstate(td, TD_VERIFYING);
615 } else if (in_ramp_time(td))
616 td_set_runstate(td, TD_RAMP);
618 td_set_runstate(td, TD_RUNNING);
620 ret = td_io_queue(td, io_u);
622 case FIO_Q_COMPLETED:
625 clear_io_u(td, io_u);
626 } else if (io_u->resid) {
627 int bytes = io_u->xfer_buflen - io_u->resid;
628 struct fio_file *f = io_u->file;
634 td_verror(td, EIO, "full resid");
639 io_u->xfer_buflen = io_u->resid;
640 io_u->xfer_buf += bytes;
641 io_u->offset += bytes;
643 if (ddir_rw(io_u->ddir))
644 td->ts.short_io_u[io_u->ddir]++;
646 if (io_u->offset == f->real_file_size)
649 requeue_io_u(td, &io_u);
652 if (__should_check_rate(td, 0) ||
653 __should_check_rate(td, 1))
654 fio_gettime(&comp_time, NULL);
656 ret = io_u_sync_complete(td, io_u, bytes_done);
663 * if the engine doesn't have a commit hook,
664 * the io_u is really queued. if it does have such
665 * a hook, it has to call io_u_queued() itself.
667 if (td->io_ops->commit == NULL)
668 io_u_queued(td, io_u);
671 requeue_io_u(td, &io_u);
672 ret2 = td_io_commit(td);
682 if (break_on_this_error(td, ddir, &ret))
686 * See if we need to complete some commands. Note that we
687 * can get BUSY even without IO queued, if the system is
690 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
691 if (full || !td->o.iodepth_batch_complete) {
692 min_evts = min(td->o.iodepth_batch_complete,
695 * if the queue is full, we MUST reap at least 1 event
697 if (full && !min_evts)
700 if (__should_check_rate(td, 0) ||
701 __should_check_rate(td, 1))
702 fio_gettime(&comp_time, NULL);
705 ret = io_u_queued_complete(td, min_evts, bytes_done);
709 } while (full && (td->cur_depth > td->o.iodepth_low));
714 if (!(bytes_done[0] + bytes_done[1]))
717 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
718 if (check_min_rate(td, &comp_time, bytes_done)) {
719 if (exitall_on_terminate)
720 fio_terminate_threads(td->groupid);
721 td_verror(td, EIO, "check_min_rate");
726 if (td->o.thinktime) {
727 unsigned long long b;
729 b = td->io_blocks[0] + td->io_blocks[1];
730 if (!(b % td->o.thinktime_blocks)) {
733 if (td->o.thinktime_spin)
734 usec_spin(td->o.thinktime_spin);
736 left = td->o.thinktime - td->o.thinktime_spin;
738 usec_sleep(td, left);
743 if (td->trim_entries)
744 log_err("fio: %d trim entries leaked?\n", td->trim_entries);
746 if (td->o.fill_device && td->error == ENOSPC) {
755 ret = io_u_queued_complete(td, i, NULL);
756 if (td->o.fill_device && td->error == ENOSPC)
760 if (should_fsync(td) && td->o.end_fsync) {
761 td_set_runstate(td, TD_FSYNCING);
763 for_each_file(td, f, i) {
764 if (!fio_file_open(f))
770 cleanup_pending_aio(td);
773 * stop job if we failed doing any IO
775 if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
779 static void cleanup_io_u(struct thread_data *td)
781 struct flist_head *entry, *n;
784 flist_for_each_safe(entry, n, &td->io_u_freelist) {
785 io_u = flist_entry(entry, struct io_u, list);
787 flist_del(&io_u->list);
788 fio_memfree(io_u, sizeof(*io_u));
794 static int init_io_u(struct thread_data *td)
797 unsigned int max_bs, min_write;
798 int cl_align, i, max_units;
801 max_units = td->o.iodepth;
802 max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
803 min_write = td->o.min_bs[DDIR_WRITE];
804 td->orig_buffer_size = (unsigned long long) max_bs
805 * (unsigned long long) max_units;
807 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
810 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
811 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
814 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
815 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
819 if (allocate_io_mem(td))
822 if (td->o.odirect || td->o.mem_align ||
823 (td->io_ops->flags & FIO_RAWIO))
824 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
828 cl_align = os_cache_line_size();
830 for (i = 0; i < max_units; i++) {
836 ptr = fio_memalign(cl_align, sizeof(*io_u));
838 log_err("fio: unable to allocate aligned memory\n");
843 memset(io_u, 0, sizeof(*io_u));
844 INIT_FLIST_HEAD(&io_u->list);
845 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
847 if (!(td->io_ops->flags & FIO_NOIO)) {
849 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
852 io_u_fill_buffer(td, io_u, min_write, max_bs);
853 if (td_write(td) && td->o.verify_pattern_bytes) {
855 * Fill the buffer with the pattern if we are
856 * going to be doing writes.
858 fill_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
863 io_u->flags = IO_U_F_FREE;
864 flist_add(&io_u->list, &td->io_u_freelist);
871 static int switch_ioscheduler(struct thread_data *td)
873 char tmp[256], tmp2[128];
877 if (td->io_ops->flags & FIO_DISKLESSIO)
880 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
882 f = fopen(tmp, "r+");
884 if (errno == ENOENT) {
885 log_err("fio: os or kernel doesn't support IO scheduler"
889 td_verror(td, errno, "fopen iosched");
896 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
897 if (ferror(f) || ret != 1) {
898 td_verror(td, errno, "fwrite");
906 * Read back and check that the selected scheduler is now the default.
908 ret = fread(tmp, 1, sizeof(tmp), f);
909 if (ferror(f) || ret < 0) {
910 td_verror(td, errno, "fread");
915 sprintf(tmp2, "[%s]", td->o.ioscheduler);
916 if (!strstr(tmp, tmp2)) {
917 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
918 td_verror(td, EINVAL, "iosched_switch");
927 static int keep_running(struct thread_data *td)
929 unsigned long long io_done;
933 if (td->o.time_based)
940 io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
942 if (io_done < td->o.size)
948 static int exec_string(const char *string)
950 int ret, newlen = strlen(string) + 1 + 8;
953 str = malloc(newlen);
954 sprintf(str, "sh -c %s", string);
958 log_err("fio: exec of cmd <%s> failed\n", str);
965 * Entry point for the thread based jobs. The process based jobs end up
966 * here as well, after a little setup.
968 static void *thread_main(void *data)
970 unsigned long long elapsed;
971 struct thread_data *td = data;
972 struct thread_options *o = &td->o;
973 pthread_condattr_t attr;
977 if (!o->use_thread) {
983 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
986 fio_server_send_start(td);
988 INIT_FLIST_HEAD(&td->io_u_freelist);
989 INIT_FLIST_HEAD(&td->io_u_busylist);
990 INIT_FLIST_HEAD(&td->io_u_requeues);
991 INIT_FLIST_HEAD(&td->io_log_list);
992 INIT_FLIST_HEAD(&td->io_hist_list);
993 INIT_FLIST_HEAD(&td->verify_list);
994 INIT_FLIST_HEAD(&td->trim_list);
995 pthread_mutex_init(&td->io_u_lock, NULL);
996 td->io_hist_tree = RB_ROOT;
998 pthread_condattr_init(&attr);
999 pthread_cond_init(&td->verify_cond, &attr);
1000 pthread_cond_init(&td->free_cond, &attr);
1002 td_set_runstate(td, TD_INITIALIZED);
1003 dprint(FD_MUTEX, "up startup_mutex\n");
1004 fio_mutex_up(startup_mutex);
1005 dprint(FD_MUTEX, "wait on td->mutex\n");
1006 fio_mutex_down(td->mutex);
1007 dprint(FD_MUTEX, "done waiting on td->mutex\n");
1010 * the ->mutex mutex is now no longer used, close it to avoid
1011 * eating a file descriptor
1013 fio_mutex_remove(td->mutex);
1017 * A new gid requires privilege, so we need to do this before setting
1020 if (o->gid != -1U && setgid(o->gid)) {
1021 td_verror(td, errno, "setgid");
1024 if (o->uid != -1U && setuid(o->uid)) {
1025 td_verror(td, errno, "setuid");
1030 * If we have a gettimeofday() thread, make sure we exclude that
1031 * thread from this job
1034 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1037 * Set affinity first, in case it has an impact on the memory
1040 if (o->cpumask_set) {
1041 ret = fio_setaffinity(td->pid, o->cpumask);
1043 td_verror(td, errno, "cpu_set_affinity");
1048 if (fio_pin_memory(td))
1052 * May alter parameters that init_io_u() will use, so we need to
1061 if (o->verify_async && verify_async_init(td))
1065 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1067 td_verror(td, errno, "ioprio_set");
1072 if (td->o.cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1076 if (nice(o->nice) == -1 && errno != 0) {
1077 td_verror(td, errno, "nice");
1081 if (o->ioscheduler && switch_ioscheduler(td))
1084 if (!o->create_serialize && setup_files(td))
1090 if (init_random_map(td))
1093 if (o->exec_prerun && exec_string(o->exec_prerun))
1097 if (pre_read_files(td) < 0)
1101 fio_verify_init(td);
1103 fio_gettime(&td->epoch, NULL);
1104 getrusage(RUSAGE_SELF, &td->ru_start);
1107 while (keep_running(td)) {
1108 fio_gettime(&td->start, NULL);
1109 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1110 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1111 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1113 if (td->o.ratemin[0] || td->o.ratemin[1]) {
1114 memcpy(&td->lastrate[0], &td->bw_sample_time,
1115 sizeof(td->bw_sample_time));
1116 memcpy(&td->lastrate[1], &td->bw_sample_time,
1117 sizeof(td->bw_sample_time));
1123 prune_io_piece_log(td);
1129 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1130 elapsed = utime_since_now(&td->start);
1131 td->ts.runtime[DDIR_READ] += elapsed;
1133 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1134 elapsed = utime_since_now(&td->start);
1135 td->ts.runtime[DDIR_WRITE] += elapsed;
1138 if (td->error || td->terminate)
1141 if (!td->o.do_verify ||
1142 td->o.verify == VERIFY_NONE ||
1143 (td->io_ops->flags & FIO_UNIDIR))
1148 fio_gettime(&td->start, NULL);
1152 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1154 if (td->error || td->terminate)
1158 update_rusage_stat(td);
1159 td->ts.runtime[0] = (td->ts.runtime[0] + 999) / 1000;
1160 td->ts.runtime[1] = (td->ts.runtime[1] + 999) / 1000;
1161 td->ts.total_run_time = mtime_since_now(&td->epoch);
1162 td->ts.io_bytes[0] = td->io_bytes[0];
1163 td->ts.io_bytes[1] = td->io_bytes[1];
1165 fio_unpin_memory(td);
1167 fio_mutex_down(writeout_mutex);
1169 if (td->o.bw_log_file) {
1170 finish_log_named(td, td->bw_log,
1171 td->o.bw_log_file, "bw");
1173 finish_log(td, td->bw_log, "bw");
1176 if (td->o.lat_log_file) {
1177 finish_log_named(td, td->lat_log,
1178 td->o.lat_log_file, "lat");
1180 finish_log(td, td->lat_log, "lat");
1183 if (td->o.lat_log_file) {
1184 finish_log_named(td, td->slat_log,
1185 td->o.lat_log_file, "slat");
1187 finish_log(td, td->slat_log, "slat");
1190 if (td->o.lat_log_file) {
1191 finish_log_named(td, td->clat_log,
1192 td->o.lat_log_file, "clat");
1194 finish_log(td, td->clat_log, "clat");
1197 if (td->o.iops_log_file) {
1198 finish_log_named(td, td->iops_log,
1199 td->o.iops_log_file, "iops");
1201 finish_log(td, td->iops_log, "iops");
1204 fio_mutex_up(writeout_mutex);
1205 if (td->o.exec_postrun)
1206 exec_string(td->o.exec_postrun);
1208 if (exitall_on_terminate)
1209 fio_terminate_threads(td->groupid);
1213 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1216 if (td->o.verify_async)
1217 verify_async_exit(td);
1219 close_and_free_files(td);
1222 cgroup_shutdown(td, &cgroup_mnt);
1224 if (o->cpumask_set) {
1225 int ret = fio_cpuset_exit(&o->cpumask);
1227 td_verror(td, ret, "fio_cpuset_exit");
1231 * do this very late, it will log file closing as well
1233 if (td->o.write_iolog_file)
1234 write_iolog_close(td);
1236 td_set_runstate(td, TD_EXITED);
1237 return (void *) (uintptr_t) td->error;
1242 * We cannot pass the td data into a forked process, so attach the td and
1243 * pass it to the thread worker.
1245 static int fork_main(int shmid, int offset)
1247 struct thread_data *td;
1251 data = shmat(shmid, NULL, 0);
1252 if (data == (void *) -1) {
1260 * HP-UX inherits shm mappings?
1265 td = data + offset * sizeof(struct thread_data);
1266 ret = thread_main(td);
1268 return (int) (uintptr_t) ret;
1272 * Run over the job map and reap the threads that have exited, if any.
1274 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1275 unsigned int *m_rate)
1277 struct thread_data *td;
1278 unsigned int cputhreads, realthreads, pending;
1282 * reap exited threads (TD_EXITED -> TD_REAPED)
1284 realthreads = pending = cputhreads = 0;
1285 for_each_td(td, i) {
1289 * ->io_ops is NULL for a thread that has closed its
1292 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1301 if (td->runstate == TD_REAPED)
1303 if (td->o.use_thread) {
1304 if (td->runstate == TD_EXITED) {
1305 td_set_runstate(td, TD_REAPED);
1312 if (td->runstate == TD_EXITED)
1316 * check if someone quit or got killed in an unusual way
1318 ret = waitpid(td->pid, &status, flags);
1320 if (errno == ECHILD) {
1321 log_err("fio: pid=%d disappeared %d\n",
1322 (int) td->pid, td->runstate);
1324 td_set_runstate(td, TD_REAPED);
1328 } else if (ret == td->pid) {
1329 if (WIFSIGNALED(status)) {
1330 int sig = WTERMSIG(status);
1333 log_err("fio: pid=%d, got signal=%d\n",
1334 (int) td->pid, sig);
1336 td_set_runstate(td, TD_REAPED);
1339 if (WIFEXITED(status)) {
1340 if (WEXITSTATUS(status) && !td->error)
1341 td->error = WEXITSTATUS(status);
1343 td_set_runstate(td, TD_REAPED);
1349 * thread is not dead, continue
1355 (*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
1356 (*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
1363 done_secs += mtime_since_now(&td->epoch) / 1000;
1366 if (*nr_running == cputhreads && !pending && realthreads)
1367 fio_terminate_threads(TERMINATE_ALL);
1371 * Main function for kicking off and reaping jobs, as needed.
1373 static void run_threads(void)
1375 struct thread_data *td;
1376 unsigned long spent;
1377 unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
1379 if (fio_gtod_offload && fio_start_gtod_thread())
1384 nr_thread = nr_process = 0;
1385 for_each_td(td, i) {
1386 if (td->o.use_thread)
1392 if (!terse_output) {
1393 log_info("Starting ");
1395 log_info("%d thread%s", nr_thread,
1396 nr_thread > 1 ? "s" : "");
1400 log_info("%d process%s", nr_process,
1401 nr_process > 1 ? "es" : "");
1407 todo = thread_number;
1410 m_rate = t_rate = 0;
1412 for_each_td(td, i) {
1413 print_status_init(td->thread_number - 1);
1415 if (!td->o.create_serialize)
1419 * do file setup here so it happens sequentially,
1420 * we don't want X number of threads getting their
1421 * client data interspersed on disk
1423 if (setup_files(td)) {
1426 log_err("fio: pid=%d, err=%d/%s\n",
1427 (int) td->pid, td->error, td->verror);
1428 td_set_runstate(td, TD_REAPED);
1435 * for sharing to work, each job must always open
1436 * its own files. so close them, if we opened them
1439 for_each_file(td, f, j) {
1440 if (fio_file_open(f))
1441 td_io_close_file(td, f);
1449 struct thread_data *map[REAL_MAX_JOBS];
1450 struct timeval this_start;
1451 int this_jobs = 0, left;
1454 * create threads (TD_NOT_CREATED -> TD_CREATED)
1456 for_each_td(td, i) {
1457 if (td->runstate != TD_NOT_CREATED)
1461 * never got a chance to start, killed by other
1462 * thread for some reason
1464 if (td->terminate) {
1469 if (td->o.start_delay) {
1470 spent = mtime_since_genesis();
1472 if (td->o.start_delay * 1000 > spent)
1476 if (td->o.stonewall && (nr_started || nr_running)) {
1477 dprint(FD_PROCESS, "%s: stonewall wait\n",
1485 * Set state to created. Thread will transition
1486 * to TD_INITIALIZED when it's done setting up.
1488 td_set_runstate(td, TD_CREATED);
1489 map[this_jobs++] = td;
1492 if (td->o.use_thread) {
1495 dprint(FD_PROCESS, "will pthread_create\n");
1496 ret = pthread_create(&td->thread, NULL,
1499 log_err("pthread_create: %s\n",
1504 ret = pthread_detach(td->thread);
1506 log_err("pthread_detach: %s",
1510 dprint(FD_PROCESS, "will fork\n");
1513 int ret = fork_main(shm_id, i);
1516 } else if (i == fio_debug_jobno)
1517 *fio_debug_jobp = pid;
1519 dprint(FD_MUTEX, "wait on startup_mutex\n");
1520 if (fio_mutex_down_timeout(startup_mutex, 10)) {
1521 log_err("fio: job startup hung? exiting.\n");
1522 fio_terminate_threads(TERMINATE_ALL);
1527 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1531 * Wait for the started threads to transition to
1534 fio_gettime(&this_start, NULL);
1536 while (left && !fio_abort) {
1537 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1542 for (i = 0; i < this_jobs; i++) {
1546 if (td->runstate == TD_INITIALIZED) {
1549 } else if (td->runstate >= TD_EXITED) {
1553 nr_running++; /* work-around... */
1559 log_err("fio: %d job%s failed to start\n", left,
1560 left > 1 ? "s" : "");
1561 for (i = 0; i < this_jobs; i++) {
1565 kill(td->pid, SIGTERM);
1571 * start created threads (TD_INITIALIZED -> TD_RUNNING).
1573 for_each_td(td, i) {
1574 if (td->runstate != TD_INITIALIZED)
1577 if (in_ramp_time(td))
1578 td_set_runstate(td, TD_RAMP);
1580 td_set_runstate(td, TD_RUNNING);
1583 m_rate += td->o.ratemin[0] + td->o.ratemin[1];
1584 t_rate += td->o.rate[0] + td->o.rate[1];
1586 fio_mutex_up(td->mutex);
1589 reap_threads(&nr_running, &t_rate, &m_rate);
1595 while (nr_running) {
1596 reap_threads(&nr_running, &t_rate, &m_rate);
1603 void wait_for_disk_thread_exit(void)
1605 fio_mutex_down(disk_thread_mutex);
1608 static void *disk_thread_main(void *data)
1612 fio_mutex_up(startup_mutex);
1614 while (threads && !ret) {
1615 usleep(DISK_UTIL_MSEC * 1000);
1618 ret = update_io_ticks();
1621 print_thread_status();
1624 fio_mutex_up(disk_thread_mutex);
1628 static int create_disk_util_thread(void)
1634 disk_thread_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
1636 ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
1638 fio_mutex_remove(disk_thread_mutex);
1639 log_err("Can't create disk util thread: %s\n", strerror(ret));
1643 ret = pthread_detach(disk_util_thread);
1645 fio_mutex_remove(disk_thread_mutex);
1646 log_err("Can't detatch disk util thread: %s\n", strerror(ret));
1650 dprint(FD_MUTEX, "wait on startup_mutex\n");
1651 fio_mutex_down(startup_mutex);
1652 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1656 int fio_backend(void)
1658 struct thread_data *td;
1662 if (load_profile(exec_profile))
1665 exec_profile = NULL;
1671 setup_log(&agg_io_log[DDIR_READ], 0, IO_LOG_TYPE_BW);
1672 setup_log(&agg_io_log[DDIR_WRITE], 0, IO_LOG_TYPE_BW);
1675 startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
1676 if (startup_mutex == NULL)
1678 writeout_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1679 if (writeout_mutex == NULL)
1683 create_disk_util_thread();
1685 cgroup_list = smalloc(sizeof(*cgroup_list));
1686 INIT_FLIST_HEAD(cgroup_list);
1693 __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1694 __finish_log(agg_io_log[DDIR_WRITE],
1695 "agg-write_bw.log");
1700 fio_options_free(td);
1703 cgroup_kill(cgroup_list);
1707 fio_mutex_remove(startup_mutex);
1708 fio_mutex_remove(writeout_mutex);
1709 fio_mutex_remove(disk_thread_mutex);