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) {
73 td_set_runstate(td, TD_EXITED);
78 static void sig_handler(int sig)
83 disk_util_timer_arm();
84 print_thread_status();
87 printf("\nfio: terminating on signal %d\n", sig);
89 terminate_threads(TERMINATE_ALL, 0);
95 * Check if we are above the minimum rate given.
97 static int check_min_rate(struct thread_data *td, struct timeval *now)
103 * allow a 2 second settle period in the beginning
105 if (mtime_since(&td->start, now) < 2000)
109 * if rate blocks is set, sample is running
111 if (td->rate_bytes) {
112 unsigned long long bytes = 0;
114 spent = mtime_since(&td->lastrate, now);
115 if (spent < td->ratecycle)
119 bytes += td->this_io_bytes[DDIR_READ];
121 bytes += td->this_io_bytes[DDIR_WRITE];
123 if (bytes < td->rate_bytes) {
124 fprintf(f_out, "%s: min rate %u not met\n", td->name, td->ratemin);
127 rate = (bytes - td->rate_bytes) / spent;
128 if (rate < td->ratemin || bytes < td->rate_bytes) {
129 fprintf(f_out, "%s: min rate %u not met, got %luKiB/sec\n", td->name, td->ratemin, rate);
133 td->rate_bytes = bytes;
136 memcpy(&td->lastrate, now, sizeof(*now));
140 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
144 if (mtime_since(&td->epoch, t) >= td->timeout * 1000)
151 * When job exits, we can cancel the in-flight IO if we are using async
152 * io. Attempt to do so.
154 static void cleanup_pending_aio(struct thread_data *td)
156 struct list_head *entry, *n;
161 * get immediately available events, if any
163 r = io_u_queued_complete(td, 0, NULL);
168 * now cancel remaining active events
170 if (td->io_ops->cancel) {
171 list_for_each_safe(entry, n, &td->io_u_busylist) {
172 io_u = list_entry(entry, struct io_u, list);
175 * if the io_u isn't in flight, then that generally
176 * means someone leaked an io_u. complain but fix
177 * it up, so we don't stall here.
179 if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
180 log_err("fio: non-busy IO on busy list\n");
183 r = td->io_ops->cancel(td, io_u);
191 r = io_u_queued_complete(td, td->cur_depth, NULL);
195 * Helper to handle the final sync of a file. Works just like the normal
196 * io path, just does everything sync.
198 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
200 struct io_u *io_u = __get_io_u(td);
206 io_u->ddir = DDIR_SYNC;
209 if (td_io_prep(td, io_u)) {
215 ret = td_io_queue(td, io_u);
217 td_verror(td, io_u->error, "td_io_queue");
220 } else if (ret == FIO_Q_QUEUED) {
221 if (io_u_queued_complete(td, 1, NULL) < 0)
223 } else if (ret == FIO_Q_COMPLETED) {
225 td_verror(td, io_u->error, "td_io_queue");
229 if (io_u_sync_complete(td, io_u, NULL) < 0)
231 } else if (ret == FIO_Q_BUSY) {
232 if (td_io_commit(td))
241 * The main verify engine. Runs over the writes we previusly submitted,
242 * reads the blocks back in, and checks the crc/md5 of the data.
244 static void do_verify(struct thread_data *td)
248 int ret, i, min_events;
251 * sync io first and invalidate cache, to make sure we really
254 for_each_file(td, f, i) {
255 if (fio_io_sync(td, f))
257 if (file_invalidate_cache(td, f))
264 td_set_runstate(td, TD_VERIFYING);
267 while (!td->terminate) {
268 io_u = __get_io_u(td);
272 if (runtime_exceeded(td, &io_u->start_time)) {
277 if (get_next_verify(td, io_u)) {
282 if (td_io_prep(td, io_u)) {
287 ret = td_io_queue(td, io_u);
290 case FIO_Q_COMPLETED:
293 if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
294 int bytes = io_u->xfer_buflen - io_u->resid;
296 io_u->xfer_buflen = io_u->resid;
297 io_u->xfer_buf += bytes;
300 ret = io_u_sync_complete(td, io_u, verify_io_u);
307 requeue_io_u(td, &io_u);
308 ret = td_io_commit(td);
312 td_verror(td, -ret, "td_io_queue");
316 if (ret < 0 || td->error)
320 * if we can queue more, do so. but check if there are
321 * completed io_u's first.
324 if (queue_full(td) || ret == FIO_Q_BUSY) {
327 if (td->cur_depth > td->iodepth_low)
328 min_events = td->cur_depth - td->iodepth_low;
332 * Reap required number of io units, if any, and do the
333 * verification on them through the callback handler
335 if (io_u_queued_complete(td, min_events, verify_io_u) < 0)
340 cleanup_pending_aio(td);
342 td_set_runstate(td, TD_RUNNING);
346 * Not really an io thread, all it does is burn CPU cycles in the specified
349 static void do_cpuio(struct thread_data *td)
352 int split = 100 / td->cpuload;
355 while (!td->terminate) {
356 fio_gettime(&e, NULL);
358 if (runtime_exceeded(td, &e))
364 usec_sleep(td, 10000);
371 * Main IO worker function. It retrieves io_u's to process and queues
372 * and reaps them, checking for rate and errors along the way.
374 static void do_io(struct thread_data *td)
380 td_set_runstate(td, TD_RUNNING);
382 while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->io_size) {
383 struct timeval comp_time;
395 memcpy(&s, &io_u->start_time, sizeof(s));
397 if (runtime_exceeded(td, &s)) {
402 ret = td_io_queue(td, io_u);
405 case FIO_Q_COMPLETED:
410 if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
411 int bytes = io_u->xfer_buflen - io_u->resid;
413 io_u->xfer_buflen = io_u->resid;
414 io_u->xfer_buf += bytes;
417 fio_gettime(&comp_time, NULL);
418 bytes_done = io_u_sync_complete(td, io_u, NULL);
424 * if the engine doesn't have a commit hook,
425 * the io_u is really queued. if it does have such
426 * a hook, it has to call io_u_queued() itself.
428 if (td->io_ops->commit == NULL)
429 io_u_queued(td, io_u);
432 requeue_io_u(td, &io_u);
433 ret = td_io_commit(td);
441 if (ret < 0 || td->error)
445 * See if we need to complete some commands
447 if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
449 if (queue_full(td) || ret == FIO_Q_BUSY) {
452 if (td->cur_depth > td->iodepth_low)
453 min_evts = td->cur_depth - td->iodepth_low;
456 fio_gettime(&comp_time, NULL);
457 bytes_done = io_u_queued_complete(td, min_evts, NULL);
466 * the rate is batched for now, it should work for batches
467 * of completions except the very first one which may look
470 usec = utime_since(&s, &comp_time);
472 rate_throttle(td, usec, bytes_done);
474 if (check_min_rate(td, &comp_time)) {
475 if (exitall_on_terminate)
476 terminate_threads(td->groupid, 0);
477 td_verror(td, ENODATA, "check_min_rate");
482 unsigned long long b;
484 b = td->io_blocks[0] + td->io_blocks[1];
485 if (!(b % td->thinktime_blocks)) {
488 if (td->thinktime_spin)
489 __usec_sleep(td->thinktime_spin);
491 left = td->thinktime - td->thinktime_spin;
493 usec_sleep(td, left);
502 cleanup_pending_aio(td);
504 if (should_fsync(td) && td->end_fsync) {
505 td_set_runstate(td, TD_FSYNCING);
506 for_each_file(td, f, i)
512 static void cleanup_io_u(struct thread_data *td)
514 struct list_head *entry, *n;
517 list_for_each_safe(entry, n, &td->io_u_freelist) {
518 io_u = list_entry(entry, struct io_u, list);
520 list_del(&io_u->list);
528 * "randomly" fill the buffer contents
530 static void fill_rand_buf(struct io_u *io_u, int max_bs)
532 int *ptr = io_u->buf;
534 while ((void *) ptr - io_u->buf < max_bs) {
535 *ptr = rand() * 0x9e370001;
540 static int init_io_u(struct thread_data *td)
547 if (td->io_ops->flags & FIO_CPUIO)
550 if (td->io_ops->flags & FIO_SYNCIO)
553 max_units = td->iodepth;
555 max_bs = max(td->max_bs[DDIR_READ], td->max_bs[DDIR_WRITE]);
556 td->orig_buffer_size = max_bs * max_units;
558 if (td->mem_type == MEM_SHMHUGE || td->mem_type == MEM_MMAPHUGE)
559 td->orig_buffer_size = (td->orig_buffer_size + td->hugepage_size - 1) & ~(td->hugepage_size - 1);
561 td->orig_buffer_size += page_mask;
563 if (allocate_io_mem(td))
566 p = ALIGN(td->orig_buffer);
567 for (i = 0; i < max_units; i++) {
568 io_u = malloc(sizeof(*io_u));
569 memset(io_u, 0, sizeof(*io_u));
570 INIT_LIST_HEAD(&io_u->list);
572 io_u->buf = p + max_bs * i;
573 if (td_write(td) || td_rw(td))
574 fill_rand_buf(io_u, max_bs);
577 io_u->flags = IO_U_F_FREE;
578 list_add(&io_u->list, &td->io_u_freelist);
586 static int switch_ioscheduler(struct thread_data *td)
588 char tmp[256], tmp2[128];
592 if (td->io_ops->flags & FIO_CPUIO)
595 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
597 f = fopen(tmp, "r+");
599 td_verror(td, errno, "fopen");
606 ret = fwrite(td->ioscheduler, strlen(td->ioscheduler), 1, f);
607 if (ferror(f) || ret != 1) {
608 td_verror(td, errno, "fwrite");
616 * Read back and check that the selected scheduler is now the default.
618 ret = fread(tmp, 1, sizeof(tmp), f);
619 if (ferror(f) || ret < 0) {
620 td_verror(td, errno, "fread");
625 sprintf(tmp2, "[%s]", td->ioscheduler);
626 if (!strstr(tmp, tmp2)) {
627 log_err("fio: io scheduler %s not found\n", td->ioscheduler);
628 td_verror(td, EINVAL, "iosched_switch");
637 static void clear_io_state(struct thread_data *td)
642 td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
643 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
646 td->last_was_sync = 0;
648 for_each_file(td, f, i) {
649 f->last_completed_pos = 0;
652 if (td->io_ops->flags & FIO_SYNCIO)
653 lseek(f->fd, SEEK_SET, 0);
656 memset(f->file_map, 0, f->num_maps * sizeof(long));
661 * Entry point for the thread based jobs. The process based jobs end up
662 * here as well, after a little setup.
664 static void *thread_main(void *data)
666 unsigned long long runtime[2];
667 struct thread_data *td = data;
674 INIT_LIST_HEAD(&td->io_u_freelist);
675 INIT_LIST_HEAD(&td->io_u_busylist);
676 INIT_LIST_HEAD(&td->io_u_requeues);
677 INIT_LIST_HEAD(&td->io_hist_list);
678 INIT_LIST_HEAD(&td->io_log_list);
683 if (fio_setaffinity(td) == -1) {
684 td_verror(td, errno, "cpu_set_affinity");
692 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
693 td_verror(td, errno, "ioprio_set");
698 if (nice(td->nice) == -1) {
699 td_verror(td, errno, "nice");
703 if (init_random_state(td))
706 if (td->ioscheduler && switch_ioscheduler(td))
709 td_set_runstate(td, TD_INITIALIZED);
710 fio_sem_up(&startup_sem);
711 fio_sem_down(&td->mutex);
713 if (!td->create_serialize && setup_files(td))
719 * Do this late, as some IO engines would like to have the
720 * files setup prior to initializing structures.
725 if (td->exec_prerun) {
726 if (system(td->exec_prerun) < 0)
730 fio_gettime(&td->epoch, NULL);
731 memcpy(&td->timeout_end, &td->epoch, sizeof(td->epoch));
732 getrusage(RUSAGE_SELF, &td->ts.ru_start);
734 runtime[0] = runtime[1] = 0;
735 while (td->loops--) {
736 fio_gettime(&td->start, NULL);
737 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
740 memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
743 prune_io_piece_log(td);
745 if (td->io_ops->flags & FIO_CPUIO)
750 if (td_read(td) && td->io_bytes[DDIR_READ])
751 runtime[DDIR_READ] += utime_since_now(&td->start);
752 if (td_write(td) && td->io_bytes[DDIR_WRITE])
753 runtime[DDIR_WRITE] += utime_since_now(&td->start);
755 if (td->error || td->terminate)
758 if (td->verify == VERIFY_NONE)
762 fio_gettime(&td->start, NULL);
766 runtime[DDIR_READ] += utime_since_now(&td->start);
768 if (td->error || td->terminate)
772 update_rusage_stat(td);
773 fio_gettime(&td->end_time, NULL);
774 td->runtime[0] = runtime[0] / 1000;
775 td->runtime[1] = runtime[1] / 1000;
778 finish_log(td, td->ts.bw_log, "bw");
780 finish_log(td, td->ts.slat_log, "slat");
782 finish_log(td, td->ts.clat_log, "clat");
783 if (td->write_iolog_file)
784 write_iolog_close(td);
785 if (td->exec_postrun) {
786 if (system(td->exec_postrun) < 0)
787 log_err("fio: postrun %s failed\n", td->exec_postrun);
790 if (exitall_on_terminate)
791 terminate_threads(td->groupid, 0);
795 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
799 td_set_runstate(td, TD_EXITED);
800 return (void *) (unsigned long) td->error;
804 * We cannot pass the td data into a forked process, so attach the td and
805 * pass it to the thread worker.
807 static int fork_main(int shmid, int offset)
809 struct thread_data *td;
812 data = shmat(shmid, NULL, 0);
813 if (data == (void *) -1) {
820 td = data + offset * sizeof(struct thread_data);
821 ret = thread_main(td);
823 return (int) (unsigned long) ret;
827 * Run over the job map and reap the threads that have exited, if any.
829 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
831 struct thread_data *td;
832 int i, cputhreads, pending, status, ret;
835 * reap exited threads (TD_EXITED -> TD_REAPED)
837 pending = cputhreads = 0;
842 * ->io_ops is NULL for a thread that has closed its
845 if (td->io_ops && td->io_ops->flags & FIO_CPUIO)
848 if (!td->pid || td->runstate == TD_REAPED)
850 if (td->use_thread) {
851 if (td->runstate == TD_EXITED) {
852 td_set_runstate(td, TD_REAPED);
859 if (td->runstate == TD_EXITED)
863 * check if someone quit or got killed in an unusual way
865 ret = waitpid(td->pid, &status, flags);
867 if (errno == ECHILD) {
868 log_err("fio: pid=%d disappeared %d\n", td->pid, td->runstate);
869 td_set_runstate(td, TD_REAPED);
873 } else if (ret == td->pid) {
874 if (WIFSIGNALED(status)) {
875 int sig = WTERMSIG(status);
877 log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
878 td_set_runstate(td, TD_REAPED);
881 if (WIFEXITED(status)) {
882 if (WEXITSTATUS(status) && !td->error)
883 td->error = WEXITSTATUS(status);
885 td_set_runstate(td, TD_REAPED);
891 * thread is not dead, continue
895 if (td->use_thread) {
898 if (pthread_join(td->thread, (void *) &ret))
899 perror("pthread_join");
903 (*m_rate) -= td->ratemin;
904 (*t_rate) -= td->rate;
910 if (*nr_running == cputhreads && !pending)
911 terminate_threads(TERMINATE_ALL, 0);
915 * Main function for kicking off and reaping jobs, as needed.
917 static void run_threads(void)
919 struct thread_data *td;
921 int i, todo, nr_running, m_rate, t_rate, nr_started;
923 if (fio_pin_memory())
927 printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
931 signal(SIGINT, sig_handler);
932 signal(SIGALRM, sig_handler);
934 todo = thread_number;
940 print_status_init(td->thread_number - 1);
942 if (!td->create_serialize) {
948 * do file setup here so it happens sequentially,
949 * we don't want X number of threads getting their
950 * client data interspersed on disk
952 if (setup_files(td)) {
955 log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
956 td_set_runstate(td, TD_REAPED);
966 struct thread_data *map[MAX_JOBS];
967 struct timeval this_start;
968 int this_jobs = 0, left;
971 * create threads (TD_NOT_CREATED -> TD_CREATED)
974 if (td->runstate != TD_NOT_CREATED)
978 * never got a chance to start, killed by other
979 * thread for some reason
986 if (td->start_delay) {
987 spent = mtime_since_genesis();
989 if (td->start_delay * 1000 > spent)
993 if (td->stonewall && (nr_started || nr_running))
997 * Set state to created. Thread will transition
998 * to TD_INITIALIZED when it's done setting up.
1000 td_set_runstate(td, TD_CREATED);
1001 map[this_jobs++] = td;
1002 fio_sem_init(&startup_sem, 1);
1005 if (td->use_thread) {
1006 if (pthread_create(&td->thread, NULL, thread_main, td)) {
1007 perror("thread_create");
1012 fio_sem_down(&startup_sem);
1014 int ret = fork_main(shm_id, i);
1022 * Wait for the started threads to transition to
1025 fio_gettime(&this_start, NULL);
1027 while (left && !fio_abort) {
1028 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1033 for (i = 0; i < this_jobs; i++) {
1037 if (td->runstate == TD_INITIALIZED) {
1040 } else if (td->runstate >= TD_EXITED) {
1044 nr_running++; /* work-around... */
1050 log_err("fio: %d jobs failed to start\n", left);
1051 for (i = 0; i < this_jobs; i++) {
1055 kill(td->pid, SIGTERM);
1061 * start created threads (TD_INITIALIZED -> TD_RUNNING).
1063 for_each_td(td, i) {
1064 if (td->runstate != TD_INITIALIZED)
1067 td_set_runstate(td, TD_RUNNING);
1070 m_rate += td->ratemin;
1073 fio_sem_up(&td->mutex);
1076 reap_threads(&nr_running, &t_rate, &m_rate);
1082 while (nr_running) {
1083 reap_threads(&nr_running, &t_rate, &m_rate);
1091 int main(int argc, char *argv[])
1096 * We need locale for number printing, if it isn't set then just
1097 * go with the US format.
1099 if (!getenv("LC_NUMERIC"))
1100 setlocale(LC_NUMERIC, "en_US");
1102 if (parse_options(argc, argv))
1105 if (!thread_number) {
1106 log_err("Nothing to do\n");
1110 ps = sysconf(_SC_PAGESIZE);
1112 log_err("Failed to get page size\n");
1119 setup_log(&agg_io_log[DDIR_READ]);
1120 setup_log(&agg_io_log[DDIR_WRITE]);
1125 disk_util_timer_arm();
1132 __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1133 __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");