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
39 unsigned long page_mask;
40 unsigned long page_size;
42 (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
45 int thread_number = 0;
51 static struct fio_sem *startup_sem;
52 static volatile int fio_abort;
53 static int exit_value;
55 struct io_log *agg_io_log[2];
57 #define TERMINATE_ALL (-1)
58 #define JOB_START_TIMEOUT (5 * 1000)
60 static inline void td_set_runstate(struct thread_data *td, int runstate)
62 td->runstate = runstate;
65 static void terminate_threads(int group_id)
67 struct thread_data *td;
71 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
73 * if the thread is running, just let it exit
75 if (td->runstate < TD_RUNNING)
76 kill(td->pid, SIGQUIT);
78 td->o.start_delay = 0;
83 static void sig_handler(int sig)
88 disk_util_timer_arm();
89 print_thread_status();
92 printf("\nfio: terminating on signal %d\n", sig);
94 terminate_threads(TERMINATE_ALL);
100 * Check if we are above the minimum rate given.
102 static int check_min_rate(struct thread_data *td, struct timeval *now)
104 unsigned long long bytes = 0;
105 unsigned long iops = 0;
110 * No minimum rate set, always ok
112 if (!td->o.ratemin && !td->o.rate_iops_min)
116 * allow a 2 second settle period in the beginning
118 if (mtime_since(&td->start, now) < 2000)
122 iops += td->io_blocks[DDIR_READ];
123 bytes += td->this_io_bytes[DDIR_READ];
126 iops += td->io_blocks[DDIR_WRITE];
127 bytes += td->this_io_bytes[DDIR_WRITE];
131 * if rate blocks is set, sample is running
133 if (td->rate_bytes || td->rate_blocks) {
134 spent = mtime_since(&td->lastrate, now);
135 if (spent < td->o.ratecycle)
140 * check bandwidth specified rate
142 if (bytes < td->rate_bytes) {
143 log_err("%s: min rate %u not met\n", td->o.name, td->o.ratemin);
146 rate = (bytes - td->rate_bytes) / spent;
147 if (rate < td->o.ratemin || bytes < td->rate_bytes) {
148 log_err("%s: min rate %u not met, got %luKiB/sec\n", td->o.name, td->o.ratemin, rate);
154 * checks iops specified rate
156 if (iops < td->o.rate_iops) {
157 log_err("%s: min iops rate %u not met\n", td->o.name, td->o.rate_iops);
160 rate = (iops - td->rate_blocks) / spent;
161 if (rate < td->o.rate_iops_min || iops < td->rate_blocks) {
162 log_err("%s: min iops rate %u not met, got %lu\n", td->o.name, td->o.rate_iops_min, rate);
168 td->rate_bytes = bytes;
169 td->rate_blocks = iops;
170 memcpy(&td->lastrate, now, sizeof(*now));
174 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
178 if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
185 * When job exits, we can cancel the in-flight IO if we are using async
186 * io. Attempt to do so.
188 static void cleanup_pending_aio(struct thread_data *td)
190 struct list_head *entry, *n;
195 * get immediately available events, if any
197 r = io_u_queued_complete(td, 0);
202 * now cancel remaining active events
204 if (td->io_ops->cancel) {
205 list_for_each_safe(entry, n, &td->io_u_busylist) {
206 io_u = list_entry(entry, struct io_u, list);
209 * if the io_u isn't in flight, then that generally
210 * means someone leaked an io_u. complain but fix
211 * it up, so we don't stall here.
213 if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
214 log_err("fio: non-busy IO on busy list\n");
217 r = td->io_ops->cancel(td, io_u);
225 r = io_u_queued_complete(td, td->cur_depth);
229 * Helper to handle the final sync of a file. Works just like the normal
230 * io path, just does everything sync.
232 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
234 struct io_u *io_u = __get_io_u(td);
240 io_u->ddir = DDIR_SYNC;
243 if (td_io_prep(td, io_u)) {
249 ret = td_io_queue(td, io_u);
251 td_verror(td, io_u->error, "td_io_queue");
254 } else if (ret == FIO_Q_QUEUED) {
255 if (io_u_queued_complete(td, 1) < 0)
257 } else if (ret == FIO_Q_COMPLETED) {
259 td_verror(td, io_u->error, "td_io_queue");
263 if (io_u_sync_complete(td, io_u) < 0)
265 } else if (ret == FIO_Q_BUSY) {
266 if (td_io_commit(td))
275 * The main verify engine. Runs over the writes we previously submitted,
276 * reads the blocks back in, and checks the crc/md5 of the data.
278 static void do_verify(struct thread_data *td)
286 * sync io first and invalidate cache, to make sure we really
289 for_each_file(td, f, i) {
290 if (!(f->flags & FIO_FILE_OPEN))
292 if (fio_io_sync(td, f))
294 if (file_invalidate_cache(td, f))
301 td_set_runstate(td, TD_VERIFYING);
304 while (!td->terminate) {
307 io_u = __get_io_u(td);
311 if (runtime_exceeded(td, &io_u->start_time)) {
316 if (get_next_verify(td, io_u)) {
321 if (td_io_prep(td, io_u)) {
326 io_u->end_io = verify_io_u;
328 ret = td_io_queue(td, io_u);
330 case FIO_Q_COMPLETED:
333 else if (io_u->resid) {
334 int bytes = io_u->xfer_buflen - io_u->resid;
335 struct fio_file *f = io_u->file;
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;
349 f->last_completed_pos = io_u->offset;
351 td->ts.short_io_u[io_u->ddir]++;
353 if (io_u->offset == f->real_file_size)
356 requeue_io_u(td, &io_u);
359 ret = io_u_sync_complete(td, io_u);
367 requeue_io_u(td, &io_u);
368 ret2 = td_io_commit(td);
374 td_verror(td, -ret, "td_io_queue");
378 if (ret < 0 || td->error)
382 * if we can queue more, do so. but check if there are
383 * completed io_u's first.
386 if (queue_full(td) || ret == FIO_Q_BUSY) {
389 if (td->cur_depth > td->o.iodepth_low)
390 min_events = td->cur_depth - td->o.iodepth_low;
394 * Reap required number of io units, if any, and do the
395 * verification on them through the callback handler
397 if (io_u_queued_complete(td, min_events) < 0)
402 min_events = td->cur_depth;
405 ret = io_u_queued_complete(td, min_events);
407 cleanup_pending_aio(td);
409 td_set_runstate(td, TD_RUNNING);
413 * Main IO worker function. It retrieves io_u's to process and queues
414 * and reaps them, checking for rate and errors along the way.
416 static void do_io(struct thread_data *td)
423 td_set_runstate(td, TD_RUNNING);
425 while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) {
426 struct timeval comp_time;
439 memcpy(&s, &io_u->start_time, sizeof(s));
441 if (runtime_exceeded(td, &s)) {
446 ret = td_io_queue(td, io_u);
448 case FIO_Q_COMPLETED:
451 else if (io_u->resid) {
452 int bytes = io_u->xfer_buflen - io_u->resid;
453 struct fio_file *f = io_u->file;
459 td_verror(td, ENODATA, "full resid");
464 io_u->xfer_buflen = io_u->resid;
465 io_u->xfer_buf += bytes;
466 io_u->offset += bytes;
467 f->last_completed_pos = io_u->offset;
469 td->ts.short_io_u[io_u->ddir]++;
471 if (io_u->offset == f->real_file_size)
474 requeue_io_u(td, &io_u);
477 fio_gettime(&comp_time, NULL);
478 bytes_done = io_u_sync_complete(td, io_u);
485 * if the engine doesn't have a commit hook,
486 * the io_u is really queued. if it does have such
487 * a hook, it has to call io_u_queued() itself.
489 if (td->io_ops->commit == NULL)
490 io_u_queued(td, io_u);
493 requeue_io_u(td, &io_u);
494 ret2 = td_io_commit(td);
504 if (ret < 0 || td->error)
508 * See if we need to complete some commands
510 if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
512 if (queue_full(td) || ret == FIO_Q_BUSY) {
515 if (td->cur_depth > td->o.iodepth_low)
516 min_evts = td->cur_depth - td->o.iodepth_low;
519 fio_gettime(&comp_time, NULL);
520 bytes_done = io_u_queued_complete(td, min_evts);
529 * the rate is batched for now, it should work for batches
530 * of completions except the very first one which may look
533 usec = utime_since(&s, &comp_time);
535 rate_throttle(td, usec, bytes_done);
537 if (check_min_rate(td, &comp_time)) {
538 if (exitall_on_terminate)
539 terminate_threads(td->groupid);
540 td_verror(td, ENODATA, "check_min_rate");
544 if (td->o.thinktime) {
545 unsigned long long b;
547 b = td->io_blocks[0] + td->io_blocks[1];
548 if (!(b % td->o.thinktime_blocks)) {
551 if (td->o.thinktime_spin)
552 __usec_sleep(td->o.thinktime_spin);
554 left = td->o.thinktime - td->o.thinktime_spin;
556 usec_sleep(td, left);
566 ret = io_u_queued_complete(td, i);
568 if (should_fsync(td) && td->o.end_fsync) {
569 td_set_runstate(td, TD_FSYNCING);
571 for_each_file(td, f, i) {
572 if (!(f->flags & FIO_FILE_OPEN))
578 cleanup_pending_aio(td);
581 static void cleanup_io_u(struct thread_data *td)
583 struct list_head *entry, *n;
586 list_for_each_safe(entry, n, &td->io_u_freelist) {
587 io_u = list_entry(entry, struct io_u, list);
589 list_del(&io_u->list);
597 * "randomly" fill the buffer contents
599 static void fill_rand_buf(struct io_u *io_u, int max_bs)
601 int *ptr = io_u->buf;
603 while ((void *) ptr - io_u->buf < max_bs) {
604 *ptr = rand() * 0x9e370001;
609 static int init_io_u(struct thread_data *td)
611 unsigned long long buf_size;
617 if (td->io_ops->flags & FIO_SYNCIO)
620 max_units = td->o.iodepth;
622 max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
623 buf_size = (unsigned long long) max_bs * (unsigned long long) max_units;
624 buf_size += page_mask;
625 if (buf_size != (size_t) buf_size) {
626 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
630 td->orig_buffer_size = buf_size;
632 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE)
633 td->orig_buffer_size = (td->orig_buffer_size + td->o.hugepage_size - 1) & ~(td->o.hugepage_size - 1);
634 else if (td->orig_buffer_size & page_mask)
635 td->orig_buffer_size = (td->orig_buffer_size + page_mask) & ~page_mask;
637 if (allocate_io_mem(td))
640 p = ALIGN(td->orig_buffer);
641 for (i = 0; i < max_units; i++) {
642 io_u = malloc(sizeof(*io_u));
643 memset(io_u, 0, sizeof(*io_u));
644 INIT_LIST_HEAD(&io_u->list);
646 io_u->buf = p + max_bs * i;
647 if (td_write(td) || td_rw(td))
648 fill_rand_buf(io_u, max_bs);
651 io_u->flags = IO_U_F_FREE;
652 list_add(&io_u->list, &td->io_u_freelist);
660 static int switch_ioscheduler(struct thread_data *td)
662 char tmp[256], tmp2[128];
666 if (td->io_ops->flags & FIO_DISKLESSIO)
669 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
671 f = fopen(tmp, "r+");
673 if (errno == ENOENT) {
674 log_err("fio: os or kernel doesn't support IO scheduler switching\n");
677 td_verror(td, errno, "fopen iosched");
684 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
685 if (ferror(f) || ret != 1) {
686 td_verror(td, errno, "fwrite");
694 * Read back and check that the selected scheduler is now the default.
696 ret = fread(tmp, 1, sizeof(tmp), f);
697 if (ferror(f) || ret < 0) {
698 td_verror(td, errno, "fread");
703 sprintf(tmp2, "[%s]", td->o.ioscheduler);
704 if (!strstr(tmp, tmp2)) {
705 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
706 td_verror(td, EINVAL, "iosched_switch");
715 static int clear_io_state(struct thread_data *td)
721 td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
722 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
726 td->rw_end_set[0] = td->rw_end_set[1] = 0;
728 td->last_was_sync = 0;
730 for_each_file(td, f, i)
731 td_io_close_file(td, f);
734 for_each_file(td, f, i) {
735 ret = td_io_open_file(td, f);
744 * Entry point for the thread based jobs. The process based jobs end up
745 * here as well, after a little setup.
747 static void *thread_main(void *data)
749 unsigned long long runtime[2];
750 struct thread_data *td = data;
751 unsigned long elapsed;
754 if (!td->o.use_thread)
759 INIT_LIST_HEAD(&td->io_u_freelist);
760 INIT_LIST_HEAD(&td->io_u_busylist);
761 INIT_LIST_HEAD(&td->io_u_requeues);
762 INIT_LIST_HEAD(&td->io_log_list);
763 INIT_LIST_HEAD(&td->io_hist_list);
764 td->io_hist_tree = RB_ROOT;
769 if (fio_setaffinity(td) == -1) {
770 td_verror(td, errno, "cpu_set_affinity");
778 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
779 td_verror(td, errno, "ioprio_set");
784 if (nice(td->o.nice) == -1) {
785 td_verror(td, errno, "nice");
789 if (td->o.ioscheduler && switch_ioscheduler(td))
792 td_set_runstate(td, TD_INITIALIZED);
793 fio_sem_up(startup_sem);
794 fio_sem_down(td->mutex);
797 * the ->mutex semaphore is now no longer used, close it to avoid
798 * eating a file descriptor
800 fio_sem_remove(td->mutex);
802 if (!td->o.create_serialize && setup_files(td))
811 if (init_random_map(td))
814 if (td->o.exec_prerun) {
815 if (system(td->o.exec_prerun) < 0)
819 fio_gettime(&td->epoch, NULL);
820 memcpy(&td->timeout_end, &td->epoch, sizeof(td->epoch));
821 getrusage(RUSAGE_SELF, &td->ts.ru_start);
823 runtime[0] = runtime[1] = 0;
825 while (td->o.loops--) {
826 fio_gettime(&td->start, NULL);
827 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
830 memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
832 if (clear_state && clear_io_state(td))
835 prune_io_piece_log(td);
841 if (td_read(td) && td->io_bytes[DDIR_READ]) {
842 if (td->rw_end_set[DDIR_READ])
843 elapsed = utime_since(&td->start, &td->rw_end[DDIR_READ]);
845 elapsed = utime_since_now(&td->start);
847 runtime[DDIR_READ] += elapsed;
849 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
850 if (td->rw_end_set[DDIR_WRITE])
851 elapsed = utime_since(&td->start, &td->rw_end[DDIR_WRITE]);
853 elapsed = utime_since_now(&td->start);
855 runtime[DDIR_WRITE] += elapsed;
858 if (td->error || td->terminate)
861 if (td->o.verify == VERIFY_NONE)
864 if (clear_io_state(td))
867 fio_gettime(&td->start, NULL);
871 runtime[DDIR_READ] += utime_since_now(&td->start);
873 if (td->error || td->terminate)
877 update_rusage_stat(td);
878 td->ts.runtime[0] = runtime[0] / 1000;
879 td->ts.runtime[1] = runtime[1] / 1000;
880 td->ts.total_run_time = mtime_since_now(&td->epoch);
881 td->ts.io_bytes[0] = td->io_bytes[0];
882 td->ts.io_bytes[1] = td->io_bytes[1];
885 finish_log(td, td->ts.bw_log, "bw");
887 finish_log(td, td->ts.slat_log, "slat");
889 finish_log(td, td->ts.clat_log, "clat");
890 if (td->o.write_iolog_file)
891 write_iolog_close(td);
892 if (td->o.exec_postrun) {
893 if (system(td->o.exec_postrun) < 0)
894 log_err("fio: postrun %s failed\n", td->o.exec_postrun);
897 if (exitall_on_terminate)
898 terminate_threads(td->groupid);
902 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
906 options_mem_free(td);
907 td_set_runstate(td, TD_EXITED);
908 return (void *) (unsigned long) td->error;
910 fio_sem_up(startup_sem);
915 * We cannot pass the td data into a forked process, so attach the td and
916 * pass it to the thread worker.
918 static int fork_main(int shmid, int offset)
920 struct thread_data *td;
923 data = shmat(shmid, NULL, 0);
924 if (data == (void *) -1) {
931 td = data + offset * sizeof(struct thread_data);
932 ret = thread_main(td);
934 return (int) (unsigned long) ret;
938 * Run over the job map and reap the threads that have exited, if any.
940 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
942 struct thread_data *td;
943 int i, cputhreads, pending, status, ret;
946 * reap exited threads (TD_EXITED -> TD_REAPED)
948 pending = cputhreads = 0;
953 * ->io_ops is NULL for a thread that has closed its
956 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
959 if (!td->pid || td->runstate == TD_REAPED)
961 if (td->o.use_thread) {
962 if (td->runstate == TD_EXITED) {
963 td_set_runstate(td, TD_REAPED);
970 if (td->runstate == TD_EXITED)
974 * check if someone quit or got killed in an unusual way
976 ret = waitpid(td->pid, &status, flags);
978 if (errno == ECHILD) {
979 log_err("fio: pid=%d disappeared %d\n", td->pid, td->runstate);
980 td_set_runstate(td, TD_REAPED);
984 } else if (ret == td->pid) {
985 if (WIFSIGNALED(status)) {
986 int sig = WTERMSIG(status);
989 log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
990 td_set_runstate(td, TD_REAPED);
993 if (WIFEXITED(status)) {
994 if (WEXITSTATUS(status) && !td->error)
995 td->error = WEXITSTATUS(status);
997 td_set_runstate(td, TD_REAPED);
1003 * thread is not dead, continue
1008 if (td->o.use_thread) {
1011 if (pthread_join(td->thread, (void *) &ret))
1012 perror("pthread_join");
1016 (*m_rate) -= td->o.ratemin;
1017 (*t_rate) -= td->o.rate;
1024 if (*nr_running == cputhreads && !pending)
1025 terminate_threads(TERMINATE_ALL);
1029 * Main function for kicking off and reaping jobs, as needed.
1031 static void run_threads(void)
1033 struct thread_data *td;
1034 unsigned long spent;
1035 int i, todo, nr_running, m_rate, t_rate, nr_started;
1037 if (fio_pin_memory())
1040 if (!terse_output) {
1041 printf("Starting ");
1043 printf("%d thread%s", nr_thread, nr_thread > 1 ? "s" : "");
1047 printf("%d process%s", nr_process, nr_process > 1 ? "es" : "");
1053 signal(SIGINT, sig_handler);
1054 signal(SIGALRM, sig_handler);
1056 todo = thread_number;
1059 m_rate = t_rate = 0;
1061 for_each_td(td, i) {
1062 print_status_init(td->thread_number - 1);
1064 if (!td->o.create_serialize) {
1070 * do file setup here so it happens sequentially,
1071 * we don't want X number of threads getting their
1072 * client data interspersed on disk
1074 if (setup_files(td)) {
1077 log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
1078 td_set_runstate(td, TD_REAPED);
1088 struct thread_data *map[MAX_JOBS];
1089 struct timeval this_start;
1090 int this_jobs = 0, left;
1093 * create threads (TD_NOT_CREATED -> TD_CREATED)
1095 for_each_td(td, i) {
1096 if (td->runstate != TD_NOT_CREATED)
1100 * never got a chance to start, killed by other
1101 * thread for some reason
1103 if (td->terminate) {
1108 if (td->o.start_delay) {
1109 spent = mtime_since_genesis();
1111 if (td->o.start_delay * 1000 > spent)
1115 if (td->o.stonewall && (nr_started || nr_running))
1119 * Set state to created. Thread will transition
1120 * to TD_INITIALIZED when it's done setting up.
1122 td_set_runstate(td, TD_CREATED);
1123 map[this_jobs++] = td;
1126 if (td->o.use_thread) {
1127 if (pthread_create(&td->thread, NULL, thread_main, td)) {
1128 perror("thread_create");
1134 int ret = fork_main(shm_id, i);
1139 fio_sem_down(startup_sem);
1143 * Wait for the started threads to transition to
1146 fio_gettime(&this_start, NULL);
1148 while (left && !fio_abort) {
1149 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1154 for (i = 0; i < this_jobs; i++) {
1158 if (td->runstate == TD_INITIALIZED) {
1161 } else if (td->runstate >= TD_EXITED) {
1165 nr_running++; /* work-around... */
1171 log_err("fio: %d jobs failed to start\n", left);
1172 for (i = 0; i < this_jobs; i++) {
1176 kill(td->pid, SIGTERM);
1182 * start created threads (TD_INITIALIZED -> TD_RUNNING).
1184 for_each_td(td, i) {
1185 if (td->runstate != TD_INITIALIZED)
1188 td_set_runstate(td, TD_RUNNING);
1191 m_rate += td->o.ratemin;
1192 t_rate += td->o.rate;
1194 fio_sem_up(td->mutex);
1197 reap_threads(&nr_running, &t_rate, &m_rate);
1203 while (nr_running) {
1204 reap_threads(&nr_running, &t_rate, &m_rate);
1212 int main(int argc, char *argv[])
1217 * We need locale for number printing, if it isn't set then just
1218 * go with the US format.
1220 if (!getenv("LC_NUMERIC"))
1221 setlocale(LC_NUMERIC, "en_US");
1223 if (parse_options(argc, argv))
1229 ps = sysconf(_SC_PAGESIZE);
1231 log_err("Failed to get page size\n");
1239 setup_log(&agg_io_log[DDIR_READ]);
1240 setup_log(&agg_io_log[DDIR_WRITE]);
1243 startup_sem = fio_sem_init(0);
1247 disk_util_timer_arm();
1254 __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1255 __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");
1259 fio_sem_remove(startup_sem);