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");
346 io_u->xfer_buflen = io_u->resid;
347 io_u->xfer_buf += bytes;
348 io_u->offset += bytes;
350 if (io_u->offset == io_u->file->real_file_size)
353 requeue_io_u(td, &io_u);
356 ret = io_u_sync_complete(td, io_u);
364 requeue_io_u(td, &io_u);
365 ret2 = td_io_commit(td);
371 td_verror(td, -ret, "td_io_queue");
375 if (ret < 0 || td->error)
379 * if we can queue more, do so. but check if there are
380 * completed io_u's first.
383 if (queue_full(td) || ret == FIO_Q_BUSY) {
386 if (td->cur_depth > td->o.iodepth_low)
387 min_events = td->cur_depth - td->o.iodepth_low;
391 * Reap required number of io units, if any, and do the
392 * verification on them through the callback handler
394 if (io_u_queued_complete(td, min_events) < 0)
399 min_events = td->cur_depth;
402 ret = io_u_queued_complete(td, min_events);
404 cleanup_pending_aio(td);
406 td_set_runstate(td, TD_RUNNING);
410 * Main IO worker function. It retrieves io_u's to process and queues
411 * and reaps them, checking for rate and errors along the way.
413 static void do_io(struct thread_data *td)
420 td_set_runstate(td, TD_RUNNING);
422 while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) {
423 struct timeval comp_time;
436 memcpy(&s, &io_u->start_time, sizeof(s));
438 if (runtime_exceeded(td, &s)) {
443 ret = td_io_queue(td, io_u);
445 case FIO_Q_COMPLETED:
448 else if (io_u->resid) {
449 int bytes = io_u->xfer_buflen - io_u->resid;
455 td_verror(td, ENODATA, "full resid");
460 io_u->xfer_buflen = io_u->resid;
461 io_u->xfer_buf += bytes;
462 io_u->offset += bytes;
464 if (io_u->offset == io_u->file->real_file_size)
467 requeue_io_u(td, &io_u);
470 fio_gettime(&comp_time, NULL);
471 bytes_done = io_u_sync_complete(td, io_u);
478 * if the engine doesn't have a commit hook,
479 * the io_u is really queued. if it does have such
480 * a hook, it has to call io_u_queued() itself.
482 if (td->io_ops->commit == NULL)
483 io_u_queued(td, io_u);
486 requeue_io_u(td, &io_u);
487 ret2 = td_io_commit(td);
497 if (ret < 0 || td->error)
501 * See if we need to complete some commands
503 if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
505 if (queue_full(td) || ret == FIO_Q_BUSY) {
508 if (td->cur_depth > td->o.iodepth_low)
509 min_evts = td->cur_depth - td->o.iodepth_low;
512 fio_gettime(&comp_time, NULL);
513 bytes_done = io_u_queued_complete(td, min_evts);
522 * the rate is batched for now, it should work for batches
523 * of completions except the very first one which may look
526 usec = utime_since(&s, &comp_time);
528 rate_throttle(td, usec, bytes_done);
530 if (check_min_rate(td, &comp_time)) {
531 if (exitall_on_terminate)
532 terminate_threads(td->groupid);
533 td_verror(td, ENODATA, "check_min_rate");
537 if (td->o.thinktime) {
538 unsigned long long b;
540 b = td->io_blocks[0] + td->io_blocks[1];
541 if (!(b % td->o.thinktime_blocks)) {
544 if (td->o.thinktime_spin)
545 __usec_sleep(td->o.thinktime_spin);
547 left = td->o.thinktime - td->o.thinktime_spin;
549 usec_sleep(td, left);
559 ret = io_u_queued_complete(td, i);
561 if (should_fsync(td) && td->o.end_fsync) {
562 td_set_runstate(td, TD_FSYNCING);
564 for_each_file(td, f, i) {
565 if (!(f->flags & FIO_FILE_OPEN))
571 cleanup_pending_aio(td);
574 static void cleanup_io_u(struct thread_data *td)
576 struct list_head *entry, *n;
579 list_for_each_safe(entry, n, &td->io_u_freelist) {
580 io_u = list_entry(entry, struct io_u, list);
582 list_del(&io_u->list);
590 * "randomly" fill the buffer contents
592 static void fill_rand_buf(struct io_u *io_u, int max_bs)
594 int *ptr = io_u->buf;
596 while ((void *) ptr - io_u->buf < max_bs) {
597 *ptr = rand() * 0x9e370001;
602 static int init_io_u(struct thread_data *td)
604 unsigned long long buf_size;
610 if (td->io_ops->flags & FIO_SYNCIO)
613 max_units = td->o.iodepth;
615 max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
616 buf_size = (unsigned long long) max_bs * (unsigned long long) max_units;
617 buf_size += page_mask;
618 if (buf_size != (size_t) buf_size) {
619 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
623 td->orig_buffer_size = buf_size;
625 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE)
626 td->orig_buffer_size = (td->orig_buffer_size + td->o.hugepage_size - 1) & ~(td->o.hugepage_size - 1);
627 else if (td->orig_buffer_size & page_mask)
628 td->orig_buffer_size = (td->orig_buffer_size + page_mask) & ~page_mask;
630 if (allocate_io_mem(td))
633 p = ALIGN(td->orig_buffer);
634 for (i = 0; i < max_units; i++) {
635 io_u = malloc(sizeof(*io_u));
636 memset(io_u, 0, sizeof(*io_u));
637 INIT_LIST_HEAD(&io_u->list);
639 io_u->buf = p + max_bs * i;
640 if (td_write(td) || td_rw(td))
641 fill_rand_buf(io_u, max_bs);
644 io_u->flags = IO_U_F_FREE;
645 list_add(&io_u->list, &td->io_u_freelist);
653 static int switch_ioscheduler(struct thread_data *td)
655 char tmp[256], tmp2[128];
659 if (td->io_ops->flags & FIO_DISKLESSIO)
662 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
664 f = fopen(tmp, "r+");
666 td_verror(td, errno, "fopen");
673 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
674 if (ferror(f) || ret != 1) {
675 td_verror(td, errno, "fwrite");
683 * Read back and check that the selected scheduler is now the default.
685 ret = fread(tmp, 1, sizeof(tmp), f);
686 if (ferror(f) || ret < 0) {
687 td_verror(td, errno, "fread");
692 sprintf(tmp2, "[%s]", td->o.ioscheduler);
693 if (!strstr(tmp, tmp2)) {
694 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
695 td_verror(td, EINVAL, "iosched_switch");
704 static int clear_io_state(struct thread_data *td)
710 td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
711 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
715 td->rw_end_set[0] = td->rw_end_set[1] = 0;
717 td->last_was_sync = 0;
719 for_each_file(td, f, i)
720 td_io_close_file(td, f);
723 for_each_file(td, f, i) {
724 ret = td_io_open_file(td, f);
733 * Entry point for the thread based jobs. The process based jobs end up
734 * here as well, after a little setup.
736 static void *thread_main(void *data)
738 unsigned long long runtime[2];
739 struct thread_data *td = data;
740 unsigned long elapsed;
743 if (!td->o.use_thread)
748 INIT_LIST_HEAD(&td->io_u_freelist);
749 INIT_LIST_HEAD(&td->io_u_busylist);
750 INIT_LIST_HEAD(&td->io_u_requeues);
751 INIT_LIST_HEAD(&td->io_log_list);
752 INIT_LIST_HEAD(&td->io_hist_list);
753 td->io_hist_tree = RB_ROOT;
758 if (fio_setaffinity(td) == -1) {
759 td_verror(td, errno, "cpu_set_affinity");
767 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
768 td_verror(td, errno, "ioprio_set");
773 if (nice(td->o.nice) == -1) {
774 td_verror(td, errno, "nice");
778 if (td->o.ioscheduler && switch_ioscheduler(td))
781 td_set_runstate(td, TD_INITIALIZED);
782 fio_sem_up(startup_sem);
783 fio_sem_down(td->mutex);
786 * the ->mutex semaphore is now no longer used, close it to avoid
787 * eating a file descriptor
789 fio_sem_remove(td->mutex);
791 if (!td->o.create_serialize && setup_files(td))
800 if (init_random_map(td))
803 if (td->o.exec_prerun) {
804 if (system(td->o.exec_prerun) < 0)
808 fio_gettime(&td->epoch, NULL);
809 memcpy(&td->timeout_end, &td->epoch, sizeof(td->epoch));
810 getrusage(RUSAGE_SELF, &td->ts.ru_start);
812 runtime[0] = runtime[1] = 0;
814 while (td->o.loops--) {
815 fio_gettime(&td->start, NULL);
816 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
819 memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
821 if (clear_state && clear_io_state(td))
824 prune_io_piece_log(td);
830 if (td_read(td) && td->io_bytes[DDIR_READ]) {
831 if (td->rw_end_set[DDIR_READ])
832 elapsed = utime_since(&td->start, &td->rw_end[DDIR_READ]);
834 elapsed = utime_since_now(&td->start);
836 runtime[DDIR_READ] += elapsed;
838 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
839 if (td->rw_end_set[DDIR_WRITE])
840 elapsed = utime_since(&td->start, &td->rw_end[DDIR_WRITE]);
842 elapsed = utime_since_now(&td->start);
844 runtime[DDIR_WRITE] += elapsed;
847 if (td->error || td->terminate)
850 if (td->o.verify == VERIFY_NONE)
853 if (clear_io_state(td))
856 fio_gettime(&td->start, NULL);
860 runtime[DDIR_READ] += utime_since_now(&td->start);
862 if (td->error || td->terminate)
866 update_rusage_stat(td);
867 td->ts.runtime[0] = runtime[0] / 1000;
868 td->ts.runtime[1] = runtime[1] / 1000;
869 td->ts.total_run_time = mtime_since_now(&td->epoch);
870 td->ts.io_bytes[0] = td->io_bytes[0];
871 td->ts.io_bytes[1] = td->io_bytes[1];
874 finish_log(td, td->ts.bw_log, "bw");
876 finish_log(td, td->ts.slat_log, "slat");
878 finish_log(td, td->ts.clat_log, "clat");
879 if (td->o.write_iolog_file)
880 write_iolog_close(td);
881 if (td->o.exec_postrun) {
882 if (system(td->o.exec_postrun) < 0)
883 log_err("fio: postrun %s failed\n", td->o.exec_postrun);
886 if (exitall_on_terminate)
887 terminate_threads(td->groupid);
891 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
895 options_mem_free(td);
896 td_set_runstate(td, TD_EXITED);
897 return (void *) (unsigned long) td->error;
899 fio_sem_up(startup_sem);
904 * We cannot pass the td data into a forked process, so attach the td and
905 * pass it to the thread worker.
907 static int fork_main(int shmid, int offset)
909 struct thread_data *td;
912 data = shmat(shmid, NULL, 0);
913 if (data == (void *) -1) {
920 td = data + offset * sizeof(struct thread_data);
921 ret = thread_main(td);
923 return (int) (unsigned long) ret;
927 * Run over the job map and reap the threads that have exited, if any.
929 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
931 struct thread_data *td;
932 int i, cputhreads, pending, status, ret;
935 * reap exited threads (TD_EXITED -> TD_REAPED)
937 pending = cputhreads = 0;
942 * ->io_ops is NULL for a thread that has closed its
945 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
948 if (!td->pid || td->runstate == TD_REAPED)
950 if (td->o.use_thread) {
951 if (td->runstate == TD_EXITED) {
952 td_set_runstate(td, TD_REAPED);
959 if (td->runstate == TD_EXITED)
963 * check if someone quit or got killed in an unusual way
965 ret = waitpid(td->pid, &status, flags);
967 if (errno == ECHILD) {
968 log_err("fio: pid=%d disappeared %d\n", td->pid, td->runstate);
969 td_set_runstate(td, TD_REAPED);
973 } else if (ret == td->pid) {
974 if (WIFSIGNALED(status)) {
975 int sig = WTERMSIG(status);
978 log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
979 td_set_runstate(td, TD_REAPED);
982 if (WIFEXITED(status)) {
983 if (WEXITSTATUS(status) && !td->error)
984 td->error = WEXITSTATUS(status);
986 td_set_runstate(td, TD_REAPED);
992 * thread is not dead, continue
996 if (td->o.use_thread) {
999 if (pthread_join(td->thread, (void *) &ret))
1000 perror("pthread_join");
1004 (*m_rate) -= td->o.ratemin;
1005 (*t_rate) -= td->o.rate;
1011 if (*nr_running == cputhreads && !pending)
1012 terminate_threads(TERMINATE_ALL);
1016 * Main function for kicking off and reaping jobs, as needed.
1018 static void run_threads(void)
1020 struct thread_data *td;
1021 unsigned long spent;
1022 int i, todo, nr_running, m_rate, t_rate, nr_started;
1024 if (fio_pin_memory())
1027 if (!terse_output) {
1028 printf("Starting ");
1030 printf("%d thread%s", nr_thread, nr_thread > 1 ? "s" : "");
1034 printf("%d process%s", nr_process, nr_process > 1 ? "es" : "");
1040 signal(SIGINT, sig_handler);
1041 signal(SIGALRM, sig_handler);
1043 todo = thread_number;
1046 m_rate = t_rate = 0;
1048 for_each_td(td, i) {
1049 print_status_init(td->thread_number - 1);
1051 if (!td->o.create_serialize) {
1057 * do file setup here so it happens sequentially,
1058 * we don't want X number of threads getting their
1059 * client data interspersed on disk
1061 if (setup_files(td)) {
1064 log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
1065 td_set_runstate(td, TD_REAPED);
1075 struct thread_data *map[MAX_JOBS];
1076 struct timeval this_start;
1077 int this_jobs = 0, left;
1080 * create threads (TD_NOT_CREATED -> TD_CREATED)
1082 for_each_td(td, i) {
1083 if (td->runstate != TD_NOT_CREATED)
1087 * never got a chance to start, killed by other
1088 * thread for some reason
1090 if (td->terminate) {
1095 if (td->o.start_delay) {
1096 spent = mtime_since_genesis();
1098 if (td->o.start_delay * 1000 > spent)
1102 if (td->o.stonewall && (nr_started || nr_running))
1106 * Set state to created. Thread will transition
1107 * to TD_INITIALIZED when it's done setting up.
1109 td_set_runstate(td, TD_CREATED);
1110 map[this_jobs++] = td;
1113 if (td->o.use_thread) {
1114 if (pthread_create(&td->thread, NULL, thread_main, td)) {
1115 perror("thread_create");
1121 int ret = fork_main(shm_id, i);
1126 fio_sem_down(startup_sem);
1130 * Wait for the started threads to transition to
1133 fio_gettime(&this_start, NULL);
1135 while (left && !fio_abort) {
1136 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1141 for (i = 0; i < this_jobs; i++) {
1145 if (td->runstate == TD_INITIALIZED) {
1148 } else if (td->runstate >= TD_EXITED) {
1152 nr_running++; /* work-around... */
1158 log_err("fio: %d jobs failed to start\n", left);
1159 for (i = 0; i < this_jobs; i++) {
1163 kill(td->pid, SIGTERM);
1169 * start created threads (TD_INITIALIZED -> TD_RUNNING).
1171 for_each_td(td, i) {
1172 if (td->runstate != TD_INITIALIZED)
1175 td_set_runstate(td, TD_RUNNING);
1178 m_rate += td->o.ratemin;
1179 t_rate += td->o.rate;
1181 fio_sem_up(td->mutex);
1184 reap_threads(&nr_running, &t_rate, &m_rate);
1190 while (nr_running) {
1191 reap_threads(&nr_running, &t_rate, &m_rate);
1199 int main(int argc, char *argv[])
1204 * We need locale for number printing, if it isn't set then just
1205 * go with the US format.
1207 if (!getenv("LC_NUMERIC"))
1208 setlocale(LC_NUMERIC, "en_US");
1210 if (parse_options(argc, argv))
1213 if (!thread_number) {
1214 log_err("Nothing to do\n");
1218 ps = sysconf(_SC_PAGESIZE);
1220 log_err("Failed to get page size\n");
1228 setup_log(&agg_io_log[DDIR_READ]);
1229 setup_log(&agg_io_log[DDIR_WRITE]);
1232 startup_sem = fio_sem_init(0);
1236 disk_util_timer_arm();
1243 __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1244 __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");
1248 fio_sem_remove(startup_sem);