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
34 #include <sys/ioctl.h>
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;
52 struct io_log *agg_io_log[2];
54 #define TERMINATE_ALL (-1)
55 #define JOB_START_TIMEOUT (5 * 1000)
57 static inline void td_set_runstate(struct thread_data *td, int runstate)
59 td->runstate = runstate;
62 static void terminate_threads(int group_id, int forced_kill)
64 struct thread_data *td;
68 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
72 td_set_runstate(td, TD_EXITED);
77 static void sig_handler(int sig)
82 disk_util_timer_arm();
83 print_thread_status();
86 printf("\nfio: terminating on signal %d\n", sig);
88 terminate_threads(TERMINATE_ALL, 0);
94 * Check if we are above the minimum rate given.
96 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 spent = mtime_since(&td->lastrate, now);
113 if (spent < td->ratecycle)
116 rate = (td->this_io_bytes[ddir] - td->rate_bytes) / spent;
117 if (rate < td->ratemin) {
118 fprintf(f_out, "%s: min rate %u not met, got %luKiB/sec\n", td->name, td->ratemin, rate);
123 td->rate_bytes = td->this_io_bytes[ddir];
124 memcpy(&td->lastrate, now, sizeof(*now));
128 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
132 if (mtime_since(&td->epoch, t) >= td->timeout * 1000)
138 static struct fio_file *get_next_file(struct thread_data *td)
140 unsigned int old_next_file = td->next_file;
144 f = &td->files[td->next_file];
147 if (td->next_file >= td->nr_files)
154 } while (td->next_file != old_next_file);
160 * When job exits, we can cancel the in-flight IO if we are using async
161 * io. Attempt to do so.
163 static void cleanup_pending_aio(struct thread_data *td)
165 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
166 struct list_head *entry, *n;
167 struct io_completion_data icd;
172 * get immediately available events, if any
174 r = td_io_getevents(td, 0, td->cur_depth, &ts);
177 ios_completed(td, &icd);
181 * now cancel remaining active events
183 if (td->io_ops->cancel) {
184 list_for_each_safe(entry, n, &td->io_u_busylist) {
185 io_u = list_entry(entry, struct io_u, list);
187 r = td->io_ops->cancel(td, io_u);
194 r = td_io_getevents(td, td->cur_depth, td->cur_depth, NULL);
197 ios_completed(td, &icd);
203 * Helper to handle the final sync of a file. Works just like the normal
204 * io path, just does everything sync.
206 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
208 struct io_u *io_u = __get_io_u(td);
209 struct io_completion_data icd;
215 io_u->ddir = DDIR_SYNC;
218 if (td_io_prep(td, io_u)) {
223 ret = td_io_queue(td, io_u);
225 td_verror(td, io_u->error);
230 ret = td_io_getevents(td, 1, td->cur_depth, NULL);
237 ios_completed(td, &icd);
239 td_verror(td, icd.error);
247 * The main verify engine. Runs over the writes we previusly submitted,
248 * reads the blocks back in, and checks the crc/md5 of the data.
250 static void do_verify(struct thread_data *td)
252 struct io_u *io_u, *v_io_u = NULL;
253 struct io_completion_data icd;
258 * sync io first and invalidate cache, to make sure we really
261 for_each_file(td, f, i) {
263 file_invalidate_cache(td, f);
266 td_set_runstate(td, TD_VERIFYING);
272 io_u = __get_io_u(td);
276 if (runtime_exceeded(td, &io_u->start_time)) {
281 if (get_next_verify(td, io_u)) {
286 f = get_next_file(td);
292 if (td_io_prep(td, io_u)) {
297 ret = td_io_queue(td, io_u);
299 td_verror(td, io_u->error);
305 * we have one pending to verify, do that while
306 * we are doing io on the next one
308 if (do_io_u_verify(td, &v_io_u))
311 ret = td_io_getevents(td, 1, 1, NULL);
318 v_io_u = td->io_ops->event(td, 0);
321 fio_gettime(&icd.time, NULL);
322 io_completed(td, v_io_u, &icd);
325 td_verror(td, icd.error);
326 put_io_u(td, v_io_u);
332 * if we can't submit more io, we need to verify now
334 if (queue_full(td) && do_io_u_verify(td, &v_io_u))
339 do_io_u_verify(td, &v_io_u);
342 cleanup_pending_aio(td);
344 td_set_runstate(td, TD_RUNNING);
348 * Not really an io thread, all it does is burn CPU cycles in the specified
351 static void do_cpuio(struct thread_data *td)
354 int split = 100 / td->cpuload;
357 while (!td->terminate) {
358 fio_gettime(&e, NULL);
360 if (runtime_exceeded(td, &e))
366 usec_sleep(td, 10000);
373 * Main IO worker function. It retrieves io_u's to process and queues
374 * and reaps them, checking for rate and errors along the way.
376 static void do_io(struct thread_data *td)
378 struct io_completion_data icd;
384 td_set_runstate(td, TD_RUNNING);
386 while (td->this_io_bytes[td->ddir] < td->io_size) {
387 struct timespec *timeout;
394 f = get_next_file(td);
398 io_u = get_io_u(td, f);
402 memcpy(&s, &io_u->start_time, sizeof(s));
405 ret = td_io_queue(td, io_u);
407 if (ret > 0 && (io_u->xfer_buflen != io_u->resid) &&
410 * short read/write. requeue.
412 io_u->xfer_buflen = io_u->resid;
413 io_u->xfer_buf += ret;
416 td_verror(td, io_u->error);
422 add_slat_sample(td, io_u->ddir, mtime_since(&io_u->start_time, &io_u->issue_time));
424 if (td->cur_depth < td->iodepth) {
425 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
434 ret = td_io_getevents(td, min_evts, td->cur_depth, timeout);
442 ios_completed(td, &icd);
444 td_verror(td, icd.error);
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, &icd.time);
455 rate_throttle(td, usec, icd.bytes_done[td->ddir], td->ddir);
457 if (check_min_rate(td, &icd.time)) {
458 if (exitall_on_terminate)
459 terminate_threads(td->groupid, 0);
460 td_verror(td, ENODATA);
464 if (runtime_exceeded(td, &icd.time))
468 unsigned long long b;
470 b = td->io_blocks[0] + td->io_blocks[1];
471 if (!(b % td->thinktime_blocks))
472 usec_sleep(td, td->thinktime);
478 cleanup_pending_aio(td);
480 if (should_fsync(td) && td->end_fsync) {
481 td_set_runstate(td, TD_FSYNCING);
482 for_each_file(td, f, i)
488 static void cleanup_io_u(struct thread_data *td)
490 struct list_head *entry, *n;
493 list_for_each_safe(entry, n, &td->io_u_freelist) {
494 io_u = list_entry(entry, struct io_u, list);
496 list_del(&io_u->list);
504 * "randomly" fill the buffer contents
506 static void fill_rand_buf(struct io_u *io_u, int max_bs)
508 int *ptr = io_u->buf;
510 while ((void *) ptr - io_u->buf < max_bs) {
511 *ptr = rand() * 0x9e370001;
516 static int init_io_u(struct thread_data *td)
523 if (td->io_ops->flags & FIO_CPUIO)
526 if (td->io_ops->flags & FIO_SYNCIO)
529 max_units = td->iodepth;
531 max_bs = max(td->max_bs[DDIR_READ], td->max_bs[DDIR_WRITE]);
532 td->orig_buffer_size = max_bs * max_units;
534 if (td->mem_type == MEM_SHMHUGE || td->mem_type == MEM_MMAPHUGE)
535 td->orig_buffer_size = (td->orig_buffer_size + td->hugepage_size - 1) & ~(td->hugepage_size - 1);
537 td->orig_buffer_size += page_mask;
539 if (allocate_io_mem(td))
542 p = ALIGN(td->orig_buffer);
543 for (i = 0; i < max_units; i++) {
544 io_u = malloc(sizeof(*io_u));
545 memset(io_u, 0, sizeof(*io_u));
546 INIT_LIST_HEAD(&io_u->list);
548 io_u->buf = p + max_bs * i;
549 if (td_write(td) || td_rw(td))
550 fill_rand_buf(io_u, max_bs);
553 list_add(&io_u->list, &td->io_u_freelist);
559 static int switch_ioscheduler(struct thread_data *td)
561 char tmp[256], tmp2[128];
565 if (td->io_ops->flags & FIO_CPUIO)
568 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
570 f = fopen(tmp, "r+");
572 td_verror(td, errno);
579 ret = fwrite(td->ioscheduler, strlen(td->ioscheduler), 1, f);
580 if (ferror(f) || ret != 1) {
581 td_verror(td, errno);
589 * Read back and check that the selected scheduler is now the default.
591 ret = fread(tmp, 1, sizeof(tmp), f);
592 if (ferror(f) || ret < 0) {
593 td_verror(td, errno);
598 sprintf(tmp2, "[%s]", td->ioscheduler);
599 if (!strstr(tmp, tmp2)) {
600 log_err("fio: io scheduler %s not found\n", td->ioscheduler);
601 td_verror(td, EINVAL);
610 static void clear_io_state(struct thread_data *td)
615 td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0;
616 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
619 for_each_file(td, f, i) {
621 if (td->io_ops->flags & FIO_SYNCIO)
622 lseek(f->fd, SEEK_SET, 0);
625 memset(f->file_map, 0, f->num_maps * sizeof(long));
630 * Entry point for the thread based jobs. The process based jobs end up
631 * here as well, after a little setup.
633 static void *thread_main(void *data)
635 unsigned long long runtime[2];
636 struct thread_data *td = data;
643 INIT_LIST_HEAD(&td->io_u_freelist);
644 INIT_LIST_HEAD(&td->io_u_busylist);
645 INIT_LIST_HEAD(&td->io_hist_list);
646 INIT_LIST_HEAD(&td->io_log_list);
651 if (fio_setaffinity(td) == -1) {
652 td_verror(td, errno);
660 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
661 td_verror(td, errno);
666 if (nice(td->nice) == -1) {
667 td_verror(td, errno);
671 if (init_random_state(td))
674 if (td->ioscheduler && switch_ioscheduler(td))
677 td_set_runstate(td, TD_INITIALIZED);
678 fio_sem_up(&startup_sem);
679 fio_sem_down(&td->mutex);
681 if (!td->create_serialize && setup_files(td))
687 * Do this late, as some IO engines would like to have the
688 * files setup prior to initializing structures.
694 system(td->exec_prerun);
696 fio_gettime(&td->epoch, NULL);
697 getrusage(RUSAGE_SELF, &td->ru_start);
699 runtime[0] = runtime[1] = 0;
700 while (td->loops--) {
701 fio_gettime(&td->start, NULL);
702 memcpy(&td->stat_sample_time, &td->start, sizeof(td->start));
705 memcpy(&td->lastrate, &td->stat_sample_time, sizeof(td->lastrate));
708 prune_io_piece_log(td);
710 if (td->io_ops->flags & FIO_CPUIO)
715 runtime[td->ddir] += utime_since_now(&td->start);
716 if (td_rw(td) && td->io_bytes[td->ddir ^ 1])
717 runtime[td->ddir ^ 1] = runtime[td->ddir];
719 if (td->error || td->terminate)
722 if (td->verify == VERIFY_NONE)
726 fio_gettime(&td->start, NULL);
730 runtime[DDIR_READ] += utime_since_now(&td->start);
732 if (td->error || td->terminate)
736 update_rusage_stat(td);
737 fio_gettime(&td->end_time, NULL);
738 td->runtime[0] = runtime[0] / 1000;
739 td->runtime[1] = runtime[1] / 1000;
742 finish_log(td, td->bw_log, "bw");
744 finish_log(td, td->slat_log, "slat");
746 finish_log(td, td->clat_log, "clat");
747 if (td->write_iolog_file)
748 write_iolog_close(td);
749 if (td->exec_postrun)
750 system(td->exec_postrun);
752 if (exitall_on_terminate)
753 terminate_threads(td->groupid, 0);
759 td_set_runstate(td, TD_EXITED);
765 * We cannot pass the td data into a forked process, so attach the td and
766 * pass it to the thread worker.
768 static void *fork_main(int shmid, int offset)
770 struct thread_data *td;
773 data = shmat(shmid, NULL, 0);
774 if (data == (void *) -1) {
779 td = data + offset * sizeof(struct thread_data);
786 * Run over the job map and reap the threads that have exited, if any.
788 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
790 struct thread_data *td;
791 int i, cputhreads, pending;
794 * reap exited threads (TD_EXITED -> TD_REAPED)
796 pending = cputhreads = 0;
799 * ->io_ops is NULL for a thread that has closed its
802 if (td->io_ops && td->io_ops->flags & FIO_CPUIO)
805 if (td->runstate != TD_EXITED) {
806 if (td->runstate < TD_RUNNING)
812 td_set_runstate(td, TD_REAPED);
814 if (td->use_thread) {
817 if (pthread_join(td->thread, (void *) &ret))
818 perror("thread_join");
820 waitpid(td->pid, NULL, 0);
823 (*m_rate) -= td->ratemin;
824 (*t_rate) -= td->rate;
827 if (*nr_running == cputhreads && !pending)
828 terminate_threads(TERMINATE_ALL, 0);
832 * Main function for kicking off and reaping jobs, as needed.
834 static void run_threads(void)
836 struct thread_data *td;
838 int i, todo, nr_running, m_rate, t_rate, nr_started;
840 if (fio_pin_memory())
844 printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
848 signal(SIGINT, sig_handler);
849 signal(SIGALRM, sig_handler);
851 todo = thread_number;
857 print_status_init(td->thread_number - 1);
859 if (!td->create_serialize) {
865 * do file setup here so it happens sequentially,
866 * we don't want X number of threads getting their
867 * client data interspersed on disk
869 if (setup_files(td)) {
870 td_set_runstate(td, TD_REAPED);
878 struct thread_data *map[MAX_JOBS];
879 struct timeval this_start;
880 int this_jobs = 0, left;
883 * create threads (TD_NOT_CREATED -> TD_CREATED)
886 if (td->runstate != TD_NOT_CREATED)
890 * never got a chance to start, killed by other
891 * thread for some reason
898 if (td->start_delay) {
899 spent = mtime_since_genesis();
901 if (td->start_delay * 1000 > spent)
905 if (td->stonewall && (nr_started || nr_running))
909 * Set state to created. Thread will transition
910 * to TD_INITIALIZED when it's done setting up.
912 td_set_runstate(td, TD_CREATED);
913 map[this_jobs++] = td;
914 fio_sem_init(&startup_sem, 1);
917 if (td->use_thread) {
918 if (pthread_create(&td->thread, NULL, thread_main, td)) {
919 perror("thread_create");
924 fio_sem_down(&startup_sem);
926 fork_main(shm_id, i);
933 * Wait for the started threads to transition to
936 fio_gettime(&this_start, NULL);
938 while (left && !fio_abort) {
939 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
944 for (i = 0; i < this_jobs; i++) {
948 if (td->runstate == TD_INITIALIZED) {
951 } else if (td->runstate >= TD_EXITED) {
955 nr_running++; /* work-around... */
961 log_err("fio: %d jobs failed to start\n", left);
962 for (i = 0; i < this_jobs; i++) {
966 kill(td->pid, SIGTERM);
972 * start created threads (TD_INITIALIZED -> TD_RUNNING).
975 if (td->runstate != TD_INITIALIZED)
978 td_set_runstate(td, TD_RUNNING);
981 m_rate += td->ratemin;
984 fio_sem_up(&td->mutex);
987 reap_threads(&nr_running, &t_rate, &m_rate);
994 reap_threads(&nr_running, &t_rate, &m_rate);
1002 int main(int argc, char *argv[])
1006 if (parse_options(argc, argv))
1009 if (!thread_number) {
1010 log_err("Nothing to do\n");
1014 ps = sysconf(_SC_PAGESIZE);
1016 log_err("Failed to get page size\n");
1023 setup_log(&agg_io_log[DDIR_READ]);
1024 setup_log(&agg_io_log[DDIR_WRITE]);
1027 disk_util_timer_arm();
1034 __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1035 __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");