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 if (io_u->offset == f->real_file_size)
354 requeue_io_u(td, &io_u);
357 ret = io_u_sync_complete(td, io_u);
365 requeue_io_u(td, &io_u);
366 ret2 = td_io_commit(td);
372 td_verror(td, -ret, "td_io_queue");
376 if (ret < 0 || td->error)
380 * if we can queue more, do so. but check if there are
381 * completed io_u's first.
384 if (queue_full(td) || ret == FIO_Q_BUSY) {
387 if (td->cur_depth > td->o.iodepth_low)
388 min_events = td->cur_depth - td->o.iodepth_low;
392 * Reap required number of io units, if any, and do the
393 * verification on them through the callback handler
395 if (io_u_queued_complete(td, min_events) < 0)
400 min_events = td->cur_depth;
403 ret = io_u_queued_complete(td, min_events);
405 cleanup_pending_aio(td);
407 td_set_runstate(td, TD_RUNNING);
411 * Main IO worker function. It retrieves io_u's to process and queues
412 * and reaps them, checking for rate and errors along the way.
414 static void do_io(struct thread_data *td)
421 td_set_runstate(td, TD_RUNNING);
423 while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) {
424 struct timeval comp_time;
437 memcpy(&s, &io_u->start_time, sizeof(s));
439 if (runtime_exceeded(td, &s)) {
444 ret = td_io_queue(td, io_u);
446 case FIO_Q_COMPLETED:
449 else if (io_u->resid) {
450 int bytes = io_u->xfer_buflen - io_u->resid;
451 struct fio_file *f = io_u->file;
457 td_verror(td, ENODATA, "full resid");
462 io_u->xfer_buflen = io_u->resid;
463 io_u->xfer_buf += bytes;
464 io_u->offset += bytes;
465 f->last_completed_pos = io_u->offset;
467 if (io_u->offset == f->real_file_size)
470 requeue_io_u(td, &io_u);
473 fio_gettime(&comp_time, NULL);
474 bytes_done = io_u_sync_complete(td, io_u);
481 * if the engine doesn't have a commit hook,
482 * the io_u is really queued. if it does have such
483 * a hook, it has to call io_u_queued() itself.
485 if (td->io_ops->commit == NULL)
486 io_u_queued(td, io_u);
489 requeue_io_u(td, &io_u);
490 ret2 = td_io_commit(td);
500 if (ret < 0 || td->error)
504 * See if we need to complete some commands
506 if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
508 if (queue_full(td) || ret == FIO_Q_BUSY) {
511 if (td->cur_depth > td->o.iodepth_low)
512 min_evts = td->cur_depth - td->o.iodepth_low;
515 fio_gettime(&comp_time, NULL);
516 bytes_done = io_u_queued_complete(td, min_evts);
525 * the rate is batched for now, it should work for batches
526 * of completions except the very first one which may look
529 usec = utime_since(&s, &comp_time);
531 rate_throttle(td, usec, bytes_done);
533 if (check_min_rate(td, &comp_time)) {
534 if (exitall_on_terminate)
535 terminate_threads(td->groupid);
536 td_verror(td, ENODATA, "check_min_rate");
540 if (td->o.thinktime) {
541 unsigned long long b;
543 b = td->io_blocks[0] + td->io_blocks[1];
544 if (!(b % td->o.thinktime_blocks)) {
547 if (td->o.thinktime_spin)
548 __usec_sleep(td->o.thinktime_spin);
550 left = td->o.thinktime - td->o.thinktime_spin;
552 usec_sleep(td, left);
562 ret = io_u_queued_complete(td, i);
564 if (should_fsync(td) && td->o.end_fsync) {
565 td_set_runstate(td, TD_FSYNCING);
567 for_each_file(td, f, i) {
568 if (!(f->flags & FIO_FILE_OPEN))
574 cleanup_pending_aio(td);
577 static void cleanup_io_u(struct thread_data *td)
579 struct list_head *entry, *n;
582 list_for_each_safe(entry, n, &td->io_u_freelist) {
583 io_u = list_entry(entry, struct io_u, list);
585 list_del(&io_u->list);
593 * "randomly" fill the buffer contents
595 static void fill_rand_buf(struct io_u *io_u, int max_bs)
597 int *ptr = io_u->buf;
599 while ((void *) ptr - io_u->buf < max_bs) {
600 *ptr = rand() * 0x9e370001;
605 static int init_io_u(struct thread_data *td)
607 unsigned long long buf_size;
613 if (td->io_ops->flags & FIO_SYNCIO)
616 max_units = td->o.iodepth;
618 max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
619 buf_size = (unsigned long long) max_bs * (unsigned long long) max_units;
620 buf_size += page_mask;
621 if (buf_size != (size_t) buf_size) {
622 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
626 td->orig_buffer_size = buf_size;
628 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE)
629 td->orig_buffer_size = (td->orig_buffer_size + td->o.hugepage_size - 1) & ~(td->o.hugepage_size - 1);
630 else if (td->orig_buffer_size & page_mask)
631 td->orig_buffer_size = (td->orig_buffer_size + page_mask) & ~page_mask;
633 if (allocate_io_mem(td))
636 p = ALIGN(td->orig_buffer);
637 for (i = 0; i < max_units; i++) {
638 io_u = malloc(sizeof(*io_u));
639 memset(io_u, 0, sizeof(*io_u));
640 INIT_LIST_HEAD(&io_u->list);
642 io_u->buf = p + max_bs * i;
643 if (td_write(td) || td_rw(td))
644 fill_rand_buf(io_u, max_bs);
647 io_u->flags = IO_U_F_FREE;
648 list_add(&io_u->list, &td->io_u_freelist);
656 static int switch_ioscheduler(struct thread_data *td)
658 char tmp[256], tmp2[128];
662 if (td->io_ops->flags & FIO_DISKLESSIO)
665 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
667 f = fopen(tmp, "r+");
669 if (errno == ENOENT) {
670 log_err("fio: os or kernel doesn't support IO scheduler switching\n");
673 td_verror(td, errno, "fopen iosched");
680 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
681 if (ferror(f) || ret != 1) {
682 td_verror(td, errno, "fwrite");
690 * Read back and check that the selected scheduler is now the default.
692 ret = fread(tmp, 1, sizeof(tmp), f);
693 if (ferror(f) || ret < 0) {
694 td_verror(td, errno, "fread");
699 sprintf(tmp2, "[%s]", td->o.ioscheduler);
700 if (!strstr(tmp, tmp2)) {
701 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
702 td_verror(td, EINVAL, "iosched_switch");
711 static int clear_io_state(struct thread_data *td)
717 td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
718 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
722 td->rw_end_set[0] = td->rw_end_set[1] = 0;
724 td->last_was_sync = 0;
726 for_each_file(td, f, i)
727 td_io_close_file(td, f);
730 for_each_file(td, f, i) {
731 ret = td_io_open_file(td, f);
740 * Entry point for the thread based jobs. The process based jobs end up
741 * here as well, after a little setup.
743 static void *thread_main(void *data)
745 unsigned long long runtime[2];
746 struct thread_data *td = data;
747 unsigned long elapsed;
750 if (!td->o.use_thread)
755 INIT_LIST_HEAD(&td->io_u_freelist);
756 INIT_LIST_HEAD(&td->io_u_busylist);
757 INIT_LIST_HEAD(&td->io_u_requeues);
758 INIT_LIST_HEAD(&td->io_log_list);
759 INIT_LIST_HEAD(&td->io_hist_list);
760 td->io_hist_tree = RB_ROOT;
765 if (fio_setaffinity(td) == -1) {
766 td_verror(td, errno, "cpu_set_affinity");
774 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
775 td_verror(td, errno, "ioprio_set");
780 if (nice(td->o.nice) == -1) {
781 td_verror(td, errno, "nice");
785 if (td->o.ioscheduler && switch_ioscheduler(td))
788 td_set_runstate(td, TD_INITIALIZED);
789 fio_sem_up(startup_sem);
790 fio_sem_down(td->mutex);
793 * the ->mutex semaphore is now no longer used, close it to avoid
794 * eating a file descriptor
796 fio_sem_remove(td->mutex);
798 if (!td->o.create_serialize && setup_files(td))
807 if (init_random_map(td))
810 if (td->o.exec_prerun) {
811 if (system(td->o.exec_prerun) < 0)
815 fio_gettime(&td->epoch, NULL);
816 memcpy(&td->timeout_end, &td->epoch, sizeof(td->epoch));
817 getrusage(RUSAGE_SELF, &td->ts.ru_start);
819 runtime[0] = runtime[1] = 0;
821 while (td->o.loops--) {
822 fio_gettime(&td->start, NULL);
823 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
826 memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
828 if (clear_state && clear_io_state(td))
831 prune_io_piece_log(td);
837 if (td_read(td) && td->io_bytes[DDIR_READ]) {
838 if (td->rw_end_set[DDIR_READ])
839 elapsed = utime_since(&td->start, &td->rw_end[DDIR_READ]);
841 elapsed = utime_since_now(&td->start);
843 runtime[DDIR_READ] += elapsed;
845 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
846 if (td->rw_end_set[DDIR_WRITE])
847 elapsed = utime_since(&td->start, &td->rw_end[DDIR_WRITE]);
849 elapsed = utime_since_now(&td->start);
851 runtime[DDIR_WRITE] += elapsed;
854 if (td->error || td->terminate)
857 if (td->o.verify == VERIFY_NONE)
860 if (clear_io_state(td))
863 fio_gettime(&td->start, NULL);
867 runtime[DDIR_READ] += utime_since_now(&td->start);
869 if (td->error || td->terminate)
873 update_rusage_stat(td);
874 td->ts.runtime[0] = runtime[0] / 1000;
875 td->ts.runtime[1] = runtime[1] / 1000;
876 td->ts.total_run_time = mtime_since_now(&td->epoch);
877 td->ts.io_bytes[0] = td->io_bytes[0];
878 td->ts.io_bytes[1] = td->io_bytes[1];
881 finish_log(td, td->ts.bw_log, "bw");
883 finish_log(td, td->ts.slat_log, "slat");
885 finish_log(td, td->ts.clat_log, "clat");
886 if (td->o.write_iolog_file)
887 write_iolog_close(td);
888 if (td->o.exec_postrun) {
889 if (system(td->o.exec_postrun) < 0)
890 log_err("fio: postrun %s failed\n", td->o.exec_postrun);
893 if (exitall_on_terminate)
894 terminate_threads(td->groupid);
898 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
902 options_mem_free(td);
903 td_set_runstate(td, TD_EXITED);
904 return (void *) (unsigned long) td->error;
906 fio_sem_up(startup_sem);
911 * We cannot pass the td data into a forked process, so attach the td and
912 * pass it to the thread worker.
914 static int fork_main(int shmid, int offset)
916 struct thread_data *td;
919 data = shmat(shmid, NULL, 0);
920 if (data == (void *) -1) {
927 td = data + offset * sizeof(struct thread_data);
928 ret = thread_main(td);
930 return (int) (unsigned long) ret;
934 * Run over the job map and reap the threads that have exited, if any.
936 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
938 struct thread_data *td;
939 int i, cputhreads, pending, status, ret;
942 * reap exited threads (TD_EXITED -> TD_REAPED)
944 pending = cputhreads = 0;
949 * ->io_ops is NULL for a thread that has closed its
952 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
955 if (!td->pid || td->runstate == TD_REAPED)
957 if (td->o.use_thread) {
958 if (td->runstate == TD_EXITED) {
959 td_set_runstate(td, TD_REAPED);
966 if (td->runstate == TD_EXITED)
970 * check if someone quit or got killed in an unusual way
972 ret = waitpid(td->pid, &status, flags);
974 if (errno == ECHILD) {
975 log_err("fio: pid=%d disappeared %d\n", td->pid, td->runstate);
976 td_set_runstate(td, TD_REAPED);
980 } else if (ret == td->pid) {
981 if (WIFSIGNALED(status)) {
982 int sig = WTERMSIG(status);
985 log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
986 td_set_runstate(td, TD_REAPED);
989 if (WIFEXITED(status)) {
990 if (WEXITSTATUS(status) && !td->error)
991 td->error = WEXITSTATUS(status);
993 td_set_runstate(td, TD_REAPED);
999 * thread is not dead, continue
1004 if (td->o.use_thread) {
1007 if (pthread_join(td->thread, (void *) &ret))
1008 perror("pthread_join");
1012 (*m_rate) -= td->o.ratemin;
1013 (*t_rate) -= td->o.rate;
1020 if (*nr_running == cputhreads && !pending)
1021 terminate_threads(TERMINATE_ALL);
1025 * Main function for kicking off and reaping jobs, as needed.
1027 static void run_threads(void)
1029 struct thread_data *td;
1030 unsigned long spent;
1031 int i, todo, nr_running, m_rate, t_rate, nr_started;
1033 if (fio_pin_memory())
1036 if (!terse_output) {
1037 printf("Starting ");
1039 printf("%d thread%s", nr_thread, nr_thread > 1 ? "s" : "");
1043 printf("%d process%s", nr_process, nr_process > 1 ? "es" : "");
1049 signal(SIGINT, sig_handler);
1050 signal(SIGALRM, sig_handler);
1052 todo = thread_number;
1055 m_rate = t_rate = 0;
1057 for_each_td(td, i) {
1058 print_status_init(td->thread_number - 1);
1060 if (!td->o.create_serialize) {
1066 * do file setup here so it happens sequentially,
1067 * we don't want X number of threads getting their
1068 * client data interspersed on disk
1070 if (setup_files(td)) {
1073 log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
1074 td_set_runstate(td, TD_REAPED);
1084 struct thread_data *map[MAX_JOBS];
1085 struct timeval this_start;
1086 int this_jobs = 0, left;
1089 * create threads (TD_NOT_CREATED -> TD_CREATED)
1091 for_each_td(td, i) {
1092 if (td->runstate != TD_NOT_CREATED)
1096 * never got a chance to start, killed by other
1097 * thread for some reason
1099 if (td->terminate) {
1104 if (td->o.start_delay) {
1105 spent = mtime_since_genesis();
1107 if (td->o.start_delay * 1000 > spent)
1111 if (td->o.stonewall && (nr_started || nr_running))
1115 * Set state to created. Thread will transition
1116 * to TD_INITIALIZED when it's done setting up.
1118 td_set_runstate(td, TD_CREATED);
1119 map[this_jobs++] = td;
1122 if (td->o.use_thread) {
1123 if (pthread_create(&td->thread, NULL, thread_main, td)) {
1124 perror("thread_create");
1130 int ret = fork_main(shm_id, i);
1135 fio_sem_down(startup_sem);
1139 * Wait for the started threads to transition to
1142 fio_gettime(&this_start, NULL);
1144 while (left && !fio_abort) {
1145 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1150 for (i = 0; i < this_jobs; i++) {
1154 if (td->runstate == TD_INITIALIZED) {
1157 } else if (td->runstate >= TD_EXITED) {
1161 nr_running++; /* work-around... */
1167 log_err("fio: %d jobs failed to start\n", left);
1168 for (i = 0; i < this_jobs; i++) {
1172 kill(td->pid, SIGTERM);
1178 * start created threads (TD_INITIALIZED -> TD_RUNNING).
1180 for_each_td(td, i) {
1181 if (td->runstate != TD_INITIALIZED)
1184 td_set_runstate(td, TD_RUNNING);
1187 m_rate += td->o.ratemin;
1188 t_rate += td->o.rate;
1190 fio_sem_up(td->mutex);
1193 reap_threads(&nr_running, &t_rate, &m_rate);
1199 while (nr_running) {
1200 reap_threads(&nr_running, &t_rate, &m_rate);
1208 int main(int argc, char *argv[])
1213 * We need locale for number printing, if it isn't set then just
1214 * go with the US format.
1216 if (!getenv("LC_NUMERIC"))
1217 setlocale(LC_NUMERIC, "en_US");
1219 if (parse_options(argc, argv))
1225 ps = sysconf(_SC_PAGESIZE);
1227 log_err("Failed to get page size\n");
1235 setup_log(&agg_io_log[DDIR_READ]);
1236 setup_log(&agg_io_log[DDIR_WRITE]);
1239 startup_sem = fio_sem_init(0);
1243 disk_util_timer_arm();
1250 __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1251 __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");
1255 fio_sem_remove(startup_sem);