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
51 unsigned long page_mask;
52 unsigned long page_size;
54 #define PAGE_ALIGN(buf) \
55 (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
58 int thread_number = 0;
63 unsigned long done_secs = 0;
65 static struct fio_mutex *startup_mutex;
66 static struct fio_mutex *writeout_mutex;
67 static volatile int fio_abort;
68 static int exit_value;
69 static pthread_t gtod_thread;
70 static pthread_t disk_util_thread;
71 static struct flist_head *cgroup_list;
72 static char *cgroup_mnt;
74 unsigned long arch_flags = 0;
76 struct io_log *agg_io_log[2];
78 #define TERMINATE_ALL (-1)
79 #define JOB_START_TIMEOUT (5 * 1000)
81 void td_set_runstate(struct thread_data *td, int runstate)
83 if (td->runstate == runstate)
86 dprint(FD_PROCESS, "pid=%d: runstate %d -> %d\n", (int) td->pid,
87 td->runstate, runstate);
88 td->runstate = runstate;
91 static void terminate_threads(int group_id)
93 struct thread_data *td;
96 dprint(FD_PROCESS, "terminate group_id=%d\n", group_id);
99 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
100 dprint(FD_PROCESS, "setting terminate on %s/%d\n",
101 td->o.name, (int) td->pid);
103 td->o.start_delay = 0;
106 * if the thread is running, just let it exit
110 else if (td->runstate < TD_RAMP)
111 kill(td->pid, SIGTERM);
113 struct ioengine_ops *ops = td->io_ops;
115 if (ops && (ops->flags & FIO_SIGTERM))
116 kill(td->pid, SIGTERM);
122 static void sig_int(int sig)
125 log_info("\nfio: terminating on signal %d\n", sig);
129 terminate_threads(TERMINATE_ALL);
133 static void *disk_thread_main(void *data)
135 fio_mutex_up(startup_mutex);
138 usleep(DISK_UTIL_MSEC * 1000);
144 print_thread_status();
150 static int create_disk_util_thread(void)
154 ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
156 log_err("Can't create disk util thread: %s\n", strerror(ret));
160 ret = pthread_detach(disk_util_thread);
162 log_err("Can't detatch disk util thread: %s\n", strerror(ret));
166 dprint(FD_MUTEX, "wait on startup_mutex\n");
167 fio_mutex_down(startup_mutex);
168 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
172 static void set_sig_handlers(void)
174 struct sigaction act;
176 memset(&act, 0, sizeof(act));
177 act.sa_handler = sig_int;
178 act.sa_flags = SA_RESTART;
179 sigaction(SIGINT, &act, NULL);
181 memset(&act, 0, sizeof(act));
182 act.sa_handler = sig_int;
183 act.sa_flags = SA_RESTART;
184 sigaction(SIGTERM, &act, NULL);
188 * Check if we are above the minimum rate given.
190 static int __check_min_rate(struct thread_data *td, struct timeval *now,
193 unsigned long long bytes = 0;
194 unsigned long iops = 0;
197 unsigned int ratemin = 0;
198 unsigned int rate_iops = 0;
199 unsigned int rate_iops_min = 0;
201 assert(ddir_rw(ddir));
203 if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
207 * allow a 2 second settle period in the beginning
209 if (mtime_since(&td->start, now) < 2000)
212 iops += td->io_blocks[ddir];
213 bytes += td->this_io_bytes[ddir];
214 ratemin += td->o.ratemin[ddir];
215 rate_iops += td->o.rate_iops[ddir];
216 rate_iops_min += td->o.rate_iops_min[ddir];
219 * if rate blocks is set, sample is running
221 if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
222 spent = mtime_since(&td->lastrate[ddir], now);
223 if (spent < td->o.ratecycle)
226 if (td->o.rate[ddir]) {
228 * check bandwidth specified rate
230 if (bytes < td->rate_bytes[ddir]) {
231 log_err("%s: min rate %u not met\n", td->o.name,
235 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
236 if (rate < ratemin ||
237 bytes < td->rate_bytes[ddir]) {
238 log_err("%s: min rate %u not met, got"
239 " %luKB/sec\n", td->o.name,
246 * checks iops specified rate
248 if (iops < rate_iops) {
249 log_err("%s: min iops rate %u not met\n",
250 td->o.name, rate_iops);
253 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
254 if (rate < rate_iops_min ||
255 iops < td->rate_blocks[ddir]) {
256 log_err("%s: min iops rate %u not met,"
257 " got %lu\n", td->o.name,
258 rate_iops_min, rate);
264 td->rate_bytes[ddir] = bytes;
265 td->rate_blocks[ddir] = iops;
266 memcpy(&td->lastrate[ddir], now, sizeof(*now));
270 static int check_min_rate(struct thread_data *td, struct timeval *now,
271 unsigned long *bytes_done)
276 ret |= __check_min_rate(td, now, 0);
278 ret |= __check_min_rate(td, now, 1);
283 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
287 if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
294 * When job exits, we can cancel the in-flight IO if we are using async
295 * io. Attempt to do so.
297 static void cleanup_pending_aio(struct thread_data *td)
299 struct flist_head *entry, *n;
304 * get immediately available events, if any
306 r = io_u_queued_complete(td, 0, NULL);
311 * now cancel remaining active events
313 if (td->io_ops->cancel) {
314 flist_for_each_safe(entry, n, &td->io_u_busylist) {
315 io_u = flist_entry(entry, struct io_u, list);
318 * if the io_u isn't in flight, then that generally
319 * means someone leaked an io_u. complain but fix
320 * it up, so we don't stall here.
322 if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
323 log_err("fio: non-busy IO on busy list\n");
326 r = td->io_ops->cancel(td, io_u);
334 r = io_u_queued_complete(td, td->cur_depth, NULL);
338 * Helper to handle the final sync of a file. Works just like the normal
339 * io path, just does everything sync.
341 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
343 struct io_u *io_u = __get_io_u(td);
349 io_u->ddir = DDIR_SYNC;
352 if (td_io_prep(td, io_u)) {
358 ret = td_io_queue(td, io_u);
360 td_verror(td, io_u->error, "td_io_queue");
363 } else if (ret == FIO_Q_QUEUED) {
364 if (io_u_queued_complete(td, 1, NULL) < 0)
366 } else if (ret == FIO_Q_COMPLETED) {
368 td_verror(td, io_u->error, "td_io_queue");
372 if (io_u_sync_complete(td, io_u, NULL) < 0)
374 } else if (ret == FIO_Q_BUSY) {
375 if (td_io_commit(td))
383 static inline void __update_tv_cache(struct thread_data *td)
385 fio_gettime(&td->tv_cache, NULL);
388 static inline void update_tv_cache(struct thread_data *td)
390 if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
391 __update_tv_cache(td);
394 static int break_on_this_error(struct thread_data *td, int *retptr)
398 if (ret < 0 || td->error) {
401 if (!td->o.continue_on_error)
409 if (td_non_fatal_error(err)) {
411 * Continue with the I/Os in case of
414 update_error_count(td, err);
418 } else if (td->o.fill_device && err == ENOSPC) {
420 * We expect to hit this error if
421 * fill_device option is set.
428 * Stop the I/O in case of a fatal
431 update_error_count(td, err);
440 * The main verify engine. Runs over the writes we previously submitted,
441 * reads the blocks back in, and checks the crc/md5 of the data.
443 static void do_verify(struct thread_data *td)
450 dprint(FD_VERIFY, "starting loop\n");
453 * sync io first and invalidate cache, to make sure we really
456 for_each_file(td, f, i) {
457 if (!fio_file_open(f))
459 if (fio_io_sync(td, f))
461 if (file_invalidate_cache(td, f))
468 td_set_runstate(td, TD_VERIFYING);
471 while (!td->terminate) {
476 if (runtime_exceeded(td, &td->tv_cache)) {
477 __update_tv_cache(td);
478 if (runtime_exceeded(td, &td->tv_cache)) {
484 io_u = __get_io_u(td);
488 if (get_next_verify(td, io_u)) {
493 if (td_io_prep(td, io_u)) {
498 if (td->o.verify_async)
499 io_u->end_io = verify_io_u_async;
501 io_u->end_io = verify_io_u;
503 ret = td_io_queue(td, io_u);
505 case FIO_Q_COMPLETED:
508 clear_io_u(td, io_u);
509 } else if (io_u->resid) {
510 int bytes = io_u->xfer_buflen - io_u->resid;
516 td_verror(td, EIO, "full resid");
521 io_u->xfer_buflen = io_u->resid;
522 io_u->xfer_buf += bytes;
523 io_u->offset += bytes;
525 if (ddir_rw(io_u->ddir))
526 td->ts.short_io_u[io_u->ddir]++;
529 if (io_u->offset == f->real_file_size)
532 requeue_io_u(td, &io_u);
535 ret = io_u_sync_complete(td, io_u, NULL);
543 requeue_io_u(td, &io_u);
544 ret2 = td_io_commit(td);
550 td_verror(td, -ret, "td_io_queue");
554 if (break_on_this_error(td, &ret))
558 * if we can queue more, do so. but check if there are
559 * completed io_u's first. Note that we can get BUSY even
560 * without IO queued, if the system is resource starved.
562 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
563 if (full || !td->o.iodepth_batch_complete) {
564 min_events = min(td->o.iodepth_batch_complete,
566 if (full && !min_events && td->o.iodepth_batch_complete != 0)
571 * Reap required number of io units, if any,
572 * and do the verification on them through
573 * the callback handler
575 if (io_u_queued_complete(td, min_events, NULL) < 0) {
579 } while (full && (td->cur_depth > td->o.iodepth_low));
586 min_events = td->cur_depth;
589 ret = io_u_queued_complete(td, min_events, NULL);
591 cleanup_pending_aio(td);
593 td_set_runstate(td, TD_RUNNING);
595 dprint(FD_VERIFY, "exiting loop\n");
599 * Main IO worker function. It retrieves io_u's to process and queues
600 * and reaps them, checking for rate and errors along the way.
602 static void do_io(struct thread_data *td)
607 if (in_ramp_time(td))
608 td_set_runstate(td, TD_RAMP);
610 td_set_runstate(td, TD_RUNNING);
612 while ( (td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
613 (!flist_empty(&td->trim_list)) ||
614 ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) ) {
615 struct timeval comp_time;
616 unsigned long bytes_done[2] = { 0, 0 };
626 if (runtime_exceeded(td, &td->tv_cache)) {
627 __update_tv_cache(td);
628 if (runtime_exceeded(td, &td->tv_cache)) {
639 * Add verification end_io handler, if asked to verify
640 * a previously written file.
642 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
644 if (td->o.verify_async)
645 io_u->end_io = verify_io_u_async;
647 io_u->end_io = verify_io_u;
648 td_set_runstate(td, TD_VERIFYING);
649 } else if (in_ramp_time(td))
650 td_set_runstate(td, TD_RAMP);
652 td_set_runstate(td, TD_RUNNING);
654 ret = td_io_queue(td, io_u);
656 case FIO_Q_COMPLETED:
659 clear_io_u(td, io_u);
660 } else if (io_u->resid) {
661 int bytes = io_u->xfer_buflen - io_u->resid;
662 struct fio_file *f = io_u->file;
668 td_verror(td, EIO, "full resid");
673 io_u->xfer_buflen = io_u->resid;
674 io_u->xfer_buf += bytes;
675 io_u->offset += bytes;
677 if (ddir_rw(io_u->ddir))
678 td->ts.short_io_u[io_u->ddir]++;
680 if (io_u->offset == f->real_file_size)
683 requeue_io_u(td, &io_u);
686 if (__should_check_rate(td, 0) ||
687 __should_check_rate(td, 1))
688 fio_gettime(&comp_time, NULL);
690 ret = io_u_sync_complete(td, io_u, bytes_done);
697 * if the engine doesn't have a commit hook,
698 * the io_u is really queued. if it does have such
699 * a hook, it has to call io_u_queued() itself.
701 if (td->io_ops->commit == NULL)
702 io_u_queued(td, io_u);
705 requeue_io_u(td, &io_u);
706 ret2 = td_io_commit(td);
716 if (break_on_this_error(td, &ret))
720 * See if we need to complete some commands. Note that we
721 * can get BUSY even without IO queued, if the system is
724 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
725 if (full || !td->o.iodepth_batch_complete) {
726 min_evts = min(td->o.iodepth_batch_complete,
728 if (full && !min_evts && td->o.iodepth_batch_complete != 0)
731 if (__should_check_rate(td, 0) ||
732 __should_check_rate(td, 1))
733 fio_gettime(&comp_time, NULL);
736 ret = io_u_queued_complete(td, min_evts, bytes_done);
740 } while (full && (td->cur_depth > td->o.iodepth_low));
745 if (!(bytes_done[0] + bytes_done[1]))
748 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
749 if (check_min_rate(td, &comp_time, bytes_done)) {
750 if (exitall_on_terminate)
751 terminate_threads(td->groupid);
752 td_verror(td, EIO, "check_min_rate");
757 if (td->o.thinktime) {
758 unsigned long long b;
760 b = td->io_blocks[0] + td->io_blocks[1];
761 if (!(b % td->o.thinktime_blocks)) {
764 if (td->o.thinktime_spin)
765 usec_spin(td->o.thinktime_spin);
767 left = td->o.thinktime - td->o.thinktime_spin;
769 usec_sleep(td, left);
774 if (td->trim_entries)
775 printf("trim entries %ld\n", td->trim_entries);
777 if (td->o.fill_device && td->error == ENOSPC) {
786 ret = io_u_queued_complete(td, i, NULL);
787 if (td->o.fill_device && td->error == ENOSPC)
791 if (should_fsync(td) && td->o.end_fsync) {
792 td_set_runstate(td, TD_FSYNCING);
794 for_each_file(td, f, i) {
795 if (!fio_file_open(f))
801 cleanup_pending_aio(td);
804 * stop job if we failed doing any IO
806 if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
810 static void cleanup_io_u(struct thread_data *td)
812 struct flist_head *entry, *n;
815 flist_for_each_safe(entry, n, &td->io_u_freelist) {
816 io_u = flist_entry(entry, struct io_u, list);
818 flist_del(&io_u->list);
819 fio_memfree(io_u, sizeof(*io_u));
825 static int init_io_u(struct thread_data *td)
829 int cl_align, i, max_units;
832 max_units = td->o.iodepth;
833 max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
834 td->orig_buffer_size = (unsigned long long) max_bs
835 * (unsigned long long) max_units;
837 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
840 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
841 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
844 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
845 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
849 if (allocate_io_mem(td))
852 if (td->o.odirect || td->o.mem_align ||
853 (td->io_ops->flags & FIO_RAWIO))
854 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
858 cl_align = os_cache_line_size();
860 for (i = 0; i < max_units; i++) {
866 ptr = fio_memalign(cl_align, sizeof(*io_u));
868 log_err("fio: unable to allocate aligned memory\n");
873 memset(io_u, 0, sizeof(*io_u));
874 INIT_FLIST_HEAD(&io_u->list);
875 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
877 if (!(td->io_ops->flags & FIO_NOIO)) {
878 io_u->buf = p + max_bs * i;
879 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
882 io_u_fill_buffer(td, io_u, max_bs);
883 if (td_write(td) && td->o.verify_pattern_bytes) {
885 * Fill the buffer with the pattern if we are
886 * going to be doing writes.
888 fill_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
893 io_u->flags = IO_U_F_FREE;
894 flist_add(&io_u->list, &td->io_u_freelist);
900 static int switch_ioscheduler(struct thread_data *td)
902 char tmp[256], tmp2[128];
906 if (td->io_ops->flags & FIO_DISKLESSIO)
909 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
911 f = fopen(tmp, "r+");
913 if (errno == ENOENT) {
914 log_err("fio: os or kernel doesn't support IO scheduler"
918 td_verror(td, errno, "fopen iosched");
925 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
926 if (ferror(f) || ret != 1) {
927 td_verror(td, errno, "fwrite");
935 * Read back and check that the selected scheduler is now the default.
937 ret = fread(tmp, 1, sizeof(tmp), f);
938 if (ferror(f) || ret < 0) {
939 td_verror(td, errno, "fread");
944 sprintf(tmp2, "[%s]", td->o.ioscheduler);
945 if (!strstr(tmp, tmp2)) {
946 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
947 td_verror(td, EINVAL, "iosched_switch");
956 static int keep_running(struct thread_data *td)
958 unsigned long long io_done;
962 if (td->o.time_based)
969 io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
971 if (io_done < td->o.size)
977 static void reset_io_counters(struct thread_data *td)
979 td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
980 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
982 td->rate_bytes[0] = td->rate_bytes[1] = 0;
983 td->rate_blocks[0] = td->rate_blocks[1] = 0;
985 td->last_was_sync = 0;
988 * reset file done count if we are to start over
990 if (td->o.time_based || td->o.loops)
991 td->nr_done_files = 0;
994 void reset_all_stats(struct thread_data *td)
999 reset_io_counters(td);
1001 for (i = 0; i < 2; i++) {
1002 td->io_bytes[i] = 0;
1003 td->io_blocks[i] = 0;
1004 td->io_issues[i] = 0;
1005 td->ts.total_io_u[i] = 0;
1008 fio_gettime(&tv, NULL);
1009 td->ts.runtime[0] = 0;
1010 td->ts.runtime[1] = 0;
1011 memcpy(&td->epoch, &tv, sizeof(tv));
1012 memcpy(&td->start, &tv, sizeof(tv));
1015 static void clear_io_state(struct thread_data *td)
1020 reset_io_counters(td);
1023 for_each_file(td, f, i)
1024 fio_file_clear_done(f);
1027 * Set the same seed to get repeatable runs
1029 td_fill_rand_seeds(td);
1032 static int exec_string(const char *string)
1034 int ret, newlen = strlen(string) + 1 + 8;
1037 str = malloc(newlen);
1038 sprintf(str, "sh -c %s", string);
1042 log_err("fio: exec of cmd <%s> failed\n", str);
1049 * Entry point for the thread based jobs. The process based jobs end up
1050 * here as well, after a little setup.
1052 static void *thread_main(void *data)
1054 unsigned long long elapsed;
1055 struct thread_data *td = data;
1056 pthread_condattr_t attr;
1059 if (!td->o.use_thread) {
1065 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1067 INIT_FLIST_HEAD(&td->io_u_freelist);
1068 INIT_FLIST_HEAD(&td->io_u_busylist);
1069 INIT_FLIST_HEAD(&td->io_u_requeues);
1070 INIT_FLIST_HEAD(&td->io_log_list);
1071 INIT_FLIST_HEAD(&td->io_hist_list);
1072 INIT_FLIST_HEAD(&td->verify_list);
1073 INIT_FLIST_HEAD(&td->trim_list);
1074 pthread_mutex_init(&td->io_u_lock, NULL);
1075 td->io_hist_tree = RB_ROOT;
1077 pthread_condattr_init(&attr);
1078 pthread_cond_init(&td->verify_cond, &attr);
1079 pthread_cond_init(&td->free_cond, &attr);
1081 td_set_runstate(td, TD_INITIALIZED);
1082 dprint(FD_MUTEX, "up startup_mutex\n");
1083 fio_mutex_up(startup_mutex);
1084 dprint(FD_MUTEX, "wait on td->mutex\n");
1085 fio_mutex_down(td->mutex);
1086 dprint(FD_MUTEX, "done waiting on td->mutex\n");
1089 * the ->mutex mutex is now no longer used, close it to avoid
1090 * eating a file descriptor
1092 fio_mutex_remove(td->mutex);
1095 * A new gid requires privilege, so we need to do this before setting
1098 if (td->o.gid != -1U && setgid(td->o.gid)) {
1099 td_verror(td, errno, "setgid");
1102 if (td->o.uid != -1U && setuid(td->o.uid)) {
1103 td_verror(td, errno, "setuid");
1108 * If we have a gettimeofday() thread, make sure we exclude that
1109 * thread from this job
1112 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1115 * Set affinity first, in case it has an impact on the memory
1118 if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1119 td_verror(td, errno, "cpu_set_affinity");
1124 * May alter parameters that init_io_u() will use, so we need to
1133 if (td->o.verify_async && verify_async_init(td))
1136 if (td->ioprio_set) {
1137 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1138 td_verror(td, errno, "ioprio_set");
1143 if (td->o.cgroup_weight && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1146 if (nice(td->o.nice) == -1) {
1147 td_verror(td, errno, "nice");
1151 if (td->o.ioscheduler && switch_ioscheduler(td))
1154 if (!td->o.create_serialize && setup_files(td))
1160 if (init_random_map(td))
1163 if (td->o.exec_prerun) {
1164 if (exec_string(td->o.exec_prerun))
1168 if (td->o.pre_read) {
1169 if (pre_read_files(td) < 0)
1173 fio_gettime(&td->epoch, NULL);
1174 getrusage(RUSAGE_SELF, &td->ts.ru_start);
1177 while (keep_running(td)) {
1178 fio_gettime(&td->start, NULL);
1179 memcpy(&td->ts.stat_sample_time[0], &td->start,
1181 memcpy(&td->ts.stat_sample_time[1], &td->start,
1183 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1185 if (td->o.ratemin[0] || td->o.ratemin[1])
1186 memcpy(&td->lastrate, &td->ts.stat_sample_time,
1187 sizeof(td->lastrate));
1192 prune_io_piece_log(td);
1198 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1199 elapsed = utime_since_now(&td->start);
1200 td->ts.runtime[DDIR_READ] += elapsed;
1202 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1203 elapsed = utime_since_now(&td->start);
1204 td->ts.runtime[DDIR_WRITE] += elapsed;
1207 if (td->error || td->terminate)
1210 if (!td->o.do_verify ||
1211 td->o.verify == VERIFY_NONE ||
1212 (td->io_ops->flags & FIO_UNIDIR))
1217 fio_gettime(&td->start, NULL);
1221 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1223 if (td->error || td->terminate)
1227 update_rusage_stat(td);
1228 td->ts.runtime[0] = (td->ts.runtime[0] + 999) / 1000;
1229 td->ts.runtime[1] = (td->ts.runtime[1] + 999) / 1000;
1230 td->ts.total_run_time = mtime_since_now(&td->epoch);
1231 td->ts.io_bytes[0] = td->io_bytes[0];
1232 td->ts.io_bytes[1] = td->io_bytes[1];
1234 fio_mutex_down(writeout_mutex);
1235 if (td->ts.bw_log) {
1236 if (td->o.bw_log_file) {
1237 finish_log_named(td, td->ts.bw_log,
1238 td->o.bw_log_file, "bw");
1240 finish_log(td, td->ts.bw_log, "bw");
1242 if (td->ts.lat_log) {
1243 if (td->o.lat_log_file) {
1244 finish_log_named(td, td->ts.lat_log,
1245 td->o.lat_log_file, "lat");
1247 finish_log(td, td->ts.lat_log, "lat");
1249 if (td->ts.slat_log) {
1250 if (td->o.lat_log_file) {
1251 finish_log_named(td, td->ts.slat_log,
1252 td->o.lat_log_file, "slat");
1254 finish_log(td, td->ts.slat_log, "slat");
1256 if (td->ts.clat_log) {
1257 if (td->o.lat_log_file) {
1258 finish_log_named(td, td->ts.clat_log,
1259 td->o.lat_log_file, "clat");
1261 finish_log(td, td->ts.clat_log, "clat");
1263 fio_mutex_up(writeout_mutex);
1264 if (td->o.exec_postrun)
1265 exec_string(td->o.exec_postrun);
1267 if (exitall_on_terminate)
1268 terminate_threads(td->groupid);
1272 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1275 if (td->o.verify_async)
1276 verify_async_exit(td);
1278 close_and_free_files(td);
1281 cgroup_shutdown(td, &cgroup_mnt);
1283 if (td->o.cpumask_set) {
1284 int ret = fio_cpuset_exit(&td->o.cpumask);
1286 td_verror(td, ret, "fio_cpuset_exit");
1290 * do this very late, it will log file closing as well
1292 if (td->o.write_iolog_file)
1293 write_iolog_close(td);
1295 options_mem_free(td);
1296 td_set_runstate(td, TD_EXITED);
1297 return (void *) (unsigned long) td->error;
1301 * We cannot pass the td data into a forked process, so attach the td and
1302 * pass it to the thread worker.
1304 static int fork_main(int shmid, int offset)
1306 struct thread_data *td;
1310 data = shmat(shmid, NULL, 0);
1311 if (data == (void *) -1) {
1319 * HP-UX inherits shm mappings?
1324 td = data + offset * sizeof(struct thread_data);
1325 ret = thread_main(td);
1327 return (int) (unsigned long) ret;
1331 * Run over the job map and reap the threads that have exited, if any.
1333 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1335 struct thread_data *td;
1336 int i, cputhreads, realthreads, pending, status, ret;
1339 * reap exited threads (TD_EXITED -> TD_REAPED)
1341 realthreads = pending = cputhreads = 0;
1342 for_each_td(td, i) {
1346 * ->io_ops is NULL for a thread that has closed its
1349 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1358 if (td->runstate == TD_REAPED)
1360 if (td->o.use_thread) {
1361 if (td->runstate == TD_EXITED) {
1362 td_set_runstate(td, TD_REAPED);
1369 if (td->runstate == TD_EXITED)
1373 * check if someone quit or got killed in an unusual way
1375 ret = waitpid(td->pid, &status, flags);
1377 if (errno == ECHILD) {
1378 log_err("fio: pid=%d disappeared %d\n",
1379 (int) td->pid, td->runstate);
1380 td_set_runstate(td, TD_REAPED);
1384 } else if (ret == td->pid) {
1385 if (WIFSIGNALED(status)) {
1386 int sig = WTERMSIG(status);
1389 log_err("fio: pid=%d, got signal=%d\n",
1390 (int) td->pid, sig);
1391 td_set_runstate(td, TD_REAPED);
1394 if (WIFEXITED(status)) {
1395 if (WEXITSTATUS(status) && !td->error)
1396 td->error = WEXITSTATUS(status);
1398 td_set_runstate(td, TD_REAPED);
1404 * thread is not dead, continue
1410 (*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
1411 (*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
1418 done_secs += mtime_since_now(&td->epoch) / 1000;
1421 if (*nr_running == cputhreads && !pending && realthreads)
1422 terminate_threads(TERMINATE_ALL);
1425 static void *gtod_thread_main(void *data)
1427 fio_mutex_up(startup_mutex);
1430 * As long as we have jobs around, update the clock. It would be nice
1431 * to have some way of NOT hammering that CPU with gettimeofday(),
1432 * but I'm not sure what to use outside of a simple CPU nop to relax
1433 * it - we don't want to lose precision.
1443 static int fio_start_gtod_thread(void)
1445 pthread_attr_t attr;
1448 pthread_attr_init(&attr);
1449 pthread_attr_setstacksize(&attr, PTHREAD_STACK_MIN);
1450 ret = pthread_create(>od_thread, &attr, gtod_thread_main, NULL);
1451 pthread_attr_destroy(&attr);
1453 log_err("Can't create gtod thread: %s\n", strerror(ret));
1457 ret = pthread_detach(gtod_thread);
1459 log_err("Can't detatch gtod thread: %s\n", strerror(ret));
1463 dprint(FD_MUTEX, "wait on startup_mutex\n");
1464 fio_mutex_down(startup_mutex);
1465 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1470 * Main function for kicking off and reaping jobs, as needed.
1472 static void run_threads(void)
1474 struct thread_data *td;
1475 unsigned long spent;
1476 int i, todo, nr_running, m_rate, t_rate, nr_started;
1478 if (fio_pin_memory())
1481 if (fio_gtod_offload && fio_start_gtod_thread())
1484 if (!terse_output) {
1485 log_info("Starting ");
1487 log_info("%d thread%s", nr_thread,
1488 nr_thread > 1 ? "s" : "");
1492 log_info("%d process%s", nr_process,
1493 nr_process > 1 ? "es" : "");
1501 todo = thread_number;
1504 m_rate = t_rate = 0;
1506 for_each_td(td, i) {
1507 print_status_init(td->thread_number - 1);
1509 if (!td->o.create_serialize)
1513 * do file setup here so it happens sequentially,
1514 * we don't want X number of threads getting their
1515 * client data interspersed on disk
1517 if (setup_files(td)) {
1520 log_err("fio: pid=%d, err=%d/%s\n",
1521 (int) td->pid, td->error, td->verror);
1522 td_set_runstate(td, TD_REAPED);
1529 * for sharing to work, each job must always open
1530 * its own files. so close them, if we opened them
1533 for_each_file(td, f, j) {
1534 if (fio_file_open(f))
1535 td_io_close_file(td, f);
1543 struct thread_data *map[REAL_MAX_JOBS];
1544 struct timeval this_start;
1545 int this_jobs = 0, left;
1548 * create threads (TD_NOT_CREATED -> TD_CREATED)
1550 for_each_td(td, i) {
1551 if (td->runstate != TD_NOT_CREATED)
1555 * never got a chance to start, killed by other
1556 * thread for some reason
1558 if (td->terminate) {
1563 if (td->o.start_delay) {
1564 spent = mtime_since_genesis();
1566 if (td->o.start_delay * 1000 > spent)
1570 if (td->o.stonewall && (nr_started || nr_running)) {
1571 dprint(FD_PROCESS, "%s: stonewall wait\n",
1579 * Set state to created. Thread will transition
1580 * to TD_INITIALIZED when it's done setting up.
1582 td_set_runstate(td, TD_CREATED);
1583 map[this_jobs++] = td;
1586 if (td->o.use_thread) {
1589 dprint(FD_PROCESS, "will pthread_create\n");
1590 ret = pthread_create(&td->thread, NULL,
1593 log_err("pthread_create: %s\n",
1598 ret = pthread_detach(td->thread);
1600 log_err("pthread_detach: %s",
1604 dprint(FD_PROCESS, "will fork\n");
1607 int ret = fork_main(shm_id, i);
1610 } else if (i == fio_debug_jobno)
1611 *fio_debug_jobp = pid;
1613 dprint(FD_MUTEX, "wait on startup_mutex\n");
1614 if (fio_mutex_down_timeout(startup_mutex, 10)) {
1615 log_err("fio: job startup hung? exiting.\n");
1616 terminate_threads(TERMINATE_ALL);
1621 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1625 * Wait for the started threads to transition to
1628 fio_gettime(&this_start, NULL);
1630 while (left && !fio_abort) {
1631 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1636 for (i = 0; i < this_jobs; i++) {
1640 if (td->runstate == TD_INITIALIZED) {
1643 } else if (td->runstate >= TD_EXITED) {
1647 nr_running++; /* work-around... */
1653 log_err("fio: %d jobs failed to start\n", left);
1654 for (i = 0; i < this_jobs; i++) {
1658 kill(td->pid, SIGTERM);
1664 * start created threads (TD_INITIALIZED -> TD_RUNNING).
1666 for_each_td(td, i) {
1667 if (td->runstate != TD_INITIALIZED)
1670 if (in_ramp_time(td))
1671 td_set_runstate(td, TD_RAMP);
1673 td_set_runstate(td, TD_RUNNING);
1676 m_rate += td->o.ratemin[0] + td->o.ratemin[1];
1677 t_rate += td->o.rate[0] + td->o.rate[1];
1679 fio_mutex_up(td->mutex);
1682 reap_threads(&nr_running, &t_rate, &m_rate);
1688 while (nr_running) {
1689 reap_threads(&nr_running, &t_rate, &m_rate);
1699 if (exec_profile && load_profile(exec_profile))
1706 setup_log(&agg_io_log[DDIR_READ]);
1707 setup_log(&agg_io_log[DDIR_WRITE]);
1710 startup_mutex = fio_mutex_init(0);
1711 if (startup_mutex == NULL)
1713 writeout_mutex = fio_mutex_init(1);
1714 if (writeout_mutex == NULL)
1718 create_disk_util_thread();
1720 cgroup_list = smalloc(sizeof(*cgroup_list));
1721 INIT_FLIST_HEAD(cgroup_list);
1728 __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1729 __finish_log(agg_io_log[DDIR_WRITE],
1730 "agg-write_bw.log");
1734 cgroup_kill(cgroup_list);
1738 fio_mutex_remove(startup_mutex);
1739 fio_mutex_remove(writeout_mutex);
1743 void reset_fio_state(void)
1752 int main(int argc, char *argv[], char *envp[])
1761 * We need locale for number printing, if it isn't set then just
1762 * go with the US format.
1764 if (!getenv("LC_NUMERIC"))
1765 setlocale(LC_NUMERIC, "en_US");
1767 ps = sysconf(_SC_PAGESIZE);
1769 log_err("Failed to get page size\n");
1776 fio_keywords_init();
1778 if (parse_options(argc, argv))