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)
99 unsigned long long bytes = 0;
104 * allow a 2 second settle period in the beginning
106 if (mtime_since(&td->start, now) < 2000)
110 bytes += td->this_io_bytes[DDIR_READ];
112 bytes += td->this_io_bytes[DDIR_WRITE];
115 * if rate blocks is set, sample is running
117 if (td->rate_bytes) {
118 spent = mtime_since(&td->lastrate, now);
119 if (spent < td->ratecycle)
122 if (bytes < td->rate_bytes) {
123 fprintf(f_out, "%s: min rate %u not met\n", td->name, td->ratemin);
126 rate = (bytes - td->rate_bytes) / spent;
127 if (rate < td->ratemin || bytes < td->rate_bytes) {
128 fprintf(f_out, "%s: min rate %u not met, got %luKiB/sec\n", td->name, td->ratemin, rate);
134 td->rate_bytes = bytes;
135 memcpy(&td->lastrate, now, sizeof(*now));
139 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
143 if (mtime_since(&td->epoch, t) >= td->timeout * 1000)
150 * When job exits, we can cancel the in-flight IO if we are using async
151 * io. Attempt to do so.
153 static void cleanup_pending_aio(struct thread_data *td)
155 struct list_head *entry, *n;
160 * get immediately available events, if any
162 r = io_u_queued_complete(td, 0, NULL);
167 * now cancel remaining active events
169 if (td->io_ops->cancel) {
170 list_for_each_safe(entry, n, &td->io_u_busylist) {
171 io_u = list_entry(entry, struct io_u, list);
174 * if the io_u isn't in flight, then that generally
175 * means someone leaked an io_u. complain but fix
176 * it up, so we don't stall here.
178 if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
179 log_err("fio: non-busy IO on busy list\n");
182 r = td->io_ops->cancel(td, io_u);
190 r = io_u_queued_complete(td, td->cur_depth, NULL);
194 * Helper to handle the final sync of a file. Works just like the normal
195 * io path, just does everything sync.
197 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
199 struct io_u *io_u = __get_io_u(td);
205 io_u->ddir = DDIR_SYNC;
208 if (td_io_prep(td, io_u)) {
214 ret = td_io_queue(td, io_u);
216 td_verror(td, io_u->error, "td_io_queue");
219 } else if (ret == FIO_Q_QUEUED) {
220 if (io_u_queued_complete(td, 1, NULL) < 0)
222 } else if (ret == FIO_Q_COMPLETED) {
224 td_verror(td, io_u->error, "td_io_queue");
228 if (io_u_sync_complete(td, io_u, NULL) < 0)
230 } else if (ret == FIO_Q_BUSY) {
231 if (td_io_commit(td))
240 * The main verify engine. Runs over the writes we previusly submitted,
241 * reads the blocks back in, and checks the crc/md5 of the data.
243 static void do_verify(struct thread_data *td)
247 int ret, i, min_events;
250 * sync io first and invalidate cache, to make sure we really
253 for_each_file(td, f, i) {
254 if (fio_io_sync(td, f))
256 if (file_invalidate_cache(td, f))
263 td_set_runstate(td, TD_VERIFYING);
266 while (!td->terminate) {
267 io_u = __get_io_u(td);
271 if (runtime_exceeded(td, &io_u->start_time)) {
276 if (get_next_verify(td, io_u)) {
281 if (td_io_prep(td, io_u)) {
286 ret = td_io_queue(td, io_u);
289 case FIO_Q_COMPLETED:
292 if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
293 int bytes = io_u->xfer_buflen - io_u->resid;
295 io_u->xfer_buflen = io_u->resid;
296 io_u->xfer_buf += bytes;
299 ret = io_u_sync_complete(td, io_u, verify_io_u);
306 requeue_io_u(td, &io_u);
307 ret = td_io_commit(td);
311 td_verror(td, -ret, "td_io_queue");
315 if (ret < 0 || td->error)
319 * if we can queue more, do so. but check if there are
320 * completed io_u's first.
323 if (queue_full(td) || ret == FIO_Q_BUSY) {
326 if (td->cur_depth > td->iodepth_low)
327 min_events = td->cur_depth - td->iodepth_low;
331 * Reap required number of io units, if any, and do the
332 * verification on them through the callback handler
334 if (io_u_queued_complete(td, min_events, verify_io_u) < 0)
339 cleanup_pending_aio(td);
341 td_set_runstate(td, TD_RUNNING);
345 * Not really an io thread, all it does is burn CPU cycles in the specified
348 static void do_cpuio(struct thread_data *td)
351 int split = 100 / td->cpuload;
354 while (!td->terminate) {
355 fio_gettime(&e, NULL);
357 if (runtime_exceeded(td, &e))
363 usec_sleep(td, 10000);
370 * Main IO worker function. It retrieves io_u's to process and queues
371 * and reaps them, checking for rate and errors along the way.
373 static void do_io(struct thread_data *td)
379 td_set_runstate(td, TD_RUNNING);
381 while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->io_size) {
382 struct timeval comp_time;
394 memcpy(&s, &io_u->start_time, sizeof(s));
396 if (runtime_exceeded(td, &s)) {
401 ret = td_io_queue(td, io_u);
404 case FIO_Q_COMPLETED:
409 if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
410 int bytes = io_u->xfer_buflen - io_u->resid;
412 io_u->xfer_buflen = io_u->resid;
413 io_u->xfer_buf += bytes;
416 fio_gettime(&comp_time, NULL);
417 bytes_done = io_u_sync_complete(td, io_u, NULL);
423 * if the engine doesn't have a commit hook,
424 * the io_u is really queued. if it does have such
425 * a hook, it has to call io_u_queued() itself.
427 if (td->io_ops->commit == NULL)
428 io_u_queued(td, io_u);
431 requeue_io_u(td, &io_u);
432 ret = td_io_commit(td);
440 if (ret < 0 || td->error)
444 * See if we need to complete some commands
446 if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
448 if (queue_full(td) || ret == FIO_Q_BUSY) {
451 if (td->cur_depth > td->iodepth_low)
452 min_evts = td->cur_depth - td->iodepth_low;
455 fio_gettime(&comp_time, NULL);
456 bytes_done = io_u_queued_complete(td, min_evts, NULL);
465 * the rate is batched for now, it should work for batches
466 * of completions except the very first one which may look
469 usec = utime_since(&s, &comp_time);
471 rate_throttle(td, usec, bytes_done);
473 if (check_min_rate(td, &comp_time)) {
474 if (exitall_on_terminate)
475 terminate_threads(td->groupid, 0);
476 td_verror(td, ENODATA, "check_min_rate");
481 unsigned long long b;
483 b = td->io_blocks[0] + td->io_blocks[1];
484 if (!(b % td->thinktime_blocks)) {
487 if (td->thinktime_spin)
488 __usec_sleep(td->thinktime_spin);
490 left = td->thinktime - td->thinktime_spin;
492 usec_sleep(td, left);
501 cleanup_pending_aio(td);
503 if (should_fsync(td) && td->end_fsync) {
504 td_set_runstate(td, TD_FSYNCING);
505 for_each_file(td, f, i)
511 static void cleanup_io_u(struct thread_data *td)
513 struct list_head *entry, *n;
516 list_for_each_safe(entry, n, &td->io_u_freelist) {
517 io_u = list_entry(entry, struct io_u, list);
519 list_del(&io_u->list);
527 * "randomly" fill the buffer contents
529 static void fill_rand_buf(struct io_u *io_u, int max_bs)
531 int *ptr = io_u->buf;
533 while ((void *) ptr - io_u->buf < max_bs) {
534 *ptr = rand() * 0x9e370001;
539 static int init_io_u(struct thread_data *td)
546 if (td->io_ops->flags & FIO_CPUIO)
549 if (td->io_ops->flags & FIO_SYNCIO)
552 max_units = td->iodepth;
554 max_bs = max(td->max_bs[DDIR_READ], td->max_bs[DDIR_WRITE]);
555 td->orig_buffer_size = max_bs * max_units;
557 if (td->mem_type == MEM_SHMHUGE || td->mem_type == MEM_MMAPHUGE)
558 td->orig_buffer_size = (td->orig_buffer_size + td->hugepage_size - 1) & ~(td->hugepage_size - 1);
560 td->orig_buffer_size += page_mask;
562 if (allocate_io_mem(td))
565 p = ALIGN(td->orig_buffer);
566 for (i = 0; i < max_units; i++) {
567 io_u = malloc(sizeof(*io_u));
568 memset(io_u, 0, sizeof(*io_u));
569 INIT_LIST_HEAD(&io_u->list);
571 io_u->buf = p + max_bs * i;
572 if (td_write(td) || td_rw(td))
573 fill_rand_buf(io_u, max_bs);
576 io_u->flags = IO_U_F_FREE;
577 list_add(&io_u->list, &td->io_u_freelist);
585 static int switch_ioscheduler(struct thread_data *td)
587 char tmp[256], tmp2[128];
591 if (td->io_ops->flags & FIO_CPUIO)
594 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
596 f = fopen(tmp, "r+");
598 td_verror(td, errno, "fopen");
605 ret = fwrite(td->ioscheduler, strlen(td->ioscheduler), 1, f);
606 if (ferror(f) || ret != 1) {
607 td_verror(td, errno, "fwrite");
615 * Read back and check that the selected scheduler is now the default.
617 ret = fread(tmp, 1, sizeof(tmp), f);
618 if (ferror(f) || ret < 0) {
619 td_verror(td, errno, "fread");
624 sprintf(tmp2, "[%s]", td->ioscheduler);
625 if (!strstr(tmp, tmp2)) {
626 log_err("fio: io scheduler %s not found\n", td->ioscheduler);
627 td_verror(td, EINVAL, "iosched_switch");
636 static void clear_io_state(struct thread_data *td)
641 td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
642 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
645 td->last_was_sync = 0;
647 for_each_file(td, f, i) {
648 f->last_completed_pos = 0;
651 if (td->io_ops->flags & FIO_SYNCIO)
652 lseek(f->fd, SEEK_SET, 0);
655 memset(f->file_map, 0, f->num_maps * sizeof(long));
660 * Entry point for the thread based jobs. The process based jobs end up
661 * here as well, after a little setup.
663 static void *thread_main(void *data)
665 unsigned long long runtime[2];
666 struct thread_data *td = data;
673 INIT_LIST_HEAD(&td->io_u_freelist);
674 INIT_LIST_HEAD(&td->io_u_busylist);
675 INIT_LIST_HEAD(&td->io_u_requeues);
676 INIT_LIST_HEAD(&td->io_hist_list);
677 INIT_LIST_HEAD(&td->io_log_list);
682 if (fio_setaffinity(td) == -1) {
683 td_verror(td, errno, "cpu_set_affinity");
691 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
692 td_verror(td, errno, "ioprio_set");
697 if (nice(td->nice) == -1) {
698 td_verror(td, errno, "nice");
702 if (init_random_state(td))
705 if (td->ioscheduler && switch_ioscheduler(td))
708 td_set_runstate(td, TD_INITIALIZED);
709 fio_sem_up(&startup_sem);
710 fio_sem_down(&td->mutex);
712 if (!td->create_serialize && setup_files(td))
718 * Do this late, as some IO engines would like to have the
719 * files setup prior to initializing structures.
724 if (td->exec_prerun) {
725 if (system(td->exec_prerun) < 0)
729 fio_gettime(&td->epoch, NULL);
730 memcpy(&td->timeout_end, &td->epoch, sizeof(td->epoch));
731 getrusage(RUSAGE_SELF, &td->ts.ru_start);
733 runtime[0] = runtime[1] = 0;
734 while (td->loops--) {
735 fio_gettime(&td->start, NULL);
736 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
739 memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
742 prune_io_piece_log(td);
744 if (td->io_ops->flags & FIO_CPUIO)
749 if (td_read(td) && td->io_bytes[DDIR_READ])
750 runtime[DDIR_READ] += utime_since_now(&td->start);
751 if (td_write(td) && td->io_bytes[DDIR_WRITE])
752 runtime[DDIR_WRITE] += utime_since_now(&td->start);
754 if (td->error || td->terminate)
757 if (td->verify == VERIFY_NONE)
761 fio_gettime(&td->start, NULL);
765 runtime[DDIR_READ] += utime_since_now(&td->start);
767 if (td->error || td->terminate)
771 update_rusage_stat(td);
772 fio_gettime(&td->end_time, NULL);
773 td->runtime[0] = runtime[0] / 1000;
774 td->runtime[1] = runtime[1] / 1000;
777 finish_log(td, td->ts.bw_log, "bw");
779 finish_log(td, td->ts.slat_log, "slat");
781 finish_log(td, td->ts.clat_log, "clat");
782 if (td->write_iolog_file)
783 write_iolog_close(td);
784 if (td->exec_postrun) {
785 if (system(td->exec_postrun) < 0)
786 log_err("fio: postrun %s failed\n", td->exec_postrun);
789 if (exitall_on_terminate)
790 terminate_threads(td->groupid, 0);
794 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
798 td_set_runstate(td, TD_EXITED);
799 return (void *) (unsigned long) td->error;
803 * We cannot pass the td data into a forked process, so attach the td and
804 * pass it to the thread worker.
806 static int fork_main(int shmid, int offset)
808 struct thread_data *td;
811 data = shmat(shmid, NULL, 0);
812 if (data == (void *) -1) {
819 td = data + offset * sizeof(struct thread_data);
820 ret = thread_main(td);
822 return (int) (unsigned long) ret;
826 * Run over the job map and reap the threads that have exited, if any.
828 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
830 struct thread_data *td;
831 int i, cputhreads, pending, status, ret;
834 * reap exited threads (TD_EXITED -> TD_REAPED)
836 pending = cputhreads = 0;
841 * ->io_ops is NULL for a thread that has closed its
844 if (td->io_ops && td->io_ops->flags & FIO_CPUIO)
847 if (!td->pid || td->runstate == TD_REAPED)
849 if (td->use_thread) {
850 if (td->runstate == TD_EXITED) {
851 td_set_runstate(td, TD_REAPED);
858 if (td->runstate == TD_EXITED)
862 * check if someone quit or got killed in an unusual way
864 ret = waitpid(td->pid, &status, flags);
866 if (errno == ECHILD) {
867 log_err("fio: pid=%d disappeared %d\n", td->pid, td->runstate);
868 td_set_runstate(td, TD_REAPED);
872 } else if (ret == td->pid) {
873 if (WIFSIGNALED(status)) {
874 int sig = WTERMSIG(status);
876 log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
877 td_set_runstate(td, TD_REAPED);
880 if (WIFEXITED(status)) {
881 if (WEXITSTATUS(status) && !td->error)
882 td->error = WEXITSTATUS(status);
884 td_set_runstate(td, TD_REAPED);
890 * thread is not dead, continue
894 if (td->use_thread) {
897 if (pthread_join(td->thread, (void *) &ret))
898 perror("pthread_join");
902 (*m_rate) -= td->ratemin;
903 (*t_rate) -= td->rate;
909 if (*nr_running == cputhreads && !pending)
910 terminate_threads(TERMINATE_ALL, 0);
914 * Main function for kicking off and reaping jobs, as needed.
916 static void run_threads(void)
918 struct thread_data *td;
920 int i, todo, nr_running, m_rate, t_rate, nr_started;
922 if (fio_pin_memory())
926 printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
930 signal(SIGINT, sig_handler);
931 signal(SIGALRM, sig_handler);
933 todo = thread_number;
939 print_status_init(td->thread_number - 1);
941 if (!td->create_serialize) {
947 * do file setup here so it happens sequentially,
948 * we don't want X number of threads getting their
949 * client data interspersed on disk
951 if (setup_files(td)) {
954 log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
955 td_set_runstate(td, TD_REAPED);
965 struct thread_data *map[MAX_JOBS];
966 struct timeval this_start;
967 int this_jobs = 0, left;
970 * create threads (TD_NOT_CREATED -> TD_CREATED)
973 if (td->runstate != TD_NOT_CREATED)
977 * never got a chance to start, killed by other
978 * thread for some reason
985 if (td->start_delay) {
986 spent = mtime_since_genesis();
988 if (td->start_delay * 1000 > spent)
992 if (td->stonewall && (nr_started || nr_running))
996 * Set state to created. Thread will transition
997 * to TD_INITIALIZED when it's done setting up.
999 td_set_runstate(td, TD_CREATED);
1000 map[this_jobs++] = td;
1001 fio_sem_init(&startup_sem, 1);
1004 if (td->use_thread) {
1005 if (pthread_create(&td->thread, NULL, thread_main, td)) {
1006 perror("thread_create");
1011 fio_sem_down(&startup_sem);
1013 int ret = fork_main(shm_id, i);
1021 * Wait for the started threads to transition to
1024 fio_gettime(&this_start, NULL);
1026 while (left && !fio_abort) {
1027 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1032 for (i = 0; i < this_jobs; i++) {
1036 if (td->runstate == TD_INITIALIZED) {
1039 } else if (td->runstate >= TD_EXITED) {
1043 nr_running++; /* work-around... */
1049 log_err("fio: %d jobs failed to start\n", left);
1050 for (i = 0; i < this_jobs; i++) {
1054 kill(td->pid, SIGTERM);
1060 * start created threads (TD_INITIALIZED -> TD_RUNNING).
1062 for_each_td(td, i) {
1063 if (td->runstate != TD_INITIALIZED)
1066 td_set_runstate(td, TD_RUNNING);
1069 m_rate += td->ratemin;
1072 fio_sem_up(&td->mutex);
1075 reap_threads(&nr_running, &t_rate, &m_rate);
1081 while (nr_running) {
1082 reap_threads(&nr_running, &t_rate, &m_rate);
1090 int main(int argc, char *argv[])
1095 * We need locale for number printing, if it isn't set then just
1096 * go with the US format.
1098 if (!getenv("LC_NUMERIC"))
1099 setlocale(LC_NUMERIC, "en_US");
1101 if (parse_options(argc, argv))
1104 if (!thread_number) {
1105 log_err("Nothing to do\n");
1109 ps = sysconf(_SC_PAGESIZE);
1111 log_err("Failed to get page size\n");
1118 setup_log(&agg_io_log[DDIR_READ]);
1119 setup_log(&agg_io_log[DDIR_WRITE]);
1124 disk_util_timer_arm();
1131 __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1132 __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");