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)
140 * When job exits, we can cancel the in-flight IO if we are using async
141 * io. Attempt to do so.
143 static void cleanup_pending_aio(struct thread_data *td)
145 struct list_head *entry, *n;
150 * get immediately available events, if any
152 r = io_u_queued_complete(td, 0, NULL);
157 * now cancel remaining active events
159 if (td->io_ops->cancel) {
160 list_for_each_safe(entry, n, &td->io_u_busylist) {
161 io_u = list_entry(entry, struct io_u, list);
164 * if the io_u isn't in flight, then that generally
165 * means someone leaked an io_u. complain but fix
166 * it up, so we don't stall here.
168 if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
169 log_err("fio: non-busy IO on busy list\n");
172 r = td->io_ops->cancel(td, io_u);
180 r = io_u_queued_complete(td, td->cur_depth, NULL);
184 * Helper to handle the final sync of a file. Works just like the normal
185 * io path, just does everything sync.
187 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
189 struct io_u *io_u = __get_io_u(td);
195 io_u->ddir = DDIR_SYNC;
198 if (td_io_prep(td, io_u)) {
204 ret = td_io_queue(td, io_u);
206 td_verror(td, io_u->error);
209 } else if (ret == FIO_Q_QUEUED) {
210 if (io_u_queued_complete(td, 1, NULL) < 0)
212 } else if (ret == FIO_Q_COMPLETED) {
214 td_verror(td, io_u->error);
218 if (io_u_sync_complete(td, io_u, NULL) < 0)
220 } else if (ret == FIO_Q_BUSY) {
221 if (td_io_commit(td))
230 * The main verify engine. Runs over the writes we previusly submitted,
231 * reads the blocks back in, and checks the crc/md5 of the data.
233 static void do_verify(struct thread_data *td)
237 int ret, i, min_events;
240 * sync io first and invalidate cache, to make sure we really
243 for_each_file(td, f, i) {
244 if (fio_io_sync(td, f))
246 if (file_invalidate_cache(td, f))
253 td_set_runstate(td, TD_VERIFYING);
256 while (!td->terminate) {
257 io_u = __get_io_u(td);
261 if (runtime_exceeded(td, &io_u->start_time)) {
266 if (get_next_verify(td, io_u)) {
271 if (td_io_prep(td, io_u)) {
276 ret = td_io_queue(td, io_u);
279 case FIO_Q_COMPLETED:
282 if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
283 int bytes = io_u->xfer_buflen - io_u->resid;
285 io_u->xfer_buflen = io_u->resid;
286 io_u->xfer_buf += bytes;
289 ret = io_u_sync_complete(td, io_u, verify_io_u);
296 requeue_io_u(td, &io_u);
297 ret = td_io_commit(td);
305 if (ret < 0 || td->error)
309 * if we can queue more, do so. but check if there are
310 * completed io_u's first.
313 if (queue_full(td) || ret == FIO_Q_BUSY) {
316 if (td->cur_depth > td->iodepth_low)
317 min_events = td->cur_depth - td->iodepth_low;
321 * Reap required number of io units, if any, and do the
322 * verification on them through the callback handler
324 if (io_u_queued_complete(td, min_events, verify_io_u) < 0)
329 cleanup_pending_aio(td);
331 td_set_runstate(td, TD_RUNNING);
335 * Not really an io thread, all it does is burn CPU cycles in the specified
338 static void do_cpuio(struct thread_data *td)
341 int split = 100 / td->cpuload;
344 while (!td->terminate) {
345 fio_gettime(&e, NULL);
347 if (runtime_exceeded(td, &e))
353 usec_sleep(td, 10000);
360 * Main IO worker function. It retrieves io_u's to process and queues
361 * and reaps them, checking for rate and errors along the way.
363 static void do_io(struct thread_data *td)
369 td_set_runstate(td, TD_RUNNING);
371 while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->io_size) {
372 struct timeval comp_time;
384 memcpy(&s, &io_u->start_time, sizeof(s));
386 if (runtime_exceeded(td, &s)) {
391 ret = td_io_queue(td, io_u);
394 case FIO_Q_COMPLETED:
399 if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
400 int bytes = io_u->xfer_buflen - io_u->resid;
402 io_u->xfer_buflen = io_u->resid;
403 io_u->xfer_buf += bytes;
406 fio_gettime(&comp_time, NULL);
407 bytes_done = io_u_sync_complete(td, io_u, NULL);
413 * if the engine doesn't have a commit hook,
414 * the io_u is really queued. if it does have such
415 * a hook, it has to call io_u_queued() itself.
417 if (td->io_ops->commit == NULL)
418 io_u_queued(td, io_u);
421 requeue_io_u(td, &io_u);
422 ret = td_io_commit(td);
430 if (ret < 0 || td->error)
434 * See if we need to complete some commands
436 if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
438 if (queue_full(td) || ret == FIO_Q_BUSY) {
441 if (td->cur_depth > td->iodepth_low)
442 min_evts = td->cur_depth - td->iodepth_low;
445 fio_gettime(&comp_time, NULL);
446 bytes_done = io_u_queued_complete(td, min_evts, NULL);
455 * the rate is batched for now, it should work for batches
456 * of completions except the very first one which may look
459 usec = utime_since(&s, &comp_time);
461 rate_throttle(td, usec, bytes_done, td->ddir);
463 if (check_min_rate(td, &comp_time)) {
464 if (exitall_on_terminate)
465 terminate_threads(td->groupid, 0);
466 td_verror(td, ENODATA);
471 unsigned long long b;
473 b = td->io_blocks[0] + td->io_blocks[1];
474 if (!(b % td->thinktime_blocks)) {
477 if (td->thinktime_spin)
478 __usec_sleep(td->thinktime_spin);
480 left = td->thinktime - td->thinktime_spin;
482 usec_sleep(td, left);
491 cleanup_pending_aio(td);
493 if (should_fsync(td) && td->end_fsync) {
494 td_set_runstate(td, TD_FSYNCING);
495 for_each_file(td, f, i)
501 static void cleanup_io_u(struct thread_data *td)
503 struct list_head *entry, *n;
506 list_for_each_safe(entry, n, &td->io_u_freelist) {
507 io_u = list_entry(entry, struct io_u, list);
509 list_del(&io_u->list);
517 * "randomly" fill the buffer contents
519 static void fill_rand_buf(struct io_u *io_u, int max_bs)
521 int *ptr = io_u->buf;
523 while ((void *) ptr - io_u->buf < max_bs) {
524 *ptr = rand() * 0x9e370001;
529 static int init_io_u(struct thread_data *td)
536 if (td->io_ops->flags & FIO_CPUIO)
539 if (td->io_ops->flags & FIO_SYNCIO)
542 max_units = td->iodepth;
544 max_bs = max(td->max_bs[DDIR_READ], td->max_bs[DDIR_WRITE]);
545 td->orig_buffer_size = max_bs * max_units;
547 if (td->mem_type == MEM_SHMHUGE || td->mem_type == MEM_MMAPHUGE)
548 td->orig_buffer_size = (td->orig_buffer_size + td->hugepage_size - 1) & ~(td->hugepage_size - 1);
550 td->orig_buffer_size += page_mask;
552 if (allocate_io_mem(td))
555 p = ALIGN(td->orig_buffer);
556 for (i = 0; i < max_units; i++) {
557 io_u = malloc(sizeof(*io_u));
558 memset(io_u, 0, sizeof(*io_u));
559 INIT_LIST_HEAD(&io_u->list);
561 io_u->buf = p + max_bs * i;
562 if (td_write(td) || td_rw(td))
563 fill_rand_buf(io_u, max_bs);
566 io_u->flags = IO_U_F_FREE;
567 list_add(&io_u->list, &td->io_u_freelist);
575 static int switch_ioscheduler(struct thread_data *td)
577 char tmp[256], tmp2[128];
581 if (td->io_ops->flags & FIO_CPUIO)
584 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
586 f = fopen(tmp, "r+");
588 td_verror(td, errno);
595 ret = fwrite(td->ioscheduler, strlen(td->ioscheduler), 1, f);
596 if (ferror(f) || ret != 1) {
597 td_verror(td, errno);
605 * Read back and check that the selected scheduler is now the default.
607 ret = fread(tmp, 1, sizeof(tmp), f);
608 if (ferror(f) || ret < 0) {
609 td_verror(td, errno);
614 sprintf(tmp2, "[%s]", td->ioscheduler);
615 if (!strstr(tmp, tmp2)) {
616 log_err("fio: io scheduler %s not found\n", td->ioscheduler);
617 td_verror(td, EINVAL);
626 static void clear_io_state(struct thread_data *td)
631 td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
632 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
635 td->last_was_sync = 0;
637 for_each_file(td, f, i) {
638 f->last_completed_pos = 0;
641 if (td->io_ops->flags & FIO_SYNCIO)
642 lseek(f->fd, SEEK_SET, 0);
645 memset(f->file_map, 0, f->num_maps * sizeof(long));
650 * Entry point for the thread based jobs. The process based jobs end up
651 * here as well, after a little setup.
653 static void *thread_main(void *data)
655 unsigned long long runtime[2];
656 struct thread_data *td = data;
663 INIT_LIST_HEAD(&td->io_u_freelist);
664 INIT_LIST_HEAD(&td->io_u_busylist);
665 INIT_LIST_HEAD(&td->io_u_requeues);
666 INIT_LIST_HEAD(&td->io_hist_list);
667 INIT_LIST_HEAD(&td->io_log_list);
672 if (fio_setaffinity(td) == -1) {
673 td_verror(td, errno);
681 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
682 td_verror(td, errno);
687 if (nice(td->nice) == -1) {
688 td_verror(td, errno);
692 if (init_random_state(td))
695 if (td->ioscheduler && switch_ioscheduler(td))
698 td_set_runstate(td, TD_INITIALIZED);
699 fio_sem_up(&startup_sem);
700 fio_sem_down(&td->mutex);
702 if (!td->create_serialize && setup_files(td))
708 * Do this late, as some IO engines would like to have the
709 * files setup prior to initializing structures.
714 if (td->exec_prerun) {
715 if (system(td->exec_prerun) < 0)
719 fio_gettime(&td->epoch, NULL);
720 memcpy(&td->timeout_end, &td->epoch, sizeof(td->epoch));
721 getrusage(RUSAGE_SELF, &td->ts.ru_start);
723 runtime[0] = runtime[1] = 0;
724 while (td->loops--) {
725 fio_gettime(&td->start, NULL);
726 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
729 memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
732 prune_io_piece_log(td);
734 if (td->io_ops->flags & FIO_CPUIO)
739 runtime[td->ddir] += utime_since_now(&td->start);
740 if (td_rw(td) && td->io_bytes[td->ddir ^ 1])
741 runtime[td->ddir ^ 1] = runtime[td->ddir];
743 if (td->error || td->terminate)
746 if (td->verify == VERIFY_NONE)
750 fio_gettime(&td->start, NULL);
754 runtime[DDIR_READ] += utime_since_now(&td->start);
756 if (td->error || td->terminate)
760 update_rusage_stat(td);
761 fio_gettime(&td->end_time, NULL);
762 td->runtime[0] = runtime[0] / 1000;
763 td->runtime[1] = runtime[1] / 1000;
766 finish_log(td, td->ts.bw_log, "bw");
768 finish_log(td, td->ts.slat_log, "slat");
770 finish_log(td, td->ts.clat_log, "clat");
771 if (td->write_iolog_file)
772 write_iolog_close(td);
773 if (td->exec_postrun) {
774 if (system(td->exec_postrun) < 0)
775 log_err("fio: postrun %s failed\n", td->exec_postrun);
778 if (exitall_on_terminate)
779 terminate_threads(td->groupid, 0);
783 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
787 td_set_runstate(td, TD_EXITED);
788 return (void *) (unsigned long) td->error;
792 * We cannot pass the td data into a forked process, so attach the td and
793 * pass it to the thread worker.
795 static int fork_main(int shmid, int offset)
797 struct thread_data *td;
800 data = shmat(shmid, NULL, 0);
801 if (data == (void *) -1) {
808 td = data + offset * sizeof(struct thread_data);
809 ret = thread_main(td);
811 return (int) (unsigned long) ret;
815 * Run over the job map and reap the threads that have exited, if any.
817 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
819 struct thread_data *td;
820 int i, cputhreads, pending, status, ret;
823 * reap exited threads (TD_EXITED -> TD_REAPED)
825 pending = cputhreads = 0;
830 * ->io_ops is NULL for a thread that has closed its
833 if (td->io_ops && td->io_ops->flags & FIO_CPUIO)
836 if (!td->pid || td->runstate == TD_REAPED)
840 if (td->runstate == TD_EXITED)
844 * check if someone quit or got killed in an unusual way
846 ret = waitpid(td->pid, &status, flags);
847 if (ret < 0 && !td->use_thread) {
848 if (errno == ECHILD) {
849 log_err("fio: pid=%d disappeared %d\n", td->pid, td->runstate);
850 td_set_runstate(td, TD_REAPED);
854 } else if (ret == td->pid) {
855 if (WIFSIGNALED(status)) {
856 int sig = WTERMSIG(status);
858 log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
859 td_set_runstate(td, TD_REAPED);
862 if (WIFEXITED(status)) {
863 if (WEXITSTATUS(status) && !td->error)
864 td->error = WEXITSTATUS(status);
865 if (td->use_thread) {
868 if (pthread_join(td->thread, (void *) &ret))
869 perror("pthread_join");
871 td_set_runstate(td, TD_REAPED);
877 * thread is not dead, continue
882 (*m_rate) -= td->ratemin;
883 (*t_rate) -= td->rate;
889 if (*nr_running == cputhreads && !pending)
890 terminate_threads(TERMINATE_ALL, 0);
894 * Main function for kicking off and reaping jobs, as needed.
896 static void run_threads(void)
898 struct thread_data *td;
900 int i, todo, nr_running, m_rate, t_rate, nr_started;
902 if (fio_pin_memory())
906 printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
910 signal(SIGINT, sig_handler);
911 signal(SIGALRM, sig_handler);
913 todo = thread_number;
919 print_status_init(td->thread_number - 1);
921 if (!td->create_serialize) {
927 * do file setup here so it happens sequentially,
928 * we don't want X number of threads getting their
929 * client data interspersed on disk
931 if (setup_files(td)) {
934 log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
935 td_set_runstate(td, TD_REAPED);
945 struct thread_data *map[MAX_JOBS];
946 struct timeval this_start;
947 int this_jobs = 0, left;
950 * create threads (TD_NOT_CREATED -> TD_CREATED)
953 if (td->runstate != TD_NOT_CREATED)
957 * never got a chance to start, killed by other
958 * thread for some reason
965 if (td->start_delay) {
966 spent = mtime_since_genesis();
968 if (td->start_delay * 1000 > spent)
972 if (td->stonewall && (nr_started || nr_running))
976 * Set state to created. Thread will transition
977 * to TD_INITIALIZED when it's done setting up.
979 td_set_runstate(td, TD_CREATED);
980 map[this_jobs++] = td;
981 fio_sem_init(&startup_sem, 1);
984 if (td->use_thread) {
985 if (pthread_create(&td->thread, NULL, thread_main, td)) {
986 perror("thread_create");
991 fio_sem_down(&startup_sem);
993 int ret = fork_main(shm_id, i);
1001 * Wait for the started threads to transition to
1004 fio_gettime(&this_start, NULL);
1006 while (left && !fio_abort) {
1007 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1012 for (i = 0; i < this_jobs; i++) {
1016 if (td->runstate == TD_INITIALIZED) {
1019 } else if (td->runstate >= TD_EXITED) {
1023 nr_running++; /* work-around... */
1029 log_err("fio: %d jobs failed to start\n", left);
1030 for (i = 0; i < this_jobs; i++) {
1034 kill(td->pid, SIGTERM);
1040 * start created threads (TD_INITIALIZED -> TD_RUNNING).
1042 for_each_td(td, i) {
1043 if (td->runstate != TD_INITIALIZED)
1046 td_set_runstate(td, TD_RUNNING);
1049 m_rate += td->ratemin;
1052 fio_sem_up(&td->mutex);
1055 reap_threads(&nr_running, &t_rate, &m_rate);
1061 while (nr_running) {
1062 reap_threads(&nr_running, &t_rate, &m_rate);
1070 int main(int argc, char *argv[])
1075 * We need locale for number printing, if it isn't set then just
1076 * go with the US format.
1078 if (!getenv("LC_NUMERIC"))
1079 setlocale(LC_NUMERIC, "en_US");
1081 if (parse_options(argc, argv))
1084 if (!thread_number) {
1085 log_err("Nothing to do\n");
1089 ps = sysconf(_SC_PAGESIZE);
1091 log_err("Failed to get page size\n");
1098 setup_log(&agg_io_log[DDIR_READ]);
1099 setup_log(&agg_io_log[DDIR_WRITE]);
1104 disk_util_timer_arm();
1111 __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1112 __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");