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 unsigned long page_mask;
41 unsigned long page_size;
43 (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
46 int thread_number = 0;
52 static struct fio_sem *startup_sem;
53 static volatile int fio_abort;
54 static int exit_value;
56 struct io_log *agg_io_log[2];
58 #define TERMINATE_ALL (-1)
59 #define JOB_START_TIMEOUT (5 * 1000)
61 static inline void td_set_runstate(struct thread_data *td, int runstate)
63 td->runstate = runstate;
66 static void terminate_threads(int group_id)
68 struct thread_data *td;
72 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
74 * if the thread is running, just let it exit
76 if (td->runstate < TD_RUNNING)
77 kill(td->pid, SIGQUIT);
84 static void sig_handler(int sig)
89 disk_util_timer_arm();
90 print_thread_status();
93 printf("\nfio: terminating on signal %d\n", sig);
95 terminate_threads(TERMINATE_ALL);
101 * Check if we are above the minimum rate given.
103 static int check_min_rate(struct thread_data *td, struct timeval *now)
105 unsigned long long bytes = 0;
110 * No minimum rate set, always ok
116 * allow a 2 second settle period in the beginning
118 if (mtime_since(&td->start, now) < 2000)
122 bytes += td->this_io_bytes[DDIR_READ];
124 bytes += td->this_io_bytes[DDIR_WRITE];
127 * if rate blocks is set, sample is running
129 if (td->rate_bytes) {
130 spent = mtime_since(&td->lastrate, now);
131 if (spent < td->ratecycle)
134 if (bytes < td->rate_bytes) {
135 fprintf(f_out, "%s: min rate %u not met\n", td->name, td->ratemin);
138 rate = (bytes - td->rate_bytes) / spent;
139 if (rate < td->ratemin || bytes < td->rate_bytes) {
140 fprintf(f_out, "%s: min rate %u not met, got %luKiB/sec\n", td->name, td->ratemin, rate);
146 td->rate_bytes = bytes;
147 memcpy(&td->lastrate, now, sizeof(*now));
151 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
155 if (mtime_since(&td->epoch, t) >= td->timeout * 1000)
162 * When job exits, we can cancel the in-flight IO if we are using async
163 * io. Attempt to do so.
165 static void cleanup_pending_aio(struct thread_data *td)
167 struct list_head *entry, *n;
172 * get immediately available events, if any
174 r = io_u_queued_complete(td, 0);
179 * now cancel remaining active events
181 if (td->io_ops->cancel) {
182 list_for_each_safe(entry, n, &td->io_u_busylist) {
183 io_u = list_entry(entry, struct io_u, list);
186 * if the io_u isn't in flight, then that generally
187 * means someone leaked an io_u. complain but fix
188 * it up, so we don't stall here.
190 if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
191 log_err("fio: non-busy IO on busy list\n");
194 r = td->io_ops->cancel(td, io_u);
202 r = io_u_queued_complete(td, td->cur_depth);
206 * Helper to handle the final sync of a file. Works just like the normal
207 * io path, just does everything sync.
209 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
211 struct io_u *io_u = __get_io_u(td);
217 io_u->ddir = DDIR_SYNC;
220 if (td_io_prep(td, io_u)) {
226 ret = td_io_queue(td, io_u);
228 td_verror(td, io_u->error, "td_io_queue");
231 } else if (ret == FIO_Q_QUEUED) {
232 if (io_u_queued_complete(td, 1) < 0)
234 } else if (ret == FIO_Q_COMPLETED) {
236 td_verror(td, io_u->error, "td_io_queue");
240 if (io_u_sync_complete(td, io_u) < 0)
242 } else if (ret == FIO_Q_BUSY) {
243 if (td_io_commit(td))
252 * The main verify engine. Runs over the writes we previously submitted,
253 * reads the blocks back in, and checks the crc/md5 of the data.
255 static void do_verify(struct thread_data *td)
263 * sync io first and invalidate cache, to make sure we really
266 for_each_file(td, f, i) {
267 if (!(f->flags & FIO_FILE_OPEN))
269 if (fio_io_sync(td, f))
271 if (file_invalidate_cache(td, f))
278 td_set_runstate(td, TD_VERIFYING);
281 while (!td->terminate) {
284 io_u = __get_io_u(td);
288 if (runtime_exceeded(td, &io_u->start_time)) {
293 if (get_next_verify(td, io_u)) {
298 if (td_io_prep(td, io_u)) {
303 io_u->end_io = verify_io_u;
305 ret = td_io_queue(td, io_u);
307 case FIO_Q_COMPLETED:
310 else if (io_u->resid) {
311 int bytes = io_u->xfer_buflen - io_u->resid;
317 td_verror(td, ENODATA, "full resid");
321 io_u->xfer_buflen = io_u->resid;
322 io_u->xfer_buf += bytes;
323 requeue_io_u(td, &io_u);
325 ret = io_u_sync_complete(td, io_u);
333 requeue_io_u(td, &io_u);
334 ret2 = td_io_commit(td);
340 td_verror(td, -ret, "td_io_queue");
344 if (ret < 0 || td->error)
348 * if we can queue more, do so. but check if there are
349 * completed io_u's first.
352 if (queue_full(td) || ret == FIO_Q_BUSY) {
355 if (td->cur_depth > td->iodepth_low)
356 min_events = td->cur_depth - td->iodepth_low;
360 * Reap required number of io units, if any, and do the
361 * verification on them through the callback handler
363 if (io_u_queued_complete(td, min_events) < 0)
368 min_events = td->cur_depth;
371 ret = io_u_queued_complete(td, min_events);
373 cleanup_pending_aio(td);
375 td_set_runstate(td, TD_RUNNING);
379 * Main IO worker function. It retrieves io_u's to process and queues
380 * and reaps them, checking for rate and errors along the way.
382 static void do_io(struct thread_data *td)
389 td_set_runstate(td, TD_RUNNING);
391 while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->io_size) {
392 struct timeval comp_time;
405 memcpy(&s, &io_u->start_time, sizeof(s));
407 if (runtime_exceeded(td, &s)) {
412 ret = td_io_queue(td, io_u);
414 case FIO_Q_COMPLETED:
417 else if (io_u->resid) {
418 int bytes = io_u->xfer_buflen - io_u->resid;
424 td_verror(td, ENODATA, "full resid");
429 io_u->xfer_buflen = io_u->resid;
430 io_u->xfer_buf += bytes;
431 requeue_io_u(td, &io_u);
433 fio_gettime(&comp_time, NULL);
434 bytes_done = io_u_sync_complete(td, io_u);
441 * if the engine doesn't have a commit hook,
442 * the io_u is really queued. if it does have such
443 * a hook, it has to call io_u_queued() itself.
445 if (td->io_ops->commit == NULL)
446 io_u_queued(td, io_u);
449 requeue_io_u(td, &io_u);
450 ret2 = td_io_commit(td);
460 if (ret < 0 || td->error)
464 * See if we need to complete some commands
466 if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
468 if (queue_full(td) || ret == FIO_Q_BUSY) {
471 if (td->cur_depth > td->iodepth_low)
472 min_evts = td->cur_depth - td->iodepth_low;
475 fio_gettime(&comp_time, NULL);
476 bytes_done = io_u_queued_complete(td, min_evts);
485 * the rate is batched for now, it should work for batches
486 * of completions except the very first one which may look
489 usec = utime_since(&s, &comp_time);
491 rate_throttle(td, usec, bytes_done);
493 if (check_min_rate(td, &comp_time)) {
494 if (exitall_on_terminate)
495 terminate_threads(td->groupid);
496 td_verror(td, ENODATA, "check_min_rate");
501 unsigned long long b;
503 b = td->io_blocks[0] + td->io_blocks[1];
504 if (!(b % td->thinktime_blocks)) {
507 if (td->thinktime_spin)
508 __usec_sleep(td->thinktime_spin);
510 left = td->thinktime - td->thinktime_spin;
512 usec_sleep(td, left);
522 ret = io_u_queued_complete(td, i);
524 if (should_fsync(td) && td->end_fsync) {
525 td_set_runstate(td, TD_FSYNCING);
527 for_each_file(td, f, i) {
528 if (!(f->flags & FIO_FILE_OPEN))
534 cleanup_pending_aio(td);
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_SYNCIO)
575 max_units = td->iodepth;
577 max_bs = max(td->max_bs[DDIR_READ], td->max_bs[DDIR_WRITE]);
578 td->orig_buffer_size = max_bs * max_units;
580 if (td->mem_type == MEM_SHMHUGE || td->mem_type == MEM_MMAPHUGE)
581 td->orig_buffer_size = (td->orig_buffer_size + td->hugepage_size - 1) & ~(td->hugepage_size - 1);
583 td->orig_buffer_size += page_mask;
585 if (allocate_io_mem(td))
588 p = ALIGN(td->orig_buffer);
589 for (i = 0; i < max_units; i++) {
590 io_u = malloc(sizeof(*io_u));
591 memset(io_u, 0, sizeof(*io_u));
592 INIT_LIST_HEAD(&io_u->list);
594 io_u->buf = p + max_bs * i;
595 if (td_write(td) || td_rw(td))
596 fill_rand_buf(io_u, max_bs);
599 io_u->flags = IO_U_F_FREE;
600 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_DISKLESSIO)
617 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
619 f = fopen(tmp, "r+");
621 td_verror(td, errno, "fopen");
628 ret = fwrite(td->ioscheduler, strlen(td->ioscheduler), 1, f);
629 if (ferror(f) || ret != 1) {
630 td_verror(td, errno, "fwrite");
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, "fread");
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, "iosched_switch");
659 static int clear_io_state(struct thread_data *td)
665 td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
666 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
670 td->last_was_sync = 0;
672 for_each_file(td, f, i)
673 td_io_close_file(td, f);
676 for_each_file(td, f, i) {
677 ret = td_io_open_file(td, f);
686 * Entry point for the thread based jobs. The process based jobs end up
687 * here as well, after a little setup.
689 static void *thread_main(void *data)
691 unsigned long long runtime[2];
692 struct thread_data *td = data;
700 INIT_LIST_HEAD(&td->io_u_freelist);
701 INIT_LIST_HEAD(&td->io_u_busylist);
702 INIT_LIST_HEAD(&td->io_u_requeues);
703 INIT_LIST_HEAD(&td->io_hist_list);
704 INIT_LIST_HEAD(&td->io_log_list);
709 if (fio_setaffinity(td) == -1) {
710 td_verror(td, errno, "cpu_set_affinity");
718 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
719 td_verror(td, errno, "ioprio_set");
724 if (nice(td->nice) == -1) {
725 td_verror(td, errno, "nice");
729 if (init_random_state(td))
732 if (td->ioscheduler && switch_ioscheduler(td))
735 td_set_runstate(td, TD_INITIALIZED);
736 fio_sem_up(startup_sem);
737 fio_sem_down(td->mutex);
740 * the ->mutex semaphore is now no longer used, close it to avoid
741 * eating a file descriptor
743 fio_sem_remove(td->mutex);
745 if (!td->create_serialize && setup_files(td))
754 if (td->exec_prerun) {
755 if (system(td->exec_prerun) < 0)
759 fio_gettime(&td->epoch, NULL);
760 memcpy(&td->timeout_end, &td->epoch, sizeof(td->epoch));
761 getrusage(RUSAGE_SELF, &td->ts.ru_start);
763 runtime[0] = runtime[1] = 0;
765 while (td->loops--) {
766 fio_gettime(&td->start, NULL);
767 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
770 memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
772 if (clear_state && clear_io_state(td))
775 prune_io_piece_log(td);
781 if (td_read(td) && td->io_bytes[DDIR_READ])
782 runtime[DDIR_READ] += utime_since_now(&td->start);
783 if (td_write(td) && td->io_bytes[DDIR_WRITE])
784 runtime[DDIR_WRITE] += utime_since_now(&td->start);
786 if (td->error || td->terminate)
789 if (td->verify == VERIFY_NONE)
792 if (clear_io_state(td))
795 fio_gettime(&td->start, NULL);
799 runtime[DDIR_READ] += utime_since_now(&td->start);
801 if (td->error || td->terminate)
805 update_rusage_stat(td);
806 td->ts.runtime[0] = runtime[0] / 1000;
807 td->ts.runtime[1] = runtime[1] / 1000;
808 td->ts.total_run_time = mtime_since_now(&td->epoch);
809 td->ts.io_bytes[0] = td->io_bytes[0];
810 td->ts.io_bytes[1] = td->io_bytes[1];
813 finish_log(td, td->ts.bw_log, "bw");
815 finish_log(td, td->ts.slat_log, "slat");
817 finish_log(td, td->ts.clat_log, "clat");
818 if (td->write_iolog_file)
819 write_iolog_close(td);
820 if (td->exec_postrun) {
821 if (system(td->exec_postrun) < 0)
822 log_err("fio: postrun %s failed\n", td->exec_postrun);
825 if (exitall_on_terminate)
826 terminate_threads(td->groupid);
830 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
834 td_set_runstate(td, TD_EXITED);
835 return (void *) (unsigned long) td->error;
837 fio_sem_up(startup_sem);
842 * We cannot pass the td data into a forked process, so attach the td and
843 * pass it to the thread worker.
845 static int fork_main(int shmid, int offset)
847 struct thread_data *td;
850 data = shmat(shmid, NULL, 0);
851 if (data == (void *) -1) {
858 td = data + offset * sizeof(struct thread_data);
859 ret = thread_main(td);
861 return (int) (unsigned long) ret;
865 * Run over the job map and reap the threads that have exited, if any.
867 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
869 struct thread_data *td;
870 int i, cputhreads, pending, status, ret;
873 * reap exited threads (TD_EXITED -> TD_REAPED)
875 pending = cputhreads = 0;
880 * ->io_ops is NULL for a thread that has closed its
883 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
886 if (!td->pid || td->runstate == TD_REAPED)
888 if (td->use_thread) {
889 if (td->runstate == TD_EXITED) {
890 td_set_runstate(td, TD_REAPED);
897 if (td->runstate == TD_EXITED)
901 * check if someone quit or got killed in an unusual way
903 ret = waitpid(td->pid, &status, flags);
905 if (errno == ECHILD) {
906 log_err("fio: pid=%d disappeared %d\n", td->pid, td->runstate);
907 td_set_runstate(td, TD_REAPED);
911 } else if (ret == td->pid) {
912 if (WIFSIGNALED(status)) {
913 int sig = WTERMSIG(status);
916 log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
917 td_set_runstate(td, TD_REAPED);
920 if (WIFEXITED(status)) {
921 if (WEXITSTATUS(status) && !td->error)
922 td->error = WEXITSTATUS(status);
924 td_set_runstate(td, TD_REAPED);
930 * thread is not dead, continue
934 if (td->use_thread) {
937 if (pthread_join(td->thread, (void *) &ret))
938 perror("pthread_join");
942 (*m_rate) -= td->ratemin;
943 (*t_rate) -= td->rate;
949 if (*nr_running == cputhreads && !pending)
950 terminate_threads(TERMINATE_ALL);
954 * Main function for kicking off and reaping jobs, as needed.
956 static void run_threads(void)
958 struct thread_data *td;
960 int i, todo, nr_running, m_rate, t_rate, nr_started;
962 if (fio_pin_memory())
968 printf("%d thread%s", nr_thread, nr_thread > 1 ? "s" : "");
972 printf("%d process%s", nr_process, nr_process > 1 ? "es" : "");
978 signal(SIGINT, sig_handler);
979 signal(SIGALRM, sig_handler);
981 todo = thread_number;
987 print_status_init(td->thread_number - 1);
989 if (!td->create_serialize) {
995 * do file setup here so it happens sequentially,
996 * we don't want X number of threads getting their
997 * client data interspersed on disk
999 if (setup_files(td)) {
1002 log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
1003 td_set_runstate(td, TD_REAPED);
1013 struct thread_data *map[MAX_JOBS];
1014 struct timeval this_start;
1015 int this_jobs = 0, left;
1018 * create threads (TD_NOT_CREATED -> TD_CREATED)
1020 for_each_td(td, i) {
1021 if (td->runstate != TD_NOT_CREATED)
1025 * never got a chance to start, killed by other
1026 * thread for some reason
1028 if (td->terminate) {
1033 if (td->start_delay) {
1034 spent = mtime_since_genesis();
1036 if (td->start_delay * 1000 > spent)
1040 if (td->stonewall && (nr_started || nr_running))
1044 * Set state to created. Thread will transition
1045 * to TD_INITIALIZED when it's done setting up.
1047 td_set_runstate(td, TD_CREATED);
1048 map[this_jobs++] = td;
1051 if (td->use_thread) {
1052 if (pthread_create(&td->thread, NULL, thread_main, td)) {
1053 perror("thread_create");
1059 int ret = fork_main(shm_id, i);
1064 fio_sem_down(startup_sem);
1068 * Wait for the started threads to transition to
1071 fio_gettime(&this_start, NULL);
1073 while (left && !fio_abort) {
1074 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1079 for (i = 0; i < this_jobs; i++) {
1083 if (td->runstate == TD_INITIALIZED) {
1086 } else if (td->runstate >= TD_EXITED) {
1090 nr_running++; /* work-around... */
1096 log_err("fio: %d jobs failed to start\n", left);
1097 for (i = 0; i < this_jobs; i++) {
1101 kill(td->pid, SIGTERM);
1107 * start created threads (TD_INITIALIZED -> TD_RUNNING).
1109 for_each_td(td, i) {
1110 if (td->runstate != TD_INITIALIZED)
1113 td_set_runstate(td, TD_RUNNING);
1116 m_rate += td->ratemin;
1119 fio_sem_up(td->mutex);
1122 reap_threads(&nr_running, &t_rate, &m_rate);
1128 while (nr_running) {
1129 reap_threads(&nr_running, &t_rate, &m_rate);
1137 int main(int argc, char *argv[])
1142 * We need locale for number printing, if it isn't set then just
1143 * go with the US format.
1145 if (!getenv("LC_NUMERIC"))
1146 setlocale(LC_NUMERIC, "en_US");
1148 if (parse_options(argc, argv))
1151 if (!thread_number) {
1152 log_err("Nothing to do\n");
1156 ps = sysconf(_SC_PAGESIZE);
1158 log_err("Failed to get page size\n");
1166 setup_log(&agg_io_log[DDIR_READ]);
1167 setup_log(&agg_io_log[DDIR_WRITE]);
1170 startup_sem = fio_sem_init(0);
1174 disk_util_timer_arm();
1181 __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1182 __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");
1186 fio_sem_remove(startup_sem);