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)
104 * allow a 2 second settle period in the beginning
106 if (mtime_since(&td->start, now) < 2000)
110 * if rate blocks is set, sample is running
112 if (td->rate_bytes) {
113 spent = mtime_since(&td->lastrate, now);
114 if (spent < td->ratecycle)
117 rate = (td->this_io_bytes[ddir] - td->rate_bytes) / spent;
118 if (rate < td->ratemin) {
119 fprintf(f_out, "%s: min rate %u not met, got %luKiB/sec\n", td->name, td->ratemin, rate);
124 td->rate_bytes = td->this_io_bytes[ddir];
125 memcpy(&td->lastrate, now, sizeof(*now));
129 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
133 if (mtime_since(&td->epoch, t) >= td->timeout * 1000)
139 static struct fio_file *get_next_file(struct thread_data *td)
141 unsigned int old_next_file = td->next_file;
145 f = &td->files[td->next_file];
148 if (td->next_file >= td->nr_files)
155 } while (td->next_file != old_next_file);
161 * When job exits, we can cancel the in-flight IO if we are using async
162 * io. Attempt to do so.
164 static void cleanup_pending_aio(struct thread_data *td)
166 struct list_head *entry, *n;
171 * get immediately available events, if any
173 io_u_queued_complete(td, 0, NULL);
176 * now cancel remaining active events
178 if (td->io_ops->cancel) {
179 list_for_each_safe(entry, n, &td->io_u_busylist) {
180 io_u = list_entry(entry, struct io_u, list);
182 r = td->io_ops->cancel(td, io_u);
189 io_u_queued_complete(td, td->cur_depth, NULL);
193 * Helper to handle the final sync of a file. Works just like the normal
194 * io path, just does everything sync.
196 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
198 struct io_u *io_u = __get_io_u(td);
204 io_u->ddir = DDIR_SYNC;
207 if (td_io_prep(td, io_u)) {
213 ret = td_io_queue(td, io_u);
215 td_verror(td, io_u->error);
218 } else if (ret == FIO_Q_QUEUED) {
219 if (io_u_queued_complete(td, 1, NULL))
221 } else if (ret == FIO_Q_COMPLETED) {
223 td_verror(td, io_u->error);
227 io_u_sync_complete(td, io_u, NULL);
228 } else if (ret == FIO_Q_BUSY) {
229 if (td_io_commit(td))
238 * The main verify engine. Runs over the writes we previusly submitted,
239 * reads the blocks back in, and checks the crc/md5 of the data.
241 static void do_verify(struct thread_data *td)
245 int ret, i, min_events;
248 * sync io first and invalidate cache, to make sure we really
251 for_each_file(td, f, i) {
253 file_invalidate_cache(td, f);
256 td_set_runstate(td, TD_VERIFYING);
259 while (!td->terminate) {
260 io_u = __get_io_u(td);
264 if (runtime_exceeded(td, &io_u->start_time))
267 if (get_next_verify(td, io_u))
270 if (td_io_prep(td, io_u))
274 ret = td_io_queue(td, io_u);
277 case FIO_Q_COMPLETED:
280 if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
281 int bytes = io_u->xfer_buflen - io_u->resid;
283 io_u->xfer_buflen = io_u->resid;
284 io_u->xfer_buf += bytes;
287 ret = io_u_sync_complete(td, io_u, verify_io_u);
294 requeue_io_u(td, &io_u);
295 ret = td_io_commit(td);
303 if (ret < 0 || td->error)
307 * if we can queue more, do so. but check if there are
308 * completed io_u's first.
311 if (queue_full(td) || ret == FIO_Q_BUSY)
315 * Reap required number of io units, if any, and do the
316 * verification on them through the callback handler
318 if (io_u_queued_complete(td, min_events, verify_io_u))
326 cleanup_pending_aio(td);
328 td_set_runstate(td, TD_RUNNING);
332 * Not really an io thread, all it does is burn CPU cycles in the specified
335 static void do_cpuio(struct thread_data *td)
338 int split = 100 / td->cpuload;
341 while (!td->terminate) {
342 fio_gettime(&e, NULL);
344 if (runtime_exceeded(td, &e))
350 usec_sleep(td, 10000);
357 * Main IO worker function. It retrieves io_u's to process and queues
358 * and reaps them, checking for rate and errors along the way.
360 static void do_io(struct thread_data *td)
367 td_set_runstate(td, TD_RUNNING);
369 while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->io_size) {
370 struct timeval comp_time;
378 f = get_next_file(td);
382 io_u = get_io_u(td, f);
386 memcpy(&s, &io_u->start_time, sizeof(s));
388 if (runtime_exceeded(td, &s)) {
393 ret = td_io_queue(td, io_u);
396 case FIO_Q_COMPLETED:
401 if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
402 int bytes = io_u->xfer_buflen - io_u->resid;
404 io_u->xfer_buflen = io_u->resid;
405 io_u->xfer_buf += bytes;
408 fio_gettime(&comp_time, NULL);
409 bytes_done = io_u_sync_complete(td, io_u, NULL);
416 requeue_io_u(td, &io_u);
417 ret = td_io_commit(td);
425 if (ret < 0 || td->error)
429 add_slat_sample(td, io_u->ddir, mtime_since(&io_u->start_time, &io_u->issue_time));
432 * See if we need to complete some commands
434 if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
436 if (queue_full(td) || ret == FIO_Q_BUSY)
439 fio_gettime(&comp_time, NULL);
440 bytes_done = io_u_queued_complete(td, min_evts, NULL);
449 * the rate is batched for now, it should work for batches
450 * of completions except the very first one which may look
453 usec = utime_since(&s, &comp_time);
455 rate_throttle(td, usec, bytes_done, td->ddir);
457 if (check_min_rate(td, &comp_time)) {
458 if (exitall_on_terminate)
459 terminate_threads(td->groupid, 0);
460 td_verror(td, ENODATA);
465 unsigned long long b;
467 b = td->io_blocks[0] + td->io_blocks[1];
468 if (!(b % td->thinktime_blocks)) {
471 if (td->thinktime_spin)
472 __usec_sleep(td->thinktime_spin);
474 left = td->thinktime - td->thinktime_spin;
476 usec_sleep(td, left);
483 cleanup_pending_aio(td);
485 if (should_fsync(td) && td->end_fsync) {
486 td_set_runstate(td, TD_FSYNCING);
487 for_each_file(td, f, i)
493 static void cleanup_io_u(struct thread_data *td)
495 struct list_head *entry, *n;
498 list_for_each_safe(entry, n, &td->io_u_freelist) {
499 io_u = list_entry(entry, struct io_u, list);
501 list_del(&io_u->list);
509 * "randomly" fill the buffer contents
511 static void fill_rand_buf(struct io_u *io_u, int max_bs)
513 int *ptr = io_u->buf;
515 while ((void *) ptr - io_u->buf < max_bs) {
516 *ptr = rand() * 0x9e370001;
521 static int init_io_u(struct thread_data *td)
528 if (td->io_ops->flags & FIO_CPUIO)
531 if (td->io_ops->flags & FIO_SYNCIO)
534 max_units = td->iodepth;
536 max_bs = max(td->max_bs[DDIR_READ], td->max_bs[DDIR_WRITE]);
537 td->orig_buffer_size = max_bs * max_units;
539 if (td->mem_type == MEM_SHMHUGE || td->mem_type == MEM_MMAPHUGE)
540 td->orig_buffer_size = (td->orig_buffer_size + td->hugepage_size - 1) & ~(td->hugepage_size - 1);
542 td->orig_buffer_size += page_mask;
544 if (allocate_io_mem(td))
547 p = ALIGN(td->orig_buffer);
548 for (i = 0; i < max_units; i++) {
549 io_u = malloc(sizeof(*io_u));
550 memset(io_u, 0, sizeof(*io_u));
551 INIT_LIST_HEAD(&io_u->list);
553 io_u->buf = p + max_bs * i;
554 if (td_write(td) || td_rw(td))
555 fill_rand_buf(io_u, max_bs);
558 list_add(&io_u->list, &td->io_u_freelist);
564 static int switch_ioscheduler(struct thread_data *td)
566 char tmp[256], tmp2[128];
570 if (td->io_ops->flags & FIO_CPUIO)
573 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
575 f = fopen(tmp, "r+");
577 td_verror(td, errno);
584 ret = fwrite(td->ioscheduler, strlen(td->ioscheduler), 1, f);
585 if (ferror(f) || ret != 1) {
586 td_verror(td, errno);
594 * Read back and check that the selected scheduler is now the default.
596 ret = fread(tmp, 1, sizeof(tmp), f);
597 if (ferror(f) || ret < 0) {
598 td_verror(td, errno);
603 sprintf(tmp2, "[%s]", td->ioscheduler);
604 if (!strstr(tmp, tmp2)) {
605 log_err("fio: io scheduler %s not found\n", td->ioscheduler);
606 td_verror(td, EINVAL);
615 static void clear_io_state(struct thread_data *td)
620 td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0;
621 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
624 for_each_file(td, f, i) {
626 if (td->io_ops->flags & FIO_SYNCIO)
627 lseek(f->fd, SEEK_SET, 0);
630 memset(f->file_map, 0, f->num_maps * sizeof(long));
635 * Entry point for the thread based jobs. The process based jobs end up
636 * here as well, after a little setup.
638 static void *thread_main(void *data)
640 unsigned long long runtime[2];
641 struct thread_data *td = data;
648 INIT_LIST_HEAD(&td->io_u_freelist);
649 INIT_LIST_HEAD(&td->io_u_busylist);
650 INIT_LIST_HEAD(&td->io_u_requeues);
651 INIT_LIST_HEAD(&td->io_hist_list);
652 INIT_LIST_HEAD(&td->io_log_list);
657 if (fio_setaffinity(td) == -1) {
658 td_verror(td, errno);
666 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
667 td_verror(td, errno);
672 if (nice(td->nice) == -1) {
673 td_verror(td, errno);
677 if (init_random_state(td))
680 if (td->ioscheduler && switch_ioscheduler(td))
683 td_set_runstate(td, TD_INITIALIZED);
684 fio_sem_up(&startup_sem);
685 fio_sem_down(&td->mutex);
687 if (!td->create_serialize && setup_files(td))
693 * Do this late, as some IO engines would like to have the
694 * files setup prior to initializing structures.
699 if (td->exec_prerun) {
700 if (system(td->exec_prerun) < 0)
704 fio_gettime(&td->epoch, NULL);
705 getrusage(RUSAGE_SELF, &td->ru_start);
707 runtime[0] = runtime[1] = 0;
708 while (td->loops--) {
709 fio_gettime(&td->start, NULL);
710 memcpy(&td->stat_sample_time, &td->start, sizeof(td->start));
713 memcpy(&td->lastrate, &td->stat_sample_time, sizeof(td->lastrate));
716 prune_io_piece_log(td);
718 if (td->io_ops->flags & FIO_CPUIO)
723 runtime[td->ddir] += utime_since_now(&td->start);
724 if (td_rw(td) && td->io_bytes[td->ddir ^ 1])
725 runtime[td->ddir ^ 1] = runtime[td->ddir];
727 if (td->error || td->terminate)
730 if (td->verify == VERIFY_NONE)
734 fio_gettime(&td->start, NULL);
738 runtime[DDIR_READ] += utime_since_now(&td->start);
740 if (td->error || td->terminate)
744 update_rusage_stat(td);
745 fio_gettime(&td->end_time, NULL);
746 td->runtime[0] = runtime[0] / 1000;
747 td->runtime[1] = runtime[1] / 1000;
750 finish_log(td, td->bw_log, "bw");
752 finish_log(td, td->slat_log, "slat");
754 finish_log(td, td->clat_log, "clat");
755 if (td->write_iolog_file)
756 write_iolog_close(td);
757 if (td->exec_postrun) {
758 if (system(td->exec_postrun) < 0)
759 log_err("fio: postrun %s failed\n", td->exec_postrun);
762 if (exitall_on_terminate)
763 terminate_threads(td->groupid, 0);
767 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
771 td_set_runstate(td, TD_EXITED);
772 return (void *) td->error;
776 * We cannot pass the td data into a forked process, so attach the td and
777 * pass it to the thread worker.
779 static int fork_main(int shmid, int offset)
781 struct thread_data *td;
784 data = shmat(shmid, NULL, 0);
785 if (data == (void *) -1) {
792 td = data + offset * sizeof(struct thread_data);
793 ret = thread_main(td);
799 * Run over the job map and reap the threads that have exited, if any.
801 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
803 struct thread_data *td;
804 int i, cputhreads, pending, status, ret;
807 * reap exited threads (TD_EXITED -> TD_REAPED)
809 pending = cputhreads = 0;
812 * ->io_ops is NULL for a thread that has closed its
815 if (td->io_ops && td->io_ops->flags & FIO_CPUIO)
818 if (td->runstate < TD_EXITED) {
820 * check if someone quit or got killed in an unusual way
822 ret = waitpid(td->pid, &status, WNOHANG);
825 else if ((ret == td->pid) && WIFSIGNALED(status)) {
826 int sig = WTERMSIG(status);
828 log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
829 td_set_runstate(td, TD_REAPED);
834 if (td->runstate != TD_EXITED) {
835 if (td->runstate < TD_RUNNING)
844 td_set_runstate(td, TD_REAPED);
846 if (td->use_thread) {
849 if (pthread_join(td->thread, (void *) &ret))
850 perror("thread_join");
854 ret = waitpid(td->pid, &status, 0);
857 else if (WIFEXITED(status) && WEXITSTATUS(status)) {
865 (*m_rate) -= td->ratemin;
866 (*t_rate) -= td->rate;
869 if (*nr_running == cputhreads && !pending)
870 terminate_threads(TERMINATE_ALL, 0);
874 * Main function for kicking off and reaping jobs, as needed.
876 static void run_threads(void)
878 struct thread_data *td;
880 int i, todo, nr_running, m_rate, t_rate, nr_started;
882 if (fio_pin_memory())
886 printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
890 signal(SIGINT, sig_handler);
891 signal(SIGALRM, sig_handler);
893 todo = thread_number;
899 print_status_init(td->thread_number - 1);
901 if (!td->create_serialize) {
907 * do file setup here so it happens sequentially,
908 * we don't want X number of threads getting their
909 * client data interspersed on disk
911 if (setup_files(td)) {
914 log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
915 td_set_runstate(td, TD_REAPED);
923 struct thread_data *map[MAX_JOBS];
924 struct timeval this_start;
925 int this_jobs = 0, left;
928 * create threads (TD_NOT_CREATED -> TD_CREATED)
931 if (td->runstate != TD_NOT_CREATED)
935 * never got a chance to start, killed by other
936 * thread for some reason
943 if (td->start_delay) {
944 spent = mtime_since_genesis();
946 if (td->start_delay * 1000 > spent)
950 if (td->stonewall && (nr_started || nr_running))
954 * Set state to created. Thread will transition
955 * to TD_INITIALIZED when it's done setting up.
957 td_set_runstate(td, TD_CREATED);
958 map[this_jobs++] = td;
959 fio_sem_init(&startup_sem, 1);
962 if (td->use_thread) {
963 if (pthread_create(&td->thread, NULL, thread_main, td)) {
964 perror("thread_create");
969 fio_sem_down(&startup_sem);
971 int ret = fork_main(shm_id, i);
979 * Wait for the started threads to transition to
982 fio_gettime(&this_start, NULL);
984 while (left && !fio_abort) {
985 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
990 for (i = 0; i < this_jobs; i++) {
994 if (td->runstate == TD_INITIALIZED) {
997 } else if (td->runstate >= TD_EXITED) {
1001 nr_running++; /* work-around... */
1007 log_err("fio: %d jobs failed to start\n", left);
1008 for (i = 0; i < this_jobs; i++) {
1012 kill(td->pid, SIGTERM);
1018 * start created threads (TD_INITIALIZED -> TD_RUNNING).
1020 for_each_td(td, i) {
1021 if (td->runstate != TD_INITIALIZED)
1024 td_set_runstate(td, TD_RUNNING);
1027 m_rate += td->ratemin;
1030 fio_sem_up(&td->mutex);
1033 reap_threads(&nr_running, &t_rate, &m_rate);
1039 while (nr_running) {
1040 reap_threads(&nr_running, &t_rate, &m_rate);
1048 int main(int argc, char *argv[])
1053 * We need locale for number printing, if it isn't set then just
1054 * go with the US format.
1056 if (!getenv("LC_NUMERIC"))
1057 setlocale(LC_NUMERIC, "en_US");
1059 if (parse_options(argc, argv))
1062 if (!thread_number) {
1063 log_err("Nothing to do\n");
1067 ps = sysconf(_SC_PAGESIZE);
1069 log_err("Failed to get page size\n");
1076 setup_log(&agg_io_log[DDIR_READ]);
1077 setup_log(&agg_io_log[DDIR_WRITE]);
1080 disk_util_timer_arm();
1087 __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1088 __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");