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 unsigned long arch_flags = 0;
75 struct io_log *agg_io_log[2];
77 #define TERMINATE_ALL (-1)
78 #define JOB_START_TIMEOUT (5 * 1000)
80 void td_set_runstate(struct thread_data *td, int runstate)
82 if (td->runstate == runstate)
85 dprint(FD_PROCESS, "pid=%d: runstate %d -> %d\n", (int) td->pid,
86 td->runstate, runstate);
87 td->runstate = runstate;
90 static void terminate_threads(int group_id)
92 struct thread_data *td;
95 dprint(FD_PROCESS, "terminate group_id=%d\n", group_id);
98 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
99 dprint(FD_PROCESS, "setting terminate on %s/%d\n",
100 td->o.name, (int) td->pid);
102 td->o.start_delay = 0;
105 * if the thread is running, just let it exit
109 else if (td->runstate < TD_RAMP)
110 kill(td->pid, SIGTERM);
112 struct ioengine_ops *ops = td->io_ops;
114 if (ops && (ops->flags & FIO_SIGTERM))
115 kill(td->pid, SIGTERM);
121 static void sig_int(int sig)
124 log_info("\nfio: terminating on signal %d\n", sig);
127 terminate_threads(TERMINATE_ALL);
131 static void *disk_thread_main(void *data)
133 fio_mutex_up(startup_mutex);
136 usleep(DISK_UTIL_MSEC * 1000);
140 print_thread_status();
146 static int create_disk_util_thread(void)
150 ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
152 log_err("Can't create disk util thread: %s\n", strerror(ret));
156 ret = pthread_detach(disk_util_thread);
158 log_err("Can't detatch disk util thread: %s\n", strerror(ret));
162 dprint(FD_MUTEX, "wait on startup_mutex\n");
163 fio_mutex_down(startup_mutex);
164 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
168 static void set_sig_handlers(void)
170 struct sigaction act;
172 memset(&act, 0, sizeof(act));
173 act.sa_handler = sig_int;
174 act.sa_flags = SA_RESTART;
175 sigaction(SIGINT, &act, NULL);
177 memset(&act, 0, sizeof(act));
178 act.sa_handler = sig_int;
179 act.sa_flags = SA_RESTART;
180 sigaction(SIGTERM, &act, NULL);
184 * Check if we are above the minimum rate given.
186 static int __check_min_rate(struct thread_data *td, struct timeval *now,
189 unsigned long long bytes = 0;
190 unsigned long iops = 0;
193 unsigned int ratemin = 0;
194 unsigned int rate_iops = 0;
195 unsigned int rate_iops_min = 0;
197 assert(ddir_rw(ddir));
199 if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
203 * allow a 2 second settle period in the beginning
205 if (mtime_since(&td->start, now) < 2000)
208 iops += td->io_blocks[ddir];
209 bytes += td->this_io_bytes[ddir];
210 ratemin += td->o.ratemin[ddir];
211 rate_iops += td->o.rate_iops[ddir];
212 rate_iops_min += td->o.rate_iops_min[ddir];
215 * if rate blocks is set, sample is running
217 if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
218 spent = mtime_since(&td->lastrate[ddir], now);
219 if (spent < td->o.ratecycle)
222 if (td->o.rate[ddir]) {
224 * check bandwidth specified rate
226 if (bytes < td->rate_bytes[ddir]) {
227 log_err("%s: min rate %u not met\n", td->o.name,
231 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
232 if (rate < ratemin ||
233 bytes < td->rate_bytes[ddir]) {
234 log_err("%s: min rate %u not met, got"
235 " %luKB/sec\n", td->o.name,
242 * checks iops specified rate
244 if (iops < rate_iops) {
245 log_err("%s: min iops rate %u not met\n",
246 td->o.name, rate_iops);
249 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
250 if (rate < rate_iops_min ||
251 iops < td->rate_blocks[ddir]) {
252 log_err("%s: min iops rate %u not met,"
253 " got %lu\n", td->o.name,
254 rate_iops_min, rate);
260 td->rate_bytes[ddir] = bytes;
261 td->rate_blocks[ddir] = iops;
262 memcpy(&td->lastrate[ddir], now, sizeof(*now));
266 static int check_min_rate(struct thread_data *td, struct timeval *now,
267 unsigned long *bytes_done)
272 ret |= __check_min_rate(td, now, 0);
274 ret |= __check_min_rate(td, now, 1);
279 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
283 if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
290 * When job exits, we can cancel the in-flight IO if we are using async
291 * io. Attempt to do so.
293 static void cleanup_pending_aio(struct thread_data *td)
295 struct flist_head *entry, *n;
300 * get immediately available events, if any
302 r = io_u_queued_complete(td, 0, NULL);
307 * now cancel remaining active events
309 if (td->io_ops->cancel) {
310 flist_for_each_safe(entry, n, &td->io_u_busylist) {
311 io_u = flist_entry(entry, struct io_u, list);
314 * if the io_u isn't in flight, then that generally
315 * means someone leaked an io_u. complain but fix
316 * it up, so we don't stall here.
318 if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
319 log_err("fio: non-busy IO on busy list\n");
322 r = td->io_ops->cancel(td, io_u);
330 r = io_u_queued_complete(td, td->cur_depth, NULL);
334 * Helper to handle the final sync of a file. Works just like the normal
335 * io path, just does everything sync.
337 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
339 struct io_u *io_u = __get_io_u(td);
345 io_u->ddir = DDIR_SYNC;
348 if (td_io_prep(td, io_u)) {
354 ret = td_io_queue(td, io_u);
356 td_verror(td, io_u->error, "td_io_queue");
359 } else if (ret == FIO_Q_QUEUED) {
360 if (io_u_queued_complete(td, 1, NULL) < 0)
362 } else if (ret == FIO_Q_COMPLETED) {
364 td_verror(td, io_u->error, "td_io_queue");
368 if (io_u_sync_complete(td, io_u, NULL) < 0)
370 } else if (ret == FIO_Q_BUSY) {
371 if (td_io_commit(td))
379 static inline void __update_tv_cache(struct thread_data *td)
381 fio_gettime(&td->tv_cache, NULL);
384 static inline void update_tv_cache(struct thread_data *td)
386 if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
387 __update_tv_cache(td);
390 static int break_on_this_error(struct thread_data *td, int *retptr)
394 if (ret < 0 || td->error) {
397 if (!td->o.continue_on_error)
405 if (td_non_fatal_error(err)) {
407 * Continue with the I/Os in case of
410 update_error_count(td, err);
414 } else if (td->o.fill_device && err == ENOSPC) {
416 * We expect to hit this error if
417 * fill_device option is set.
424 * Stop the I/O in case of a fatal
427 update_error_count(td, err);
436 * The main verify engine. Runs over the writes we previously submitted,
437 * reads the blocks back in, and checks the crc/md5 of the data.
439 static void do_verify(struct thread_data *td)
446 dprint(FD_VERIFY, "starting loop\n");
449 * sync io first and invalidate cache, to make sure we really
452 for_each_file(td, f, i) {
453 if (!fio_file_open(f))
455 if (fio_io_sync(td, f))
457 if (file_invalidate_cache(td, f))
464 td_set_runstate(td, TD_VERIFYING);
467 while (!td->terminate) {
472 if (runtime_exceeded(td, &td->tv_cache)) {
473 __update_tv_cache(td);
474 if (runtime_exceeded(td, &td->tv_cache)) {
480 io_u = __get_io_u(td);
484 if (get_next_verify(td, io_u)) {
489 if (td_io_prep(td, io_u)) {
494 if (td->o.verify_async)
495 io_u->end_io = verify_io_u_async;
497 io_u->end_io = verify_io_u;
499 ret = td_io_queue(td, io_u);
501 case FIO_Q_COMPLETED:
504 clear_io_u(td, io_u);
505 } else if (io_u->resid) {
506 int bytes = io_u->xfer_buflen - io_u->resid;
512 td_verror(td, EIO, "full resid");
517 io_u->xfer_buflen = io_u->resid;
518 io_u->xfer_buf += bytes;
519 io_u->offset += bytes;
521 if (ddir_rw(io_u->ddir))
522 td->ts.short_io_u[io_u->ddir]++;
525 if (io_u->offset == f->real_file_size)
528 requeue_io_u(td, &io_u);
531 ret = io_u_sync_complete(td, io_u, NULL);
539 requeue_io_u(td, &io_u);
540 ret2 = td_io_commit(td);
546 td_verror(td, -ret, "td_io_queue");
550 if (break_on_this_error(td, &ret))
554 * if we can queue more, do so. but check if there are
555 * completed io_u's first. Note that we can get BUSY even
556 * without IO queued, if the system is resource starved.
558 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
559 if (full || !td->o.iodepth_batch_complete) {
560 min_events = min(td->o.iodepth_batch_complete,
562 if (full && !min_events && td->o.iodepth_batch_complete != 0)
567 * Reap required number of io units, if any,
568 * and do the verification on them through
569 * the callback handler
571 if (io_u_queued_complete(td, min_events, NULL) < 0) {
575 } while (full && (td->cur_depth > td->o.iodepth_low));
582 min_events = td->cur_depth;
585 ret = io_u_queued_complete(td, min_events, NULL);
587 cleanup_pending_aio(td);
589 td_set_runstate(td, TD_RUNNING);
591 dprint(FD_VERIFY, "exiting loop\n");
595 * Main IO worker function. It retrieves io_u's to process and queues
596 * and reaps them, checking for rate and errors along the way.
598 static void do_io(struct thread_data *td)
603 if (in_ramp_time(td))
604 td_set_runstate(td, TD_RAMP);
606 td_set_runstate(td, TD_RUNNING);
608 while ( (td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
609 (!flist_empty(&td->trim_list)) ||
610 ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) ) {
611 struct timeval comp_time;
612 unsigned long bytes_done[2] = { 0, 0 };
622 if (runtime_exceeded(td, &td->tv_cache)) {
623 __update_tv_cache(td);
624 if (runtime_exceeded(td, &td->tv_cache)) {
635 * Add verification end_io handler, if asked to verify
636 * a previously written file.
638 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
640 if (td->o.verify_async)
641 io_u->end_io = verify_io_u_async;
643 io_u->end_io = verify_io_u;
644 td_set_runstate(td, TD_VERIFYING);
645 } else if (in_ramp_time(td))
646 td_set_runstate(td, TD_RAMP);
648 td_set_runstate(td, TD_RUNNING);
650 ret = td_io_queue(td, io_u);
652 case FIO_Q_COMPLETED:
655 clear_io_u(td, io_u);
656 } else if (io_u->resid) {
657 int bytes = io_u->xfer_buflen - io_u->resid;
658 struct fio_file *f = io_u->file;
664 td_verror(td, EIO, "full resid");
669 io_u->xfer_buflen = io_u->resid;
670 io_u->xfer_buf += bytes;
671 io_u->offset += bytes;
673 if (ddir_rw(io_u->ddir))
674 td->ts.short_io_u[io_u->ddir]++;
676 if (io_u->offset == f->real_file_size)
679 requeue_io_u(td, &io_u);
682 if (__should_check_rate(td, 0) ||
683 __should_check_rate(td, 1))
684 fio_gettime(&comp_time, NULL);
686 ret = io_u_sync_complete(td, io_u, bytes_done);
693 * if the engine doesn't have a commit hook,
694 * the io_u is really queued. if it does have such
695 * a hook, it has to call io_u_queued() itself.
697 if (td->io_ops->commit == NULL)
698 io_u_queued(td, io_u);
701 requeue_io_u(td, &io_u);
702 ret2 = td_io_commit(td);
712 if (break_on_this_error(td, &ret))
716 * See if we need to complete some commands. Note that we
717 * can get BUSY even without IO queued, if the system is
720 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
721 if (full || !td->o.iodepth_batch_complete) {
722 min_evts = min(td->o.iodepth_batch_complete,
724 if (full && !min_evts && td->o.iodepth_batch_complete != 0)
727 if (__should_check_rate(td, 0) ||
728 __should_check_rate(td, 1))
729 fio_gettime(&comp_time, NULL);
732 ret = io_u_queued_complete(td, min_evts, bytes_done);
736 } while (full && (td->cur_depth > td->o.iodepth_low));
741 if (!(bytes_done[0] + bytes_done[1]))
744 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
745 if (check_min_rate(td, &comp_time, bytes_done)) {
746 if (exitall_on_terminate)
747 terminate_threads(td->groupid);
748 td_verror(td, EIO, "check_min_rate");
753 if (td->o.thinktime) {
754 unsigned long long b;
756 b = td->io_blocks[0] + td->io_blocks[1];
757 if (!(b % td->o.thinktime_blocks)) {
760 if (td->o.thinktime_spin)
761 usec_spin(td->o.thinktime_spin);
763 left = td->o.thinktime - td->o.thinktime_spin;
765 usec_sleep(td, left);
770 if (td->trim_entries)
771 printf("trim entries %ld\n", td->trim_entries);
773 if (td->o.fill_device && td->error == ENOSPC) {
782 ret = io_u_queued_complete(td, i, NULL);
783 if (td->o.fill_device && td->error == ENOSPC)
787 if (should_fsync(td) && td->o.end_fsync) {
788 td_set_runstate(td, TD_FSYNCING);
790 for_each_file(td, f, i) {
791 if (!fio_file_open(f))
797 cleanup_pending_aio(td);
800 * stop job if we failed doing any IO
802 if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
806 static void cleanup_io_u(struct thread_data *td)
808 struct flist_head *entry, *n;
811 flist_for_each_safe(entry, n, &td->io_u_freelist) {
812 io_u = flist_entry(entry, struct io_u, list);
814 flist_del(&io_u->list);
815 fio_memfree(io_u, sizeof(*io_u));
821 static int init_io_u(struct thread_data *td)
825 int cl_align, i, max_units;
828 max_units = td->o.iodepth;
829 max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
830 td->orig_buffer_size = (unsigned long long) max_bs
831 * (unsigned long long) max_units;
833 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
836 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
837 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
840 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
841 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
845 if (allocate_io_mem(td))
848 if (td->o.odirect || td->o.mem_align ||
849 (td->io_ops->flags & FIO_RAWIO))
850 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
854 cl_align = os_cache_line_size();
856 for (i = 0; i < max_units; i++) {
862 ptr = fio_memalign(cl_align, sizeof(*io_u));
864 log_err("fio: unable to allocate aligned memory\n");
869 memset(io_u, 0, sizeof(*io_u));
870 INIT_FLIST_HEAD(&io_u->list);
871 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
873 if (!(td->io_ops->flags & FIO_NOIO)) {
874 io_u->buf = p + max_bs * i;
875 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
878 io_u_fill_buffer(td, io_u, max_bs);
879 if (td_write(td) && td->o.verify_pattern_bytes) {
881 * Fill the buffer with the pattern if we are
882 * going to be doing writes.
884 fill_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
889 io_u->flags = IO_U_F_FREE;
890 flist_add(&io_u->list, &td->io_u_freelist);
896 static int switch_ioscheduler(struct thread_data *td)
898 char tmp[256], tmp2[128];
902 if (td->io_ops->flags & FIO_DISKLESSIO)
905 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
907 f = fopen(tmp, "r+");
909 if (errno == ENOENT) {
910 log_err("fio: os or kernel doesn't support IO scheduler"
914 td_verror(td, errno, "fopen iosched");
921 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
922 if (ferror(f) || ret != 1) {
923 td_verror(td, errno, "fwrite");
931 * Read back and check that the selected scheduler is now the default.
933 ret = fread(tmp, 1, sizeof(tmp), f);
934 if (ferror(f) || ret < 0) {
935 td_verror(td, errno, "fread");
940 sprintf(tmp2, "[%s]", td->o.ioscheduler);
941 if (!strstr(tmp, tmp2)) {
942 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
943 td_verror(td, EINVAL, "iosched_switch");
952 static int keep_running(struct thread_data *td)
954 unsigned long long io_done;
958 if (td->o.time_based)
965 io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
967 if (io_done < td->o.size)
973 static void reset_io_counters(struct thread_data *td)
975 td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
976 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
978 td->rate_bytes[0] = td->rate_bytes[1] = 0;
979 td->rate_blocks[0] = td->rate_blocks[1] = 0;
981 td->last_was_sync = 0;
984 * reset file done count if we are to start over
986 if (td->o.time_based || td->o.loops)
987 td->nr_done_files = 0;
990 void reset_all_stats(struct thread_data *td)
995 reset_io_counters(td);
997 for (i = 0; i < 2; i++) {
999 td->io_blocks[i] = 0;
1000 td->io_issues[i] = 0;
1001 td->ts.total_io_u[i] = 0;
1004 fio_gettime(&tv, NULL);
1005 td->ts.runtime[0] = 0;
1006 td->ts.runtime[1] = 0;
1007 memcpy(&td->epoch, &tv, sizeof(tv));
1008 memcpy(&td->start, &tv, sizeof(tv));
1011 static void clear_io_state(struct thread_data *td)
1016 reset_io_counters(td);
1019 for_each_file(td, f, i)
1020 fio_file_clear_done(f);
1023 * Set the same seed to get repeatable runs
1025 td_fill_rand_seeds(td);
1028 static int exec_string(const char *string)
1030 int ret, newlen = strlen(string) + 1 + 8;
1033 str = malloc(newlen);
1034 sprintf(str, "sh -c %s", string);
1038 log_err("fio: exec of cmd <%s> failed\n", str);
1045 * Entry point for the thread based jobs. The process based jobs end up
1046 * here as well, after a little setup.
1048 static void *thread_main(void *data)
1050 unsigned long long elapsed;
1051 struct thread_data *td = data;
1052 pthread_condattr_t attr;
1055 if (!td->o.use_thread) {
1061 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1063 INIT_FLIST_HEAD(&td->io_u_freelist);
1064 INIT_FLIST_HEAD(&td->io_u_busylist);
1065 INIT_FLIST_HEAD(&td->io_u_requeues);
1066 INIT_FLIST_HEAD(&td->io_log_list);
1067 INIT_FLIST_HEAD(&td->io_hist_list);
1068 INIT_FLIST_HEAD(&td->verify_list);
1069 INIT_FLIST_HEAD(&td->trim_list);
1070 pthread_mutex_init(&td->io_u_lock, NULL);
1071 td->io_hist_tree = RB_ROOT;
1073 pthread_condattr_init(&attr);
1074 pthread_cond_init(&td->verify_cond, &attr);
1075 pthread_cond_init(&td->free_cond, &attr);
1077 td_set_runstate(td, TD_INITIALIZED);
1078 dprint(FD_MUTEX, "up startup_mutex\n");
1079 fio_mutex_up(startup_mutex);
1080 dprint(FD_MUTEX, "wait on td->mutex\n");
1081 fio_mutex_down(td->mutex);
1082 dprint(FD_MUTEX, "done waiting on td->mutex\n");
1085 * the ->mutex mutex is now no longer used, close it to avoid
1086 * eating a file descriptor
1088 fio_mutex_remove(td->mutex);
1091 * A new gid requires privilege, so we need to do this before setting
1094 if (td->o.gid != -1U && setgid(td->o.gid)) {
1095 td_verror(td, errno, "setgid");
1098 if (td->o.uid != -1U && setuid(td->o.uid)) {
1099 td_verror(td, errno, "setuid");
1104 * If we have a gettimeofday() thread, make sure we exclude that
1105 * thread from this job
1108 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1111 * Set affinity first, in case it has an impact on the memory
1114 if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1115 td_verror(td, errno, "cpu_set_affinity");
1120 * May alter parameters that init_io_u() will use, so we need to
1129 if (td->o.verify_async && verify_async_init(td))
1132 if (td->ioprio_set) {
1133 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1134 td_verror(td, errno, "ioprio_set");
1139 if (td->o.cgroup_weight && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1142 if (nice(td->o.nice) == -1) {
1143 td_verror(td, errno, "nice");
1147 if (td->o.ioscheduler && switch_ioscheduler(td))
1150 if (!td->o.create_serialize && setup_files(td))
1156 if (init_random_map(td))
1159 if (td->o.exec_prerun) {
1160 if (exec_string(td->o.exec_prerun))
1164 if (td->o.pre_read) {
1165 if (pre_read_files(td) < 0)
1169 fio_gettime(&td->epoch, NULL);
1170 getrusage(RUSAGE_SELF, &td->ts.ru_start);
1173 while (keep_running(td)) {
1174 fio_gettime(&td->start, NULL);
1175 memcpy(&td->ts.stat_sample_time[0], &td->start,
1177 memcpy(&td->ts.stat_sample_time[1], &td->start,
1179 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1181 if (td->o.ratemin[0] || td->o.ratemin[1])
1182 memcpy(&td->lastrate, &td->ts.stat_sample_time,
1183 sizeof(td->lastrate));
1188 prune_io_piece_log(td);
1194 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1195 elapsed = utime_since_now(&td->start);
1196 td->ts.runtime[DDIR_READ] += elapsed;
1198 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1199 elapsed = utime_since_now(&td->start);
1200 td->ts.runtime[DDIR_WRITE] += elapsed;
1203 if (td->error || td->terminate)
1206 if (!td->o.do_verify ||
1207 td->o.verify == VERIFY_NONE ||
1208 (td->io_ops->flags & FIO_UNIDIR))
1213 fio_gettime(&td->start, NULL);
1217 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1219 if (td->error || td->terminate)
1223 update_rusage_stat(td);
1224 td->ts.runtime[0] = (td->ts.runtime[0] + 999) / 1000;
1225 td->ts.runtime[1] = (td->ts.runtime[1] + 999) / 1000;
1226 td->ts.total_run_time = mtime_since_now(&td->epoch);
1227 td->ts.io_bytes[0] = td->io_bytes[0];
1228 td->ts.io_bytes[1] = td->io_bytes[1];
1230 fio_mutex_down(writeout_mutex);
1231 if (td->ts.bw_log) {
1232 if (td->o.bw_log_file) {
1233 finish_log_named(td, td->ts.bw_log,
1234 td->o.bw_log_file, "bw");
1236 finish_log(td, td->ts.bw_log, "bw");
1238 if (td->ts.lat_log) {
1239 if (td->o.lat_log_file) {
1240 finish_log_named(td, td->ts.lat_log,
1241 td->o.lat_log_file, "lat");
1243 finish_log(td, td->ts.lat_log, "lat");
1245 if (td->ts.slat_log) {
1246 if (td->o.lat_log_file) {
1247 finish_log_named(td, td->ts.slat_log,
1248 td->o.lat_log_file, "slat");
1250 finish_log(td, td->ts.slat_log, "slat");
1252 if (td->ts.clat_log) {
1253 if (td->o.lat_log_file) {
1254 finish_log_named(td, td->ts.clat_log,
1255 td->o.lat_log_file, "clat");
1257 finish_log(td, td->ts.clat_log, "clat");
1259 fio_mutex_up(writeout_mutex);
1260 if (td->o.exec_postrun)
1261 exec_string(td->o.exec_postrun);
1263 if (exitall_on_terminate)
1264 terminate_threads(td->groupid);
1268 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1271 if (td->o.verify_async)
1272 verify_async_exit(td);
1274 close_and_free_files(td);
1277 cgroup_shutdown(td, &cgroup_mnt);
1279 if (td->o.cpumask_set) {
1280 int ret = fio_cpuset_exit(&td->o.cpumask);
1282 td_verror(td, ret, "fio_cpuset_exit");
1286 * do this very late, it will log file closing as well
1288 if (td->o.write_iolog_file)
1289 write_iolog_close(td);
1291 options_mem_free(td);
1292 td_set_runstate(td, TD_EXITED);
1293 return (void *) (unsigned long) td->error;
1297 * We cannot pass the td data into a forked process, so attach the td and
1298 * pass it to the thread worker.
1300 static int fork_main(int shmid, int offset)
1302 struct thread_data *td;
1306 data = shmat(shmid, NULL, 0);
1307 if (data == (void *) -1) {
1315 * HP-UX inherits shm mappings?
1320 td = data + offset * sizeof(struct thread_data);
1321 ret = thread_main(td);
1323 return (int) (unsigned long) ret;
1327 * Run over the job map and reap the threads that have exited, if any.
1329 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1331 struct thread_data *td;
1332 int i, cputhreads, realthreads, pending, status, ret;
1335 * reap exited threads (TD_EXITED -> TD_REAPED)
1337 realthreads = pending = cputhreads = 0;
1338 for_each_td(td, i) {
1342 * ->io_ops is NULL for a thread that has closed its
1345 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1354 if (td->runstate == TD_REAPED)
1356 if (td->o.use_thread) {
1357 if (td->runstate == TD_EXITED) {
1358 td_set_runstate(td, TD_REAPED);
1365 if (td->runstate == TD_EXITED)
1369 * check if someone quit or got killed in an unusual way
1371 ret = waitpid(td->pid, &status, flags);
1373 if (errno == ECHILD) {
1374 log_err("fio: pid=%d disappeared %d\n",
1375 (int) td->pid, td->runstate);
1376 td_set_runstate(td, TD_REAPED);
1380 } else if (ret == td->pid) {
1381 if (WIFSIGNALED(status)) {
1382 int sig = WTERMSIG(status);
1385 log_err("fio: pid=%d, got signal=%d\n",
1386 (int) td->pid, sig);
1387 td_set_runstate(td, TD_REAPED);
1390 if (WIFEXITED(status)) {
1391 if (WEXITSTATUS(status) && !td->error)
1392 td->error = WEXITSTATUS(status);
1394 td_set_runstate(td, TD_REAPED);
1400 * thread is not dead, continue
1406 (*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
1407 (*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
1414 done_secs += mtime_since_now(&td->epoch) / 1000;
1417 if (*nr_running == cputhreads && !pending && realthreads)
1418 terminate_threads(TERMINATE_ALL);
1421 static void *gtod_thread_main(void *data)
1423 fio_mutex_up(startup_mutex);
1426 * As long as we have jobs around, update the clock. It would be nice
1427 * to have some way of NOT hammering that CPU with gettimeofday(),
1428 * but I'm not sure what to use outside of a simple CPU nop to relax
1429 * it - we don't want to lose precision.
1439 static int fio_start_gtod_thread(void)
1441 pthread_attr_t attr;
1444 pthread_attr_init(&attr);
1445 pthread_attr_setstacksize(&attr, PTHREAD_STACK_MIN);
1446 ret = pthread_create(>od_thread, &attr, gtod_thread_main, NULL);
1447 pthread_attr_destroy(&attr);
1449 log_err("Can't create gtod thread: %s\n", strerror(ret));
1453 ret = pthread_detach(gtod_thread);
1455 log_err("Can't detatch gtod thread: %s\n", strerror(ret));
1459 dprint(FD_MUTEX, "wait on startup_mutex\n");
1460 fio_mutex_down(startup_mutex);
1461 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1466 * Main function for kicking off and reaping jobs, as needed.
1468 static void run_threads(void)
1470 struct thread_data *td;
1471 unsigned long spent;
1472 int i, todo, nr_running, m_rate, t_rate, nr_started;
1474 if (fio_pin_memory())
1477 if (fio_gtod_offload && fio_start_gtod_thread())
1480 if (!terse_output) {
1481 log_info("Starting ");
1483 log_info("%d thread%s", nr_thread,
1484 nr_thread > 1 ? "s" : "");
1488 log_info("%d process%s", nr_process,
1489 nr_process > 1 ? "es" : "");
1497 todo = thread_number;
1500 m_rate = t_rate = 0;
1502 for_each_td(td, i) {
1503 print_status_init(td->thread_number - 1);
1505 if (!td->o.create_serialize)
1509 * do file setup here so it happens sequentially,
1510 * we don't want X number of threads getting their
1511 * client data interspersed on disk
1513 if (setup_files(td)) {
1516 log_err("fio: pid=%d, err=%d/%s\n",
1517 (int) td->pid, td->error, td->verror);
1518 td_set_runstate(td, TD_REAPED);
1525 * for sharing to work, each job must always open
1526 * its own files. so close them, if we opened them
1529 for_each_file(td, f, j) {
1530 if (fio_file_open(f))
1531 td_io_close_file(td, f);
1539 struct thread_data *map[REAL_MAX_JOBS];
1540 struct timeval this_start;
1541 int this_jobs = 0, left;
1544 * create threads (TD_NOT_CREATED -> TD_CREATED)
1546 for_each_td(td, i) {
1547 if (td->runstate != TD_NOT_CREATED)
1551 * never got a chance to start, killed by other
1552 * thread for some reason
1554 if (td->terminate) {
1559 if (td->o.start_delay) {
1560 spent = mtime_since_genesis();
1562 if (td->o.start_delay * 1000 > spent)
1566 if (td->o.stonewall && (nr_started || nr_running)) {
1567 dprint(FD_PROCESS, "%s: stonewall wait\n",
1575 * Set state to created. Thread will transition
1576 * to TD_INITIALIZED when it's done setting up.
1578 td_set_runstate(td, TD_CREATED);
1579 map[this_jobs++] = td;
1582 if (td->o.use_thread) {
1585 dprint(FD_PROCESS, "will pthread_create\n");
1586 ret = pthread_create(&td->thread, NULL,
1589 log_err("pthread_create: %s\n",
1594 ret = pthread_detach(td->thread);
1596 log_err("pthread_detach: %s",
1600 dprint(FD_PROCESS, "will fork\n");
1603 int ret = fork_main(shm_id, i);
1606 } else if (i == fio_debug_jobno)
1607 *fio_debug_jobp = pid;
1609 dprint(FD_MUTEX, "wait on startup_mutex\n");
1610 if (fio_mutex_down_timeout(startup_mutex, 10)) {
1611 log_err("fio: job startup hung? exiting.\n");
1612 terminate_threads(TERMINATE_ALL);
1617 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1621 * Wait for the started threads to transition to
1624 fio_gettime(&this_start, NULL);
1626 while (left && !fio_abort) {
1627 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1632 for (i = 0; i < this_jobs; i++) {
1636 if (td->runstate == TD_INITIALIZED) {
1639 } else if (td->runstate >= TD_EXITED) {
1643 nr_running++; /* work-around... */
1649 log_err("fio: %d jobs failed to start\n", left);
1650 for (i = 0; i < this_jobs; i++) {
1654 kill(td->pid, SIGTERM);
1660 * start created threads (TD_INITIALIZED -> TD_RUNNING).
1662 for_each_td(td, i) {
1663 if (td->runstate != TD_INITIALIZED)
1666 if (in_ramp_time(td))
1667 td_set_runstate(td, TD_RAMP);
1669 td_set_runstate(td, TD_RUNNING);
1672 m_rate += td->o.ratemin[0] + td->o.ratemin[1];
1673 t_rate += td->o.rate[0] + td->o.rate[1];
1675 fio_mutex_up(td->mutex);
1678 reap_threads(&nr_running, &t_rate, &m_rate);
1684 while (nr_running) {
1685 reap_threads(&nr_running, &t_rate, &m_rate);
1693 int main(int argc, char *argv[], char *envp[])
1702 * We need locale for number printing, if it isn't set then just
1703 * go with the US format.
1705 if (!getenv("LC_NUMERIC"))
1706 setlocale(LC_NUMERIC, "en_US");
1708 ps = sysconf(_SC_PAGESIZE);
1710 log_err("Failed to get page size\n");
1717 fio_keywords_init();
1719 if (parse_options(argc, argv))
1722 if (exec_profile && load_profile(exec_profile))
1729 setup_log(&agg_io_log[DDIR_READ]);
1730 setup_log(&agg_io_log[DDIR_WRITE]);
1733 startup_mutex = fio_mutex_init(0);
1734 if (startup_mutex == NULL)
1736 writeout_mutex = fio_mutex_init(1);
1737 if (writeout_mutex == NULL)
1741 create_disk_util_thread();
1743 cgroup_list = smalloc(sizeof(*cgroup_list));
1744 INIT_FLIST_HEAD(cgroup_list);
1751 __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1752 __finish_log(agg_io_log[DDIR_WRITE],
1753 "agg-write_bw.log");
1757 cgroup_kill(cgroup_list);
1761 fio_mutex_remove(startup_mutex);
1762 fio_mutex_remove(writeout_mutex);