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)
65 struct thread_data *td;
69 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
71 * if the thread is running, just let it exit
73 if (td->runstate < TD_RUNNING)
74 kill(td->pid, SIGQUIT);
81 static void sig_handler(int sig)
86 disk_util_timer_arm();
87 print_thread_status();
90 printf("\nfio: terminating on signal %d\n", sig);
92 terminate_threads(TERMINATE_ALL);
98 * Check if we are above the minimum rate given.
100 static int check_min_rate(struct thread_data *td, struct timeval *now)
102 unsigned long long bytes = 0;
107 * No minimum rate set, always ok
113 * allow a 2 second settle period in the beginning
115 if (mtime_since(&td->start, now) < 2000)
119 bytes += td->this_io_bytes[DDIR_READ];
121 bytes += td->this_io_bytes[DDIR_WRITE];
124 * if rate blocks is set, sample is running
126 if (td->rate_bytes) {
127 spent = mtime_since(&td->lastrate, now);
128 if (spent < td->ratecycle)
131 if (bytes < td->rate_bytes) {
132 fprintf(f_out, "%s: min rate %u not met\n", td->name, td->ratemin);
135 rate = (bytes - td->rate_bytes) / spent;
136 if (rate < td->ratemin || bytes < td->rate_bytes) {
137 fprintf(f_out, "%s: min rate %u not met, got %luKiB/sec\n", td->name, td->ratemin, rate);
143 td->rate_bytes = bytes;
144 memcpy(&td->lastrate, now, sizeof(*now));
148 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
152 if (mtime_since(&td->epoch, t) >= td->timeout * 1000)
159 * When job exits, we can cancel the in-flight IO if we are using async
160 * io. Attempt to do so.
162 static void cleanup_pending_aio(struct thread_data *td)
164 struct list_head *entry, *n;
169 * get immediately available events, if any
171 r = io_u_queued_complete(td, 0);
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);
183 * if the io_u isn't in flight, then that generally
184 * means someone leaked an io_u. complain but fix
185 * it up, so we don't stall here.
187 if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
188 log_err("fio: non-busy IO on busy list\n");
191 r = td->io_ops->cancel(td, io_u);
199 r = io_u_queued_complete(td, td->cur_depth);
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);
214 io_u->ddir = DDIR_SYNC;
217 if (td_io_prep(td, io_u)) {
223 ret = td_io_queue(td, io_u);
225 td_verror(td, io_u->error, "td_io_queue");
228 } else if (ret == FIO_Q_QUEUED) {
229 if (io_u_queued_complete(td, 1) < 0)
231 } else if (ret == FIO_Q_COMPLETED) {
233 td_verror(td, io_u->error, "td_io_queue");
237 if (io_u_sync_complete(td, io_u) < 0)
239 } else if (ret == FIO_Q_BUSY) {
240 if (td_io_commit(td))
249 * The main verify engine. Runs over the writes we previously submitted,
250 * reads the blocks back in, and checks the crc/md5 of the data.
252 static void do_verify(struct thread_data *td)
256 int ret, i, min_events;
259 * sync io first and invalidate cache, to make sure we really
262 for_each_file(td, f, i) {
263 if (fio_io_sync(td, f))
265 if (file_invalidate_cache(td, f))
272 td_set_runstate(td, TD_VERIFYING);
275 while (!td->terminate) {
278 io_u = __get_io_u(td);
282 if (runtime_exceeded(td, &io_u->start_time)) {
287 if (get_next_verify(td, io_u)) {
292 if (td_io_prep(td, io_u)) {
297 io_u->end_io = verify_io_u;
299 ret = td_io_queue(td, io_u);
301 case FIO_Q_COMPLETED:
304 else if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
305 int bytes = io_u->xfer_buflen - io_u->resid;
307 io_u->xfer_buflen = io_u->resid;
308 io_u->xfer_buf += bytes;
309 requeue_io_u(td, &io_u);
311 ret = io_u_sync_complete(td, io_u);
319 requeue_io_u(td, &io_u);
320 ret2 = td_io_commit(td);
326 td_verror(td, -ret, "td_io_queue");
330 if (ret < 0 || td->error)
334 * if we can queue more, do so. but check if there are
335 * completed io_u's first.
338 if (queue_full(td) || ret == FIO_Q_BUSY) {
341 if (td->cur_depth > td->iodepth_low)
342 min_events = td->cur_depth - td->iodepth_low;
346 * Reap required number of io units, if any, and do the
347 * verification on them through the callback handler
349 if (io_u_queued_complete(td, min_events) < 0)
354 min_events = td->cur_depth;
357 ret = io_u_queued_complete(td, min_events);
359 cleanup_pending_aio(td);
361 td_set_runstate(td, TD_RUNNING);
365 * Not really an io thread, all it does is burn CPU cycles in the specified
368 static void do_cpuio(struct thread_data *td)
371 int split = 100 / td->cpuload;
374 while (!td->terminate) {
375 fio_gettime(&e, NULL);
377 if (runtime_exceeded(td, &e))
383 usec_sleep(td, 10000);
390 * Main IO worker function. It retrieves io_u's to process and queues
391 * and reaps them, checking for rate and errors along the way.
393 static void do_io(struct thread_data *td)
399 td_set_runstate(td, TD_RUNNING);
401 while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->io_size) {
402 struct timeval comp_time;
415 memcpy(&s, &io_u->start_time, sizeof(s));
417 if (runtime_exceeded(td, &s)) {
422 ret = td_io_queue(td, io_u);
424 case FIO_Q_COMPLETED:
427 else if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
428 int bytes = io_u->xfer_buflen - io_u->resid;
430 io_u->xfer_buflen = io_u->resid;
431 io_u->xfer_buf += bytes;
432 requeue_io_u(td, &io_u);
434 fio_gettime(&comp_time, NULL);
435 bytes_done = io_u_sync_complete(td, io_u);
442 * if the engine doesn't have a commit hook,
443 * the io_u is really queued. if it does have such
444 * a hook, it has to call io_u_queued() itself.
446 if (td->io_ops->commit == NULL)
447 io_u_queued(td, io_u);
450 requeue_io_u(td, &io_u);
451 ret2 = td_io_commit(td);
461 if (ret < 0 || td->error)
465 * See if we need to complete some commands
467 if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
469 if (queue_full(td) || ret == FIO_Q_BUSY) {
472 if (td->cur_depth > td->iodepth_low)
473 min_evts = td->cur_depth - td->iodepth_low;
476 fio_gettime(&comp_time, NULL);
477 bytes_done = io_u_queued_complete(td, min_evts);
486 * the rate is batched for now, it should work for batches
487 * of completions except the very first one which may look
490 usec = utime_since(&s, &comp_time);
492 rate_throttle(td, usec, bytes_done);
494 if (check_min_rate(td, &comp_time)) {
495 if (exitall_on_terminate)
496 terminate_threads(td->groupid);
497 td_verror(td, ENODATA, "check_min_rate");
502 unsigned long long b;
504 b = td->io_blocks[0] + td->io_blocks[1];
505 if (!(b % td->thinktime_blocks)) {
508 if (td->thinktime_spin)
509 __usec_sleep(td->thinktime_spin);
511 left = td->thinktime - td->thinktime_spin;
513 usec_sleep(td, left);
523 ret = io_u_queued_complete(td, i);
525 if (should_fsync(td) && td->end_fsync) {
526 td_set_runstate(td, TD_FSYNCING);
527 for_each_file(td, f, i)
531 cleanup_pending_aio(td);
534 static void cleanup_io_u(struct thread_data *td)
536 struct list_head *entry, *n;
539 list_for_each_safe(entry, n, &td->io_u_freelist) {
540 io_u = list_entry(entry, struct io_u, list);
542 list_del(&io_u->list);
550 * "randomly" fill the buffer contents
552 static void fill_rand_buf(struct io_u *io_u, int max_bs)
554 int *ptr = io_u->buf;
556 while ((void *) ptr - io_u->buf < max_bs) {
557 *ptr = rand() * 0x9e370001;
562 static int init_io_u(struct thread_data *td)
569 if (td->io_ops->flags & FIO_CPUIO)
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_CPUIO)
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 void clear_io_state(struct thread_data *td)
664 td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
665 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
668 td->last_was_sync = 0;
670 for_each_file(td, f, i) {
671 f->last_completed_pos = 0;
674 if (td->io_ops->flags & FIO_SYNCIO)
675 lseek(f->fd, SEEK_SET, 0);
678 memset(f->file_map, 0, f->num_maps * sizeof(long));
683 * Entry point for the thread based jobs. The process based jobs end up
684 * here as well, after a little setup.
686 static void *thread_main(void *data)
688 unsigned long long runtime[2];
689 struct thread_data *td = data;
696 INIT_LIST_HEAD(&td->io_u_freelist);
697 INIT_LIST_HEAD(&td->io_u_busylist);
698 INIT_LIST_HEAD(&td->io_u_requeues);
699 INIT_LIST_HEAD(&td->io_hist_list);
700 INIT_LIST_HEAD(&td->io_log_list);
705 if (fio_setaffinity(td) == -1) {
706 td_verror(td, errno, "cpu_set_affinity");
714 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
715 td_verror(td, errno, "ioprio_set");
720 if (nice(td->nice) == -1) {
721 td_verror(td, errno, "nice");
725 if (init_random_state(td))
728 if (td->ioscheduler && switch_ioscheduler(td))
731 td_set_runstate(td, TD_INITIALIZED);
732 fio_sem_up(&startup_sem);
733 fio_sem_down(&td->mutex);
735 if (!td->create_serialize && setup_files(td))
744 if (td->exec_prerun) {
745 if (system(td->exec_prerun) < 0)
749 fio_gettime(&td->epoch, NULL);
750 memcpy(&td->timeout_end, &td->epoch, sizeof(td->epoch));
751 getrusage(RUSAGE_SELF, &td->ts.ru_start);
753 runtime[0] = runtime[1] = 0;
754 while (td->loops--) {
755 fio_gettime(&td->start, NULL);
756 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
759 memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
762 prune_io_piece_log(td);
764 if (td->io_ops->flags & FIO_CPUIO)
769 if (td_read(td) && td->io_bytes[DDIR_READ])
770 runtime[DDIR_READ] += utime_since_now(&td->start);
771 if (td_write(td) && td->io_bytes[DDIR_WRITE])
772 runtime[DDIR_WRITE] += utime_since_now(&td->start);
774 if (td->error || td->terminate)
777 if (td->verify == VERIFY_NONE)
781 fio_gettime(&td->start, NULL);
785 runtime[DDIR_READ] += utime_since_now(&td->start);
787 if (td->error || td->terminate)
791 update_rusage_stat(td);
792 td->ts.runtime[0] = runtime[0] / 1000;
793 td->ts.runtime[1] = runtime[1] / 1000;
794 td->ts.total_run_time = mtime_since_now(&td->epoch);
795 td->ts.io_bytes[0] = td->io_bytes[0];
796 td->ts.io_bytes[1] = td->io_bytes[1];
799 finish_log(td, td->ts.bw_log, "bw");
801 finish_log(td, td->ts.slat_log, "slat");
803 finish_log(td, td->ts.clat_log, "clat");
804 if (td->write_iolog_file)
805 write_iolog_close(td);
806 if (td->exec_postrun) {
807 if (system(td->exec_postrun) < 0)
808 log_err("fio: postrun %s failed\n", td->exec_postrun);
811 if (exitall_on_terminate)
812 terminate_threads(td->groupid);
816 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
820 td_set_runstate(td, TD_EXITED);
821 return (void *) (unsigned long) td->error;
825 * We cannot pass the td data into a forked process, so attach the td and
826 * pass it to the thread worker.
828 static int fork_main(int shmid, int offset)
830 struct thread_data *td;
833 data = shmat(shmid, NULL, 0);
834 if (data == (void *) -1) {
841 td = data + offset * sizeof(struct thread_data);
842 ret = thread_main(td);
844 return (int) (unsigned long) ret;
848 * Run over the job map and reap the threads that have exited, if any.
850 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
852 struct thread_data *td;
853 int i, cputhreads, pending, status, ret;
856 * reap exited threads (TD_EXITED -> TD_REAPED)
858 pending = cputhreads = 0;
863 * ->io_ops is NULL for a thread that has closed its
866 if (td->io_ops && td->io_ops->flags & FIO_CPUIO)
869 if (!td->pid || td->runstate == TD_REAPED)
871 if (td->use_thread) {
872 if (td->runstate == TD_EXITED) {
873 td_set_runstate(td, TD_REAPED);
880 if (td->runstate == TD_EXITED)
884 * check if someone quit or got killed in an unusual way
886 ret = waitpid(td->pid, &status, flags);
888 if (errno == ECHILD) {
889 log_err("fio: pid=%d disappeared %d\n", td->pid, td->runstate);
890 td_set_runstate(td, TD_REAPED);
894 } else if (ret == td->pid) {
895 if (WIFSIGNALED(status)) {
896 int sig = WTERMSIG(status);
899 log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
900 td_set_runstate(td, TD_REAPED);
903 if (WIFEXITED(status)) {
904 if (WEXITSTATUS(status) && !td->error)
905 td->error = WEXITSTATUS(status);
907 td_set_runstate(td, TD_REAPED);
913 * thread is not dead, continue
917 if (td->use_thread) {
920 if (pthread_join(td->thread, (void *) &ret))
921 perror("pthread_join");
925 (*m_rate) -= td->ratemin;
926 (*t_rate) -= td->rate;
932 if (*nr_running == cputhreads && !pending)
933 terminate_threads(TERMINATE_ALL);
937 * Main function for kicking off and reaping jobs, as needed.
939 static void run_threads(void)
941 struct thread_data *td;
943 int i, todo, nr_running, m_rate, t_rate, nr_started;
945 if (fio_pin_memory())
949 printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
953 signal(SIGINT, sig_handler);
954 signal(SIGALRM, sig_handler);
956 todo = thread_number;
962 print_status_init(td->thread_number - 1);
964 if (!td->create_serialize) {
970 * do file setup here so it happens sequentially,
971 * we don't want X number of threads getting their
972 * client data interspersed on disk
974 if (setup_files(td)) {
977 log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
978 td_set_runstate(td, TD_REAPED);
988 struct thread_data *map[MAX_JOBS];
989 struct timeval this_start;
990 int this_jobs = 0, left;
993 * create threads (TD_NOT_CREATED -> TD_CREATED)
996 if (td->runstate != TD_NOT_CREATED)
1000 * never got a chance to start, killed by other
1001 * thread for some reason
1003 if (td->terminate) {
1008 if (td->start_delay) {
1009 spent = mtime_since_genesis();
1011 if (td->start_delay * 1000 > spent)
1015 if (td->stonewall && (nr_started || nr_running))
1019 * Set state to created. Thread will transition
1020 * to TD_INITIALIZED when it's done setting up.
1022 td_set_runstate(td, TD_CREATED);
1023 map[this_jobs++] = td;
1024 fio_sem_init(&startup_sem, 1);
1027 if (td->use_thread) {
1028 if (pthread_create(&td->thread, NULL, thread_main, td)) {
1029 perror("thread_create");
1034 fio_sem_down(&startup_sem);
1036 int ret = fork_main(shm_id, i);
1044 * Wait for the started threads to transition to
1047 fio_gettime(&this_start, NULL);
1049 while (left && !fio_abort) {
1050 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1055 for (i = 0; i < this_jobs; i++) {
1059 if (td->runstate == TD_INITIALIZED) {
1062 } else if (td->runstate >= TD_EXITED) {
1066 nr_running++; /* work-around... */
1072 log_err("fio: %d jobs failed to start\n", left);
1073 for (i = 0; i < this_jobs; i++) {
1077 kill(td->pid, SIGTERM);
1083 * start created threads (TD_INITIALIZED -> TD_RUNNING).
1085 for_each_td(td, i) {
1086 if (td->runstate != TD_INITIALIZED)
1089 td_set_runstate(td, TD_RUNNING);
1092 m_rate += td->ratemin;
1095 fio_sem_up(&td->mutex);
1098 reap_threads(&nr_running, &t_rate, &m_rate);
1104 while (nr_running) {
1105 reap_threads(&nr_running, &t_rate, &m_rate);
1113 int main(int argc, char *argv[])
1118 * We need locale for number printing, if it isn't set then just
1119 * go with the US format.
1121 if (!getenv("LC_NUMERIC"))
1122 setlocale(LC_NUMERIC, "en_US");
1124 if (parse_options(argc, argv))
1127 if (!thread_number) {
1128 log_err("Nothing to do\n");
1132 ps = sysconf(_SC_PAGESIZE);
1134 log_err("Failed to get page size\n");
1141 setup_log(&agg_io_log[DDIR_READ]);
1142 setup_log(&agg_io_log[DDIR_WRITE]);
1147 disk_util_timer_arm();
1154 __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1155 __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");