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 static unsigned long page_mask;
42 (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
45 int thread_number = 0;
49 static volatile int startup_sem;
50 static volatile int fio_abort;
51 static int exit_value;
53 struct io_log *agg_io_log[2];
55 #define TERMINATE_ALL (-1)
56 #define JOB_START_TIMEOUT (5 * 1000)
58 static inline void td_set_runstate(struct thread_data *td, int runstate)
60 td->runstate = runstate;
63 static void terminate_threads(int group_id, int forced_kill)
65 struct thread_data *td;
69 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
70 kill(td->pid, SIGQUIT);
74 td_set_runstate(td, TD_EXITED);
79 static void sig_handler(int sig)
84 disk_util_timer_arm();
85 print_thread_status();
88 printf("\nfio: terminating on signal %d\n", sig);
90 terminate_threads(TERMINATE_ALL, 0);
96 * Check if we are above the minimum rate given.
98 static int check_min_rate(struct thread_data *td, struct timeval *now)
100 unsigned long long bytes = 0;
105 * No minimum rate set, always ok
111 * allow a 2 second settle period in the beginning
113 if (mtime_since(&td->start, now) < 2000)
117 bytes += td->this_io_bytes[DDIR_READ];
119 bytes += td->this_io_bytes[DDIR_WRITE];
122 * if rate blocks is set, sample is running
124 if (td->rate_bytes) {
125 spent = mtime_since(&td->lastrate, now);
126 if (spent < td->ratecycle)
129 if (bytes < td->rate_bytes) {
130 fprintf(f_out, "%s: min rate %u not met\n", td->name, td->ratemin);
133 rate = (bytes - td->rate_bytes) / spent;
134 if (rate < td->ratemin || bytes < td->rate_bytes) {
135 fprintf(f_out, "%s: min rate %u not met, got %luKiB/sec\n", td->name, td->ratemin, rate);
141 td->rate_bytes = bytes;
142 memcpy(&td->lastrate, now, sizeof(*now));
146 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
150 if (mtime_since(&td->epoch, t) >= td->timeout * 1000)
157 * When job exits, we can cancel the in-flight IO if we are using async
158 * io. Attempt to do so.
160 static void cleanup_pending_aio(struct thread_data *td)
162 struct list_head *entry, *n;
167 * get immediately available events, if any
169 r = io_u_queued_complete(td, 0);
174 * now cancel remaining active events
176 if (td->io_ops->cancel) {
177 list_for_each_safe(entry, n, &td->io_u_busylist) {
178 io_u = list_entry(entry, struct io_u, list);
181 * if the io_u isn't in flight, then that generally
182 * means someone leaked an io_u. complain but fix
183 * it up, so we don't stall here.
185 if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
186 log_err("fio: non-busy IO on busy list\n");
189 r = td->io_ops->cancel(td, io_u);
197 r = io_u_queued_complete(td, td->cur_depth);
201 * Helper to handle the final sync of a file. Works just like the normal
202 * io path, just does everything sync.
204 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
206 struct io_u *io_u = __get_io_u(td);
212 io_u->ddir = DDIR_SYNC;
215 if (td_io_prep(td, io_u)) {
221 ret = td_io_queue(td, io_u);
223 td_verror(td, io_u->error, "td_io_queue");
226 } else if (ret == FIO_Q_QUEUED) {
227 if (io_u_queued_complete(td, 1) < 0)
229 } else if (ret == FIO_Q_COMPLETED) {
231 td_verror(td, io_u->error, "td_io_queue");
235 if (io_u_sync_complete(td, io_u) < 0)
237 } else if (ret == FIO_Q_BUSY) {
238 if (td_io_commit(td))
247 * The main verify engine. Runs over the writes we previously submitted,
248 * reads the blocks back in, and checks the crc/md5 of the data.
250 static void do_verify(struct thread_data *td)
254 int ret, i, min_events;
257 * sync io first and invalidate cache, to make sure we really
260 for_each_file(td, f, i) {
261 if (fio_io_sync(td, f))
263 if (file_invalidate_cache(td, f))
270 td_set_runstate(td, TD_VERIFYING);
273 while (!td->terminate) {
276 io_u = __get_io_u(td);
280 if (runtime_exceeded(td, &io_u->start_time)) {
285 if (get_next_verify(td, io_u)) {
290 if (td_io_prep(td, io_u)) {
295 io_u->end_io = verify_io_u;
297 ret = td_io_queue(td, io_u);
300 case FIO_Q_COMPLETED:
303 else if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
304 int bytes = io_u->xfer_buflen - io_u->resid;
306 io_u->xfer_buflen = io_u->resid;
307 io_u->xfer_buf += bytes;
310 ret = io_u_sync_complete(td, io_u);
317 requeue_io_u(td, &io_u);
318 ret2 = td_io_commit(td);
324 td_verror(td, -ret, "td_io_queue");
328 if (ret < 0 || td->error)
332 * if we can queue more, do so. but check if there are
333 * completed io_u's first.
336 if (queue_full(td) || ret == FIO_Q_BUSY) {
339 if (td->cur_depth > td->iodepth_low)
340 min_events = td->cur_depth - td->iodepth_low;
344 * Reap required number of io units, if any, and do the
345 * verification on them through the callback handler
347 if (io_u_queued_complete(td, min_events) < 0)
352 min_events = td->cur_depth;
355 ret = io_u_queued_complete(td, min_events);
357 cleanup_pending_aio(td);
359 td_set_runstate(td, TD_RUNNING);
363 * Not really an io thread, all it does is burn CPU cycles in the specified
366 static void do_cpuio(struct thread_data *td)
369 int split = 100 / td->cpuload;
372 while (!td->terminate) {
373 fio_gettime(&e, NULL);
375 if (runtime_exceeded(td, &e))
381 usec_sleep(td, 10000);
388 * Main IO worker function. It retrieves io_u's to process and queues
389 * and reaps them, checking for rate and errors along the way.
391 static void do_io(struct thread_data *td)
397 td_set_runstate(td, TD_RUNNING);
399 while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->io_size) {
400 struct timeval comp_time;
413 memcpy(&s, &io_u->start_time, sizeof(s));
415 if (runtime_exceeded(td, &s)) {
420 ret = td_io_queue(td, io_u);
423 case FIO_Q_COMPLETED:
426 else if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
427 int bytes = io_u->xfer_buflen - io_u->resid;
429 io_u->xfer_buflen = io_u->resid;
430 io_u->xfer_buf += bytes;
433 fio_gettime(&comp_time, NULL);
434 bytes_done = io_u_sync_complete(td, io_u);
440 * if the engine doesn't have a commit hook,
441 * the io_u is really queued. if it does have such
442 * a hook, it has to call io_u_queued() itself.
444 if (td->io_ops->commit == NULL)
445 io_u_queued(td, io_u);
448 requeue_io_u(td, &io_u);
449 ret2 = td_io_commit(td);
459 if (ret < 0 || td->error)
463 * See if we need to complete some commands
465 if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
467 if (queue_full(td) || ret == FIO_Q_BUSY) {
470 if (td->cur_depth > td->iodepth_low)
471 min_evts = td->cur_depth - td->iodepth_low;
474 fio_gettime(&comp_time, NULL);
475 bytes_done = io_u_queued_complete(td, min_evts);
484 * the rate is batched for now, it should work for batches
485 * of completions except the very first one which may look
488 usec = utime_since(&s, &comp_time);
490 rate_throttle(td, usec, bytes_done);
492 if (check_min_rate(td, &comp_time)) {
493 if (exitall_on_terminate)
494 terminate_threads(td->groupid, 0);
495 td_verror(td, ENODATA, "check_min_rate");
500 unsigned long long b;
502 b = td->io_blocks[0] + td->io_blocks[1];
503 if (!(b % td->thinktime_blocks)) {
506 if (td->thinktime_spin)
507 __usec_sleep(td->thinktime_spin);
509 left = td->thinktime - td->thinktime_spin;
511 usec_sleep(td, left);
521 ret = io_u_queued_complete(td, i);
523 if (should_fsync(td) && td->end_fsync) {
524 td_set_runstate(td, TD_FSYNCING);
525 for_each_file(td, f, i)
529 cleanup_pending_aio(td);
532 static void cleanup_io_u(struct thread_data *td)
534 struct list_head *entry, *n;
537 list_for_each_safe(entry, n, &td->io_u_freelist) {
538 io_u = list_entry(entry, struct io_u, list);
540 list_del(&io_u->list);
548 * "randomly" fill the buffer contents
550 static void fill_rand_buf(struct io_u *io_u, int max_bs)
552 int *ptr = io_u->buf;
554 while ((void *) ptr - io_u->buf < max_bs) {
555 *ptr = rand() * 0x9e370001;
560 static int init_io_u(struct thread_data *td)
567 if (td->io_ops->flags & FIO_CPUIO)
570 if (td->io_ops->flags & FIO_SYNCIO)
573 max_units = td->iodepth;
575 max_bs = max(td->max_bs[DDIR_READ], td->max_bs[DDIR_WRITE]);
576 td->orig_buffer_size = max_bs * max_units;
578 if (td->mem_type == MEM_SHMHUGE || td->mem_type == MEM_MMAPHUGE)
579 td->orig_buffer_size = (td->orig_buffer_size + td->hugepage_size - 1) & ~(td->hugepage_size - 1);
581 td->orig_buffer_size += page_mask;
583 if (allocate_io_mem(td))
586 p = ALIGN(td->orig_buffer);
587 for (i = 0; i < max_units; i++) {
588 io_u = malloc(sizeof(*io_u));
589 memset(io_u, 0, sizeof(*io_u));
590 INIT_LIST_HEAD(&io_u->list);
592 io_u->buf = p + max_bs * i;
593 if (td_write(td) || td_rw(td))
594 fill_rand_buf(io_u, max_bs);
597 io_u->flags = IO_U_F_FREE;
598 list_add(&io_u->list, &td->io_u_freelist);
606 static int switch_ioscheduler(struct thread_data *td)
608 char tmp[256], tmp2[128];
612 if (td->io_ops->flags & FIO_CPUIO)
615 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
617 f = fopen(tmp, "r+");
619 td_verror(td, errno, "fopen");
626 ret = fwrite(td->ioscheduler, strlen(td->ioscheduler), 1, f);
627 if (ferror(f) || ret != 1) {
628 td_verror(td, errno, "fwrite");
636 * Read back and check that the selected scheduler is now the default.
638 ret = fread(tmp, 1, sizeof(tmp), f);
639 if (ferror(f) || ret < 0) {
640 td_verror(td, errno, "fread");
645 sprintf(tmp2, "[%s]", td->ioscheduler);
646 if (!strstr(tmp, tmp2)) {
647 log_err("fio: io scheduler %s not found\n", td->ioscheduler);
648 td_verror(td, EINVAL, "iosched_switch");
657 static void clear_io_state(struct thread_data *td)
662 td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
663 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
666 td->last_was_sync = 0;
668 for_each_file(td, f, i) {
669 f->last_completed_pos = 0;
672 if (td->io_ops->flags & FIO_SYNCIO)
673 lseek(f->fd, SEEK_SET, 0);
676 memset(f->file_map, 0, f->num_maps * sizeof(long));
681 * Entry point for the thread based jobs. The process based jobs end up
682 * here as well, after a little setup.
684 static void *thread_main(void *data)
686 unsigned long long runtime[2];
687 struct thread_data *td = data;
694 INIT_LIST_HEAD(&td->io_u_freelist);
695 INIT_LIST_HEAD(&td->io_u_busylist);
696 INIT_LIST_HEAD(&td->io_u_requeues);
697 INIT_LIST_HEAD(&td->io_hist_list);
698 INIT_LIST_HEAD(&td->io_log_list);
703 if (fio_setaffinity(td) == -1) {
704 td_verror(td, errno, "cpu_set_affinity");
712 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
713 td_verror(td, errno, "ioprio_set");
718 if (nice(td->nice) == -1) {
719 td_verror(td, errno, "nice");
723 if (init_random_state(td))
726 if (td->ioscheduler && switch_ioscheduler(td))
729 td_set_runstate(td, TD_INITIALIZED);
730 fio_sem_up(&startup_sem);
731 fio_sem_down(&td->mutex);
733 if (!td->create_serialize && setup_files(td))
739 * Do this late, as some IO engines would like to have the
740 * files setup prior to initializing structures.
745 if (td->exec_prerun) {
746 if (system(td->exec_prerun) < 0)
750 fio_gettime(&td->epoch, NULL);
751 memcpy(&td->timeout_end, &td->epoch, sizeof(td->epoch));
752 getrusage(RUSAGE_SELF, &td->ts.ru_start);
754 runtime[0] = runtime[1] = 0;
755 while (td->loops--) {
756 fio_gettime(&td->start, NULL);
757 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
760 memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
763 prune_io_piece_log(td);
765 if (td->io_ops->flags & FIO_CPUIO)
770 if (td_read(td) && td->io_bytes[DDIR_READ])
771 runtime[DDIR_READ] += utime_since_now(&td->start);
772 if (td_write(td) && td->io_bytes[DDIR_WRITE])
773 runtime[DDIR_WRITE] += utime_since_now(&td->start);
775 if (td->error || td->terminate)
778 if (td->verify == VERIFY_NONE)
782 fio_gettime(&td->start, NULL);
786 runtime[DDIR_READ] += utime_since_now(&td->start);
788 if (td->error || td->terminate)
792 update_rusage_stat(td);
793 fio_gettime(&td->end_time, NULL);
794 td->runtime[0] = runtime[0] / 1000;
795 td->runtime[1] = runtime[1] / 1000;
798 finish_log(td, td->ts.bw_log, "bw");
800 finish_log(td, td->ts.slat_log, "slat");
802 finish_log(td, td->ts.clat_log, "clat");
803 if (td->write_iolog_file)
804 write_iolog_close(td);
805 if (td->exec_postrun) {
806 if (system(td->exec_postrun) < 0)
807 log_err("fio: postrun %s failed\n", td->exec_postrun);
810 if (exitall_on_terminate)
811 terminate_threads(td->groupid, 0);
815 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
819 td_set_runstate(td, TD_EXITED);
820 return (void *) (unsigned long) td->error;
824 * We cannot pass the td data into a forked process, so attach the td and
825 * pass it to the thread worker.
827 static int fork_main(int shmid, int offset)
829 struct thread_data *td;
832 data = shmat(shmid, NULL, 0);
833 if (data == (void *) -1) {
840 td = data + offset * sizeof(struct thread_data);
841 ret = thread_main(td);
843 return (int) (unsigned long) ret;
847 * Run over the job map and reap the threads that have exited, if any.
849 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
851 struct thread_data *td;
852 int i, cputhreads, pending, status, ret;
855 * reap exited threads (TD_EXITED -> TD_REAPED)
857 pending = cputhreads = 0;
862 * ->io_ops is NULL for a thread that has closed its
865 if (td->io_ops && td->io_ops->flags & FIO_CPUIO)
868 if (!td->pid || td->runstate == TD_REAPED)
870 if (td->use_thread) {
871 if (td->runstate == TD_EXITED) {
872 td_set_runstate(td, TD_REAPED);
879 if (td->runstate == TD_EXITED)
883 * check if someone quit or got killed in an unusual way
885 ret = waitpid(td->pid, &status, flags);
887 if (errno == ECHILD) {
888 log_err("fio: pid=%d disappeared %d\n", td->pid, td->runstate);
889 td_set_runstate(td, TD_REAPED);
893 } else if (ret == td->pid) {
894 if (WIFSIGNALED(status)) {
895 int sig = WTERMSIG(status);
898 log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
899 td_set_runstate(td, TD_REAPED);
902 if (WIFEXITED(status)) {
903 if (WEXITSTATUS(status) && !td->error)
904 td->error = WEXITSTATUS(status);
906 td_set_runstate(td, TD_REAPED);
912 * thread is not dead, continue
916 if (td->use_thread) {
919 if (pthread_join(td->thread, (void *) &ret))
920 perror("pthread_join");
924 (*m_rate) -= td->ratemin;
925 (*t_rate) -= td->rate;
931 if (*nr_running == cputhreads && !pending)
932 terminate_threads(TERMINATE_ALL, 0);
936 * Main function for kicking off and reaping jobs, as needed.
938 static void run_threads(void)
940 struct thread_data *td;
942 int i, todo, nr_running, m_rate, t_rate, nr_started;
944 if (fio_pin_memory())
948 printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
952 signal(SIGINT, sig_handler);
953 signal(SIGALRM, sig_handler);
955 todo = thread_number;
961 print_status_init(td->thread_number - 1);
963 if (!td->create_serialize) {
969 * do file setup here so it happens sequentially,
970 * we don't want X number of threads getting their
971 * client data interspersed on disk
973 if (setup_files(td)) {
976 log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
977 td_set_runstate(td, TD_REAPED);
987 struct thread_data *map[MAX_JOBS];
988 struct timeval this_start;
989 int this_jobs = 0, left;
992 * create threads (TD_NOT_CREATED -> TD_CREATED)
995 if (td->runstate != TD_NOT_CREATED)
999 * never got a chance to start, killed by other
1000 * thread for some reason
1002 if (td->terminate) {
1007 if (td->start_delay) {
1008 spent = mtime_since_genesis();
1010 if (td->start_delay * 1000 > spent)
1014 if (td->stonewall && (nr_started || nr_running))
1018 * Set state to created. Thread will transition
1019 * to TD_INITIALIZED when it's done setting up.
1021 td_set_runstate(td, TD_CREATED);
1022 map[this_jobs++] = td;
1023 fio_sem_init(&startup_sem, 1);
1026 if (td->use_thread) {
1027 if (pthread_create(&td->thread, NULL, thread_main, td)) {
1028 perror("thread_create");
1033 fio_sem_down(&startup_sem);
1035 int ret = fork_main(shm_id, i);
1043 * Wait for the started threads to transition to
1046 fio_gettime(&this_start, NULL);
1048 while (left && !fio_abort) {
1049 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1054 for (i = 0; i < this_jobs; i++) {
1058 if (td->runstate == TD_INITIALIZED) {
1061 } else if (td->runstate >= TD_EXITED) {
1065 nr_running++; /* work-around... */
1071 log_err("fio: %d jobs failed to start\n", left);
1072 for (i = 0; i < this_jobs; i++) {
1076 kill(td->pid, SIGTERM);
1082 * start created threads (TD_INITIALIZED -> TD_RUNNING).
1084 for_each_td(td, i) {
1085 if (td->runstate != TD_INITIALIZED)
1088 td_set_runstate(td, TD_RUNNING);
1091 m_rate += td->ratemin;
1094 fio_sem_up(&td->mutex);
1097 reap_threads(&nr_running, &t_rate, &m_rate);
1103 while (nr_running) {
1104 reap_threads(&nr_running, &t_rate, &m_rate);
1112 int main(int argc, char *argv[])
1117 * We need locale for number printing, if it isn't set then just
1118 * go with the US format.
1120 if (!getenv("LC_NUMERIC"))
1121 setlocale(LC_NUMERIC, "en_US");
1123 if (parse_options(argc, argv))
1126 if (!thread_number) {
1127 log_err("Nothing to do\n");
1131 ps = sysconf(_SC_PAGESIZE);
1133 log_err("Failed to get page size\n");
1140 setup_log(&agg_io_log[DDIR_READ]);
1141 setup_log(&agg_io_log[DDIR_WRITE]);
1146 disk_util_timer_arm();
1153 __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1154 __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");