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 timespec ts = { .tv_sec = 0, .tv_nsec = 0};
167 struct list_head *entry, *n;
168 struct io_completion_data icd;
173 * get immediately available events, if any
175 r = td_io_getevents(td, 0, td->cur_depth, &ts);
177 init_icd(&icd, NULL, r);
178 ios_completed(td, &icd);
182 * now cancel remaining active events
184 if (td->io_ops->cancel) {
185 list_for_each_safe(entry, n, &td->io_u_busylist) {
186 io_u = list_entry(entry, struct io_u, list);
188 r = td->io_ops->cancel(td, io_u);
195 r = td_io_getevents(td, td->cur_depth, td->cur_depth, NULL);
197 init_icd(&icd, NULL, r);
198 ios_completed(td, &icd);
204 * Helper to handle the final sync of a file. Works just like the normal
205 * io path, just does everything sync.
207 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
209 struct io_u *io_u = __get_io_u(td);
210 struct io_completion_data icd;
216 io_u->ddir = DDIR_SYNC;
219 if (td_io_prep(td, io_u)) {
224 ret = td_io_queue(td, io_u);
226 td_verror(td, io_u->error);
229 } else if (ret == FIO_Q_QUEUED) {
230 ret = td_io_getevents(td, 1, td->cur_depth, NULL);
236 init_icd(&icd, NULL, ret);
237 ios_completed(td, &icd);
239 td_verror(td, icd.error);
242 } else if (ret == FIO_Q_COMPLETED) {
244 td_verror(td, io_u->error);
248 init_icd(&icd, NULL, 1);
249 io_completed(td, io_u, &icd);
257 * The main verify engine. Runs over the writes we previusly submitted,
258 * reads the blocks back in, and checks the crc/md5 of the data.
260 static void do_verify(struct thread_data *td)
264 int ret, i, min_events;
267 * sync io first and invalidate cache, to make sure we really
270 for_each_file(td, f, i) {
272 file_invalidate_cache(td, f);
275 td_set_runstate(td, TD_VERIFYING);
278 while (!td->terminate) {
279 struct io_completion_data icd;
280 struct timespec *timeout;
282 io_u = __get_io_u(td);
286 if (runtime_exceeded(td, &io_u->start_time))
289 if (get_next_verify(td, io_u))
292 if (td_io_prep(td, io_u))
296 ret = td_io_queue(td, io_u);
299 case FIO_Q_COMPLETED:
302 if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
303 int bytes = io_u->xfer_buflen - io_u->resid;
305 io_u->xfer_buflen = io_u->resid;
306 io_u->xfer_buf += bytes;
309 init_icd(&icd, verify_io_u, 1);
310 io_completed(td, io_u, &icd);
329 * if we can queue more, do so. but check if there are
330 * completed io_u's first.
332 if (queue_full(td)) {
345 * Reap required number of io units, if any, and do the
346 * verification on them through the callback handler
348 ret = td_io_getevents(td, min_events, td->cur_depth, timeout);
354 init_icd(&icd, verify_io_u, ret);
355 ios_completed(td, &icd);
358 td_verror(td, icd.error);
367 cleanup_pending_aio(td);
369 td_set_runstate(td, TD_RUNNING);
373 * Not really an io thread, all it does is burn CPU cycles in the specified
376 static void do_cpuio(struct thread_data *td)
379 int split = 100 / td->cpuload;
382 while (!td->terminate) {
383 fio_gettime(&e, NULL);
385 if (runtime_exceeded(td, &e))
391 usec_sleep(td, 10000);
398 * Main IO worker function. It retrieves io_u's to process and queues
399 * and reaps them, checking for rate and errors along the way.
401 static void do_io(struct thread_data *td)
403 struct io_completion_data icd;
409 td_set_runstate(td, TD_RUNNING);
411 while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->io_size) {
412 struct timespec *timeout;
419 f = get_next_file(td);
423 io_u = get_io_u(td, f);
427 memcpy(&s, &io_u->start_time, sizeof(s));
429 ret = td_io_queue(td, io_u);
432 case FIO_Q_COMPLETED:
437 if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
438 int bytes = io_u->xfer_buflen - io_u->resid;
440 io_u->xfer_buflen = io_u->resid;
441 io_u->xfer_buf += bytes;
444 init_icd(&icd, NULL, 1);
445 io_completed(td, io_u, &icd);
459 add_slat_sample(td, io_u->ddir, mtime_since(&io_u->start_time, &io_u->issue_time));
461 if (ret == FIO_Q_QUEUED) {
462 if (td->cur_depth < td->iodepth) {
474 ret = td_io_getevents(td, min_evts, td->cur_depth, timeout);
481 init_icd(&icd, NULL, ret);
482 ios_completed(td, &icd);
484 td_verror(td, icd.error);
490 * the rate is batched for now, it should work for batches
491 * of completions except the very first one which may look
494 usec = utime_since(&s, &icd.time);
496 rate_throttle(td, usec, icd.bytes_done[td->ddir], td->ddir);
498 if (check_min_rate(td, &icd.time)) {
499 if (exitall_on_terminate)
500 terminate_threads(td->groupid, 0);
501 td_verror(td, ENODATA);
505 if (runtime_exceeded(td, &icd.time))
509 unsigned long long b;
511 b = td->io_blocks[0] + td->io_blocks[1];
512 if (!(b % td->thinktime_blocks)) {
515 if (td->thinktime_spin)
516 __usec_sleep(td->thinktime_spin);
518 left = td->thinktime - td->thinktime_spin;
520 usec_sleep(td, left);
527 cleanup_pending_aio(td);
529 if (should_fsync(td) && td->end_fsync) {
530 td_set_runstate(td, TD_FSYNCING);
531 for_each_file(td, f, i)
537 static void cleanup_io_u(struct thread_data *td)
539 struct list_head *entry, *n;
542 list_for_each_safe(entry, n, &td->io_u_freelist) {
543 io_u = list_entry(entry, struct io_u, list);
545 list_del(&io_u->list);
553 * "randomly" fill the buffer contents
555 static void fill_rand_buf(struct io_u *io_u, int max_bs)
557 int *ptr = io_u->buf;
559 while ((void *) ptr - io_u->buf < max_bs) {
560 *ptr = rand() * 0x9e370001;
565 static int init_io_u(struct thread_data *td)
572 if (td->io_ops->flags & FIO_CPUIO)
575 if (td->io_ops->flags & FIO_SYNCIO)
578 max_units = td->iodepth;
580 max_bs = max(td->max_bs[DDIR_READ], td->max_bs[DDIR_WRITE]);
581 td->orig_buffer_size = max_bs * max_units;
583 if (td->mem_type == MEM_SHMHUGE || td->mem_type == MEM_MMAPHUGE)
584 td->orig_buffer_size = (td->orig_buffer_size + td->hugepage_size - 1) & ~(td->hugepage_size - 1);
586 td->orig_buffer_size += page_mask;
588 if (allocate_io_mem(td))
591 p = ALIGN(td->orig_buffer);
592 for (i = 0; i < max_units; i++) {
593 io_u = malloc(sizeof(*io_u));
594 memset(io_u, 0, sizeof(*io_u));
595 INIT_LIST_HEAD(&io_u->list);
597 io_u->buf = p + max_bs * i;
598 if (td_write(td) || td_rw(td))
599 fill_rand_buf(io_u, max_bs);
602 list_add(&io_u->list, &td->io_u_freelist);
608 static int switch_ioscheduler(struct thread_data *td)
610 char tmp[256], tmp2[128];
614 if (td->io_ops->flags & FIO_CPUIO)
617 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
619 f = fopen(tmp, "r+");
621 td_verror(td, errno);
628 ret = fwrite(td->ioscheduler, strlen(td->ioscheduler), 1, f);
629 if (ferror(f) || ret != 1) {
630 td_verror(td, errno);
638 * Read back and check that the selected scheduler is now the default.
640 ret = fread(tmp, 1, sizeof(tmp), f);
641 if (ferror(f) || ret < 0) {
642 td_verror(td, errno);
647 sprintf(tmp2, "[%s]", td->ioscheduler);
648 if (!strstr(tmp, tmp2)) {
649 log_err("fio: io scheduler %s not found\n", td->ioscheduler);
650 td_verror(td, EINVAL);
659 static void clear_io_state(struct thread_data *td)
664 td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0;
665 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
668 for_each_file(td, f, i) {
670 if (td->io_ops->flags & FIO_SYNCIO)
671 lseek(f->fd, SEEK_SET, 0);
674 memset(f->file_map, 0, f->num_maps * sizeof(long));
679 * Entry point for the thread based jobs. The process based jobs end up
680 * here as well, after a little setup.
682 static void *thread_main(void *data)
684 unsigned long long runtime[2];
685 struct thread_data *td = data;
692 INIT_LIST_HEAD(&td->io_u_freelist);
693 INIT_LIST_HEAD(&td->io_u_busylist);
694 INIT_LIST_HEAD(&td->io_hist_list);
695 INIT_LIST_HEAD(&td->io_log_list);
700 if (fio_setaffinity(td) == -1) {
701 td_verror(td, errno);
709 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
710 td_verror(td, errno);
715 if (nice(td->nice) == -1) {
716 td_verror(td, errno);
720 if (init_random_state(td))
723 if (td->ioscheduler && switch_ioscheduler(td))
726 td_set_runstate(td, TD_INITIALIZED);
727 fio_sem_up(&startup_sem);
728 fio_sem_down(&td->mutex);
730 if (!td->create_serialize && setup_files(td))
736 * Do this late, as some IO engines would like to have the
737 * files setup prior to initializing structures.
742 if (td->exec_prerun) {
743 if (system(td->exec_prerun) < 0)
747 fio_gettime(&td->epoch, NULL);
748 getrusage(RUSAGE_SELF, &td->ru_start);
750 runtime[0] = runtime[1] = 0;
751 while (td->loops--) {
752 fio_gettime(&td->start, NULL);
753 memcpy(&td->stat_sample_time, &td->start, sizeof(td->start));
756 memcpy(&td->lastrate, &td->stat_sample_time, sizeof(td->lastrate));
759 prune_io_piece_log(td);
761 if (td->io_ops->flags & FIO_CPUIO)
766 runtime[td->ddir] += utime_since_now(&td->start);
767 if (td_rw(td) && td->io_bytes[td->ddir ^ 1])
768 runtime[td->ddir ^ 1] = runtime[td->ddir];
770 if (td->error || td->terminate)
773 if (td->verify == VERIFY_NONE)
777 fio_gettime(&td->start, NULL);
781 runtime[DDIR_READ] += utime_since_now(&td->start);
783 if (td->error || td->terminate)
787 update_rusage_stat(td);
788 fio_gettime(&td->end_time, NULL);
789 td->runtime[0] = runtime[0] / 1000;
790 td->runtime[1] = runtime[1] / 1000;
793 finish_log(td, td->bw_log, "bw");
795 finish_log(td, td->slat_log, "slat");
797 finish_log(td, td->clat_log, "clat");
798 if (td->write_iolog_file)
799 write_iolog_close(td);
800 if (td->exec_postrun) {
801 if (system(td->exec_postrun) < 0)
802 log_err("fio: postrun %s failed\n", td->exec_postrun);
805 if (exitall_on_terminate)
806 terminate_threads(td->groupid, 0);
810 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
814 td_set_runstate(td, TD_EXITED);
815 return (void *) td->error;
819 * We cannot pass the td data into a forked process, so attach the td and
820 * pass it to the thread worker.
822 static int fork_main(int shmid, int offset)
824 struct thread_data *td;
827 data = shmat(shmid, NULL, 0);
828 if (data == (void *) -1) {
835 td = data + offset * sizeof(struct thread_data);
836 ret = thread_main(td);
842 * Run over the job map and reap the threads that have exited, if any.
844 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
846 struct thread_data *td;
847 int i, cputhreads, pending, status, ret;
850 * reap exited threads (TD_EXITED -> TD_REAPED)
852 pending = cputhreads = 0;
855 * ->io_ops is NULL for a thread that has closed its
858 if (td->io_ops && td->io_ops->flags & FIO_CPUIO)
861 if (td->runstate < TD_EXITED) {
863 * check if someone quit or got killed in an unusual way
865 ret = waitpid(td->pid, &status, WNOHANG);
868 else if ((ret == td->pid) && WIFSIGNALED(status)) {
869 int sig = WTERMSIG(status);
871 log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
872 td_set_runstate(td, TD_REAPED);
877 if (td->runstate != TD_EXITED) {
878 if (td->runstate < TD_RUNNING)
887 td_set_runstate(td, TD_REAPED);
889 if (td->use_thread) {
892 if (pthread_join(td->thread, (void *) &ret))
893 perror("thread_join");
897 ret = waitpid(td->pid, &status, 0);
900 else if (WIFEXITED(status) && WEXITSTATUS(status)) {
908 (*m_rate) -= td->ratemin;
909 (*t_rate) -= td->rate;
912 if (*nr_running == cputhreads && !pending)
913 terminate_threads(TERMINATE_ALL, 0);
917 * Main function for kicking off and reaping jobs, as needed.
919 static void run_threads(void)
921 struct thread_data *td;
923 int i, todo, nr_running, m_rate, t_rate, nr_started;
925 if (fio_pin_memory())
929 printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
933 signal(SIGINT, sig_handler);
934 signal(SIGALRM, sig_handler);
936 todo = thread_number;
942 print_status_init(td->thread_number - 1);
944 if (!td->create_serialize) {
950 * do file setup here so it happens sequentially,
951 * we don't want X number of threads getting their
952 * client data interspersed on disk
954 if (setup_files(td)) {
957 log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
958 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]);
1123 disk_util_timer_arm();
1130 __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1131 __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");