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
40 unsigned long page_mask;
41 unsigned long page_size;
43 (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
46 int thread_number = 0;
52 static struct fio_sem *startup_sem;
53 static volatile int fio_abort;
54 static int exit_value;
56 struct io_log *agg_io_log[2];
58 #define TERMINATE_ALL (-1)
59 #define JOB_START_TIMEOUT (5 * 1000)
61 static inline void td_set_runstate(struct thread_data *td, int runstate)
63 td->runstate = runstate;
66 static void terminate_threads(int group_id)
68 struct thread_data *td;
72 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
74 * if the thread is running, just let it exit
76 if (td->runstate < TD_RUNNING)
77 kill(td->pid, SIGQUIT);
79 td->o.start_delay = 0;
84 static void sig_handler(int sig)
89 disk_util_timer_arm();
90 print_thread_status();
93 printf("\nfio: terminating on signal %d\n", sig);
95 terminate_threads(TERMINATE_ALL);
101 * Check if we are above the minimum rate given.
103 static int check_min_rate(struct thread_data *td, struct timeval *now)
105 unsigned long long bytes = 0;
106 unsigned long iops = 0;
111 * No minimum rate set, always ok
113 if (!td->o.ratemin && !td->o.rate_iops_min)
117 * allow a 2 second settle period in the beginning
119 if (mtime_since(&td->start, now) < 2000)
123 iops += td->io_blocks[DDIR_READ];
124 bytes += td->this_io_bytes[DDIR_READ];
127 iops += td->io_blocks[DDIR_WRITE];
128 bytes += td->this_io_bytes[DDIR_WRITE];
132 * if rate blocks is set, sample is running
134 if (td->rate_bytes || td->rate_blocks) {
135 spent = mtime_since(&td->lastrate, now);
136 if (spent < td->o.ratecycle)
141 * check bandwidth specified rate
143 if (bytes < td->rate_bytes) {
144 log_err("%s: min rate %u not met\n", td->o.name, td->o.ratemin);
147 rate = (bytes - td->rate_bytes) / spent;
148 if (rate < td->o.ratemin || bytes < td->rate_bytes) {
149 log_err("%s: min rate %u not met, got %luKiB/sec\n", td->o.name, td->o.ratemin, rate);
155 * checks iops specified rate
157 if (iops < td->o.rate_iops) {
158 log_err("%s: min iops rate %u not met\n", td->o.name, td->o.rate_iops);
161 rate = (iops - td->rate_blocks) / spent;
162 if (rate < td->o.rate_iops_min || iops < td->rate_blocks) {
163 log_err("%s: min iops rate %u not met, got %lu\n", td->o.name, td->o.rate_iops_min, rate);
169 td->rate_bytes = bytes;
170 td->rate_blocks = iops;
171 memcpy(&td->lastrate, now, sizeof(*now));
175 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
179 if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
186 * When job exits, we can cancel the in-flight IO if we are using async
187 * io. Attempt to do so.
189 static void cleanup_pending_aio(struct thread_data *td)
191 struct list_head *entry, *n;
196 * get immediately available events, if any
198 r = io_u_queued_complete(td, 0);
203 * now cancel remaining active events
205 if (td->io_ops->cancel) {
206 list_for_each_safe(entry, n, &td->io_u_busylist) {
207 io_u = list_entry(entry, struct io_u, list);
210 * if the io_u isn't in flight, then that generally
211 * means someone leaked an io_u. complain but fix
212 * it up, so we don't stall here.
214 if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
215 log_err("fio: non-busy IO on busy list\n");
218 r = td->io_ops->cancel(td, io_u);
226 r = io_u_queued_complete(td, td->cur_depth);
230 * Helper to handle the final sync of a file. Works just like the normal
231 * io path, just does everything sync.
233 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
235 struct io_u *io_u = __get_io_u(td);
241 io_u->ddir = DDIR_SYNC;
244 if (td_io_prep(td, io_u)) {
250 ret = td_io_queue(td, io_u);
252 td_verror(td, io_u->error, "td_io_queue");
255 } else if (ret == FIO_Q_QUEUED) {
256 if (io_u_queued_complete(td, 1) < 0)
258 } else if (ret == FIO_Q_COMPLETED) {
260 td_verror(td, io_u->error, "td_io_queue");
264 if (io_u_sync_complete(td, io_u) < 0)
266 } else if (ret == FIO_Q_BUSY) {
267 if (td_io_commit(td))
276 * The main verify engine. Runs over the writes we previously submitted,
277 * reads the blocks back in, and checks the crc/md5 of the data.
279 static void do_verify(struct thread_data *td)
287 * sync io first and invalidate cache, to make sure we really
290 for_each_file(td, f, i) {
291 if (!(f->flags & FIO_FILE_OPEN))
293 if (fio_io_sync(td, f))
295 if (file_invalidate_cache(td, f))
302 td_set_runstate(td, TD_VERIFYING);
305 while (!td->terminate) {
308 io_u = __get_io_u(td);
312 if (runtime_exceeded(td, &io_u->start_time)) {
317 if (get_next_verify(td, io_u)) {
322 if (td_io_prep(td, io_u)) {
327 io_u->end_io = verify_io_u;
329 ret = td_io_queue(td, io_u);
331 case FIO_Q_COMPLETED:
334 else if (io_u->resid) {
335 int bytes = io_u->xfer_buflen - io_u->resid;
341 td_verror(td, ENODATA, "full resid");
345 io_u->xfer_buflen = io_u->resid;
346 io_u->xfer_buf += bytes;
347 requeue_io_u(td, &io_u);
349 ret = io_u_sync_complete(td, io_u);
357 requeue_io_u(td, &io_u);
358 ret2 = td_io_commit(td);
364 td_verror(td, -ret, "td_io_queue");
368 if (ret < 0 || td->error)
372 * if we can queue more, do so. but check if there are
373 * completed io_u's first.
376 if (queue_full(td) || ret == FIO_Q_BUSY) {
379 if (td->cur_depth > td->o.iodepth_low)
380 min_events = td->cur_depth - td->o.iodepth_low;
384 * Reap required number of io units, if any, and do the
385 * verification on them through the callback handler
387 if (io_u_queued_complete(td, min_events) < 0)
392 min_events = td->cur_depth;
395 ret = io_u_queued_complete(td, min_events);
397 cleanup_pending_aio(td);
399 td_set_runstate(td, TD_RUNNING);
403 * Main IO worker function. It retrieves io_u's to process and queues
404 * and reaps them, checking for rate and errors along the way.
406 static void do_io(struct thread_data *td)
413 td_set_runstate(td, TD_RUNNING);
415 while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->io_size) {
416 struct timeval comp_time;
429 memcpy(&s, &io_u->start_time, sizeof(s));
431 if (runtime_exceeded(td, &s)) {
436 ret = td_io_queue(td, io_u);
438 case FIO_Q_COMPLETED:
441 else if (io_u->resid) {
442 int bytes = io_u->xfer_buflen - io_u->resid;
448 td_verror(td, ENODATA, "full resid");
453 io_u->xfer_buflen = io_u->resid;
454 io_u->xfer_buf += bytes;
455 requeue_io_u(td, &io_u);
457 fio_gettime(&comp_time, NULL);
458 bytes_done = io_u_sync_complete(td, io_u);
465 * if the engine doesn't have a commit hook,
466 * the io_u is really queued. if it does have such
467 * a hook, it has to call io_u_queued() itself.
469 if (td->io_ops->commit == NULL)
470 io_u_queued(td, io_u);
473 requeue_io_u(td, &io_u);
474 ret2 = td_io_commit(td);
484 if (ret < 0 || td->error)
488 * See if we need to complete some commands
490 if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
492 if (queue_full(td) || ret == FIO_Q_BUSY) {
495 if (td->cur_depth > td->o.iodepth_low)
496 min_evts = td->cur_depth - td->o.iodepth_low;
499 fio_gettime(&comp_time, NULL);
500 bytes_done = io_u_queued_complete(td, min_evts);
509 * the rate is batched for now, it should work for batches
510 * of completions except the very first one which may look
513 usec = utime_since(&s, &comp_time);
515 rate_throttle(td, usec, bytes_done);
517 if (check_min_rate(td, &comp_time)) {
518 if (exitall_on_terminate)
519 terminate_threads(td->groupid);
520 td_verror(td, ENODATA, "check_min_rate");
524 if (td->o.thinktime) {
525 unsigned long long b;
527 b = td->io_blocks[0] + td->io_blocks[1];
528 if (!(b % td->o.thinktime_blocks)) {
531 if (td->o.thinktime_spin)
532 __usec_sleep(td->o.thinktime_spin);
534 left = td->o.thinktime - td->o.thinktime_spin;
536 usec_sleep(td, left);
546 ret = io_u_queued_complete(td, i);
548 if (should_fsync(td) && td->o.end_fsync) {
549 td_set_runstate(td, TD_FSYNCING);
551 for_each_file(td, f, i) {
552 if (!(f->flags & FIO_FILE_OPEN))
558 cleanup_pending_aio(td);
561 static void cleanup_io_u(struct thread_data *td)
563 struct list_head *entry, *n;
566 list_for_each_safe(entry, n, &td->io_u_freelist) {
567 io_u = list_entry(entry, struct io_u, list);
569 list_del(&io_u->list);
577 * "randomly" fill the buffer contents
579 static void fill_rand_buf(struct io_u *io_u, int max_bs)
581 int *ptr = io_u->buf;
583 while ((void *) ptr - io_u->buf < max_bs) {
584 *ptr = rand() * 0x9e370001;
589 static int init_io_u(struct thread_data *td)
591 unsigned long long buf_size;
597 if (td->io_ops->flags & FIO_SYNCIO)
600 max_units = td->o.iodepth;
602 max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
603 buf_size = (unsigned long long) max_bs * (unsigned long long) max_units;
604 buf_size += page_mask;
605 if (buf_size != (size_t) buf_size) {
606 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
610 td->orig_buffer_size = buf_size;
612 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE)
613 td->orig_buffer_size = (td->orig_buffer_size + td->o.hugepage_size - 1) & ~(td->o.hugepage_size - 1);
614 else if (td->orig_buffer_size & page_mask)
615 td->orig_buffer_size = (td->orig_buffer_size + page_mask) & ~page_mask;
617 if (allocate_io_mem(td))
620 p = ALIGN(td->orig_buffer);
621 for (i = 0; i < max_units; i++) {
622 io_u = malloc(sizeof(*io_u));
623 memset(io_u, 0, sizeof(*io_u));
624 INIT_LIST_HEAD(&io_u->list);
626 io_u->buf = p + max_bs * i;
627 if (td_write(td) || td_rw(td))
628 fill_rand_buf(io_u, max_bs);
631 io_u->flags = IO_U_F_FREE;
632 list_add(&io_u->list, &td->io_u_freelist);
640 static int switch_ioscheduler(struct thread_data *td)
642 char tmp[256], tmp2[128];
646 if (td->io_ops->flags & FIO_DISKLESSIO)
649 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
651 f = fopen(tmp, "r+");
653 td_verror(td, errno, "fopen");
660 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
661 if (ferror(f) || ret != 1) {
662 td_verror(td, errno, "fwrite");
670 * Read back and check that the selected scheduler is now the default.
672 ret = fread(tmp, 1, sizeof(tmp), f);
673 if (ferror(f) || ret < 0) {
674 td_verror(td, errno, "fread");
679 sprintf(tmp2, "[%s]", td->o.ioscheduler);
680 if (!strstr(tmp, tmp2)) {
681 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
682 td_verror(td, EINVAL, "iosched_switch");
691 static int clear_io_state(struct thread_data *td)
697 td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
698 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
702 td->rw_end_set[0] = td->rw_end_set[1] = 0;
704 td->last_was_sync = 0;
706 for_each_file(td, f, i)
707 td_io_close_file(td, f);
710 for_each_file(td, f, i) {
711 ret = td_io_open_file(td, f);
720 * Entry point for the thread based jobs. The process based jobs end up
721 * here as well, after a little setup.
723 static void *thread_main(void *data)
725 unsigned long long runtime[2];
726 struct thread_data *td = data;
727 unsigned long elapsed;
730 if (!td->o.use_thread)
735 INIT_LIST_HEAD(&td->io_u_freelist);
736 INIT_LIST_HEAD(&td->io_u_busylist);
737 INIT_LIST_HEAD(&td->io_u_requeues);
738 INIT_LIST_HEAD(&td->io_log_list);
743 if (fio_setaffinity(td) == -1) {
744 td_verror(td, errno, "cpu_set_affinity");
752 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
753 td_verror(td, errno, "ioprio_set");
758 if (nice(td->o.nice) == -1) {
759 td_verror(td, errno, "nice");
763 if (td->o.ioscheduler && switch_ioscheduler(td))
766 td_set_runstate(td, TD_INITIALIZED);
767 fio_sem_up(startup_sem);
768 fio_sem_down(td->mutex);
771 * the ->mutex semaphore is now no longer used, close it to avoid
772 * eating a file descriptor
774 fio_sem_remove(td->mutex);
776 if (!td->o.create_serialize && setup_files(td))
785 if (init_random_map(td))
788 if (td->o.exec_prerun) {
789 if (system(td->o.exec_prerun) < 0)
793 fio_gettime(&td->epoch, NULL);
794 memcpy(&td->timeout_end, &td->epoch, sizeof(td->epoch));
795 getrusage(RUSAGE_SELF, &td->ts.ru_start);
797 runtime[0] = runtime[1] = 0;
799 while (td->o.loops--) {
800 fio_gettime(&td->start, NULL);
801 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
804 memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
806 if (clear_state && clear_io_state(td))
809 prune_io_piece_log(td);
815 if (td_read(td) && td->io_bytes[DDIR_READ]) {
816 if (td->rw_end_set[DDIR_READ])
817 elapsed = utime_since(&td->start, &td->rw_end[DDIR_READ]);
819 elapsed = utime_since_now(&td->start);
821 runtime[DDIR_READ] += elapsed;
823 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
824 if (td->rw_end_set[DDIR_WRITE])
825 elapsed = utime_since(&td->start, &td->rw_end[DDIR_WRITE]);
827 elapsed = utime_since_now(&td->start);
829 runtime[DDIR_WRITE] += elapsed;
832 if (td->error || td->terminate)
835 if (td->o.verify == VERIFY_NONE)
838 if (clear_io_state(td))
841 fio_gettime(&td->start, NULL);
845 runtime[DDIR_READ] += utime_since_now(&td->start);
847 if (td->error || td->terminate)
851 update_rusage_stat(td);
852 td->ts.runtime[0] = runtime[0] / 1000;
853 td->ts.runtime[1] = runtime[1] / 1000;
854 td->ts.total_run_time = mtime_since_now(&td->epoch);
855 td->ts.io_bytes[0] = td->io_bytes[0];
856 td->ts.io_bytes[1] = td->io_bytes[1];
859 finish_log(td, td->ts.bw_log, "bw");
861 finish_log(td, td->ts.slat_log, "slat");
863 finish_log(td, td->ts.clat_log, "clat");
864 if (td->o.write_iolog_file)
865 write_iolog_close(td);
866 if (td->o.exec_postrun) {
867 if (system(td->o.exec_postrun) < 0)
868 log_err("fio: postrun %s failed\n", td->o.exec_postrun);
871 if (exitall_on_terminate)
872 terminate_threads(td->groupid);
876 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
880 options_mem_free(td);
881 td_set_runstate(td, TD_EXITED);
882 return (void *) (unsigned long) td->error;
884 fio_sem_up(startup_sem);
889 * We cannot pass the td data into a forked process, so attach the td and
890 * pass it to the thread worker.
892 static int fork_main(int shmid, int offset)
894 struct thread_data *td;
897 data = shmat(shmid, NULL, 0);
898 if (data == (void *) -1) {
905 td = data + offset * sizeof(struct thread_data);
906 ret = thread_main(td);
908 return (int) (unsigned long) ret;
912 * Run over the job map and reap the threads that have exited, if any.
914 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
916 struct thread_data *td;
917 int i, cputhreads, pending, status, ret;
920 * reap exited threads (TD_EXITED -> TD_REAPED)
922 pending = cputhreads = 0;
927 * ->io_ops is NULL for a thread that has closed its
930 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
933 if (!td->pid || td->runstate == TD_REAPED)
935 if (td->o.use_thread) {
936 if (td->runstate == TD_EXITED) {
937 td_set_runstate(td, TD_REAPED);
944 if (td->runstate == TD_EXITED)
948 * check if someone quit or got killed in an unusual way
950 ret = waitpid(td->pid, &status, flags);
952 if (errno == ECHILD) {
953 log_err("fio: pid=%d disappeared %d\n", td->pid, td->runstate);
954 td_set_runstate(td, TD_REAPED);
958 } else if (ret == td->pid) {
959 if (WIFSIGNALED(status)) {
960 int sig = WTERMSIG(status);
963 log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
964 td_set_runstate(td, TD_REAPED);
967 if (WIFEXITED(status)) {
968 if (WEXITSTATUS(status) && !td->error)
969 td->error = WEXITSTATUS(status);
971 td_set_runstate(td, TD_REAPED);
977 * thread is not dead, continue
981 if (td->o.use_thread) {
984 if (pthread_join(td->thread, (void *) &ret))
985 perror("pthread_join");
989 (*m_rate) -= td->o.ratemin;
990 (*t_rate) -= td->o.rate;
996 if (*nr_running == cputhreads && !pending)
997 terminate_threads(TERMINATE_ALL);
1001 * Main function for kicking off and reaping jobs, as needed.
1003 static void run_threads(void)
1005 struct thread_data *td;
1006 unsigned long spent;
1007 int i, todo, nr_running, m_rate, t_rate, nr_started;
1009 if (fio_pin_memory())
1012 if (!terse_output) {
1013 printf("Starting ");
1015 printf("%d thread%s", nr_thread, nr_thread > 1 ? "s" : "");
1019 printf("%d process%s", nr_process, nr_process > 1 ? "es" : "");
1025 signal(SIGINT, sig_handler);
1026 signal(SIGALRM, sig_handler);
1028 todo = thread_number;
1031 m_rate = t_rate = 0;
1033 for_each_td(td, i) {
1034 print_status_init(td->thread_number - 1);
1036 if (!td->o.create_serialize) {
1042 * do file setup here so it happens sequentially,
1043 * we don't want X number of threads getting their
1044 * client data interspersed on disk
1046 if (setup_files(td)) {
1049 log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
1050 td_set_runstate(td, TD_REAPED);
1060 struct thread_data *map[MAX_JOBS];
1061 struct timeval this_start;
1062 int this_jobs = 0, left;
1065 * create threads (TD_NOT_CREATED -> TD_CREATED)
1067 for_each_td(td, i) {
1068 if (td->runstate != TD_NOT_CREATED)
1072 * never got a chance to start, killed by other
1073 * thread for some reason
1075 if (td->terminate) {
1080 if (td->o.start_delay) {
1081 spent = mtime_since_genesis();
1083 if (td->o.start_delay * 1000 > spent)
1087 if (td->o.stonewall && (nr_started || nr_running))
1091 * Set state to created. Thread will transition
1092 * to TD_INITIALIZED when it's done setting up.
1094 td_set_runstate(td, TD_CREATED);
1095 map[this_jobs++] = td;
1098 if (td->o.use_thread) {
1099 if (pthread_create(&td->thread, NULL, thread_main, td)) {
1100 perror("thread_create");
1106 int ret = fork_main(shm_id, i);
1111 fio_sem_down(startup_sem);
1115 * Wait for the started threads to transition to
1118 fio_gettime(&this_start, NULL);
1120 while (left && !fio_abort) {
1121 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1126 for (i = 0; i < this_jobs; i++) {
1130 if (td->runstate == TD_INITIALIZED) {
1133 } else if (td->runstate >= TD_EXITED) {
1137 nr_running++; /* work-around... */
1143 log_err("fio: %d jobs failed to start\n", left);
1144 for (i = 0; i < this_jobs; i++) {
1148 kill(td->pid, SIGTERM);
1154 * start created threads (TD_INITIALIZED -> TD_RUNNING).
1156 for_each_td(td, i) {
1157 if (td->runstate != TD_INITIALIZED)
1160 td_set_runstate(td, TD_RUNNING);
1163 m_rate += td->o.ratemin;
1164 t_rate += td->o.rate;
1166 fio_sem_up(td->mutex);
1169 reap_threads(&nr_running, &t_rate, &m_rate);
1175 while (nr_running) {
1176 reap_threads(&nr_running, &t_rate, &m_rate);
1184 int main(int argc, char *argv[])
1189 * We need locale for number printing, if it isn't set then just
1190 * go with the US format.
1192 if (!getenv("LC_NUMERIC"))
1193 setlocale(LC_NUMERIC, "en_US");
1195 if (parse_options(argc, argv))
1198 if (!thread_number) {
1199 log_err("Nothing to do\n");
1203 ps = sysconf(_SC_PAGESIZE);
1205 log_err("Failed to get page size\n");
1213 setup_log(&agg_io_log[DDIR_READ]);
1214 setup_log(&agg_io_log[DDIR_WRITE]);
1217 startup_sem = fio_sem_init(0);
1221 disk_util_timer_arm();
1228 __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1229 __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");
1233 fio_sem_remove(startup_sem);