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;
51 static struct fio_sem *startup_sem;
52 static volatile int fio_abort;
53 static int exit_value;
55 struct io_log *agg_io_log[2];
57 #define TERMINATE_ALL (-1)
58 #define JOB_START_TIMEOUT (5 * 1000)
60 static inline void td_set_runstate(struct thread_data *td, int runstate)
62 td->runstate = runstate;
65 static void terminate_threads(int group_id)
67 struct thread_data *td;
71 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
73 * if the thread is running, just let it exit
75 if (td->runstate < TD_RUNNING)
76 kill(td->pid, SIGQUIT);
83 static void sig_handler(int sig)
88 disk_util_timer_arm();
89 print_thread_status();
92 printf("\nfio: terminating on signal %d\n", sig);
94 terminate_threads(TERMINATE_ALL);
100 * Check if we are above the minimum rate given.
102 static int check_min_rate(struct thread_data *td, struct timeval *now)
104 unsigned long long bytes = 0;
109 * No minimum rate set, always ok
115 * allow a 2 second settle period in the beginning
117 if (mtime_since(&td->start, now) < 2000)
121 bytes += td->this_io_bytes[DDIR_READ];
123 bytes += td->this_io_bytes[DDIR_WRITE];
126 * if rate blocks is set, sample is running
128 if (td->rate_bytes) {
129 spent = mtime_since(&td->lastrate, now);
130 if (spent < td->ratecycle)
133 if (bytes < td->rate_bytes) {
134 fprintf(f_out, "%s: min rate %u not met\n", td->name, td->ratemin);
137 rate = (bytes - td->rate_bytes) / spent;
138 if (rate < td->ratemin || bytes < td->rate_bytes) {
139 fprintf(f_out, "%s: min rate %u not met, got %luKiB/sec\n", td->name, td->ratemin, rate);
145 td->rate_bytes = bytes;
146 memcpy(&td->lastrate, now, sizeof(*now));
150 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
154 if (mtime_since(&td->epoch, t) >= td->timeout * 1000)
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 list_head *entry, *n;
171 * get immediately available events, if any
173 r = io_u_queued_complete(td, 0);
178 * now cancel remaining active events
180 if (td->io_ops->cancel) {
181 list_for_each_safe(entry, n, &td->io_u_busylist) {
182 io_u = list_entry(entry, struct io_u, list);
185 * if the io_u isn't in flight, then that generally
186 * means someone leaked an io_u. complain but fix
187 * it up, so we don't stall here.
189 if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
190 log_err("fio: non-busy IO on busy list\n");
193 r = td->io_ops->cancel(td, io_u);
201 r = io_u_queued_complete(td, td->cur_depth);
205 * Helper to handle the final sync of a file. Works just like the normal
206 * io path, just does everything sync.
208 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
210 struct io_u *io_u = __get_io_u(td);
216 io_u->ddir = DDIR_SYNC;
219 if (td_io_prep(td, io_u)) {
225 ret = td_io_queue(td, io_u);
227 td_verror(td, io_u->error, "td_io_queue");
230 } else if (ret == FIO_Q_QUEUED) {
231 if (io_u_queued_complete(td, 1) < 0)
233 } else if (ret == FIO_Q_COMPLETED) {
235 td_verror(td, io_u->error, "td_io_queue");
239 if (io_u_sync_complete(td, io_u) < 0)
241 } else if (ret == FIO_Q_BUSY) {
242 if (td_io_commit(td))
251 * The main verify engine. Runs over the writes we previously submitted,
252 * reads the blocks back in, and checks the crc/md5 of the data.
254 static void do_verify(struct thread_data *td)
258 int ret, i, min_events;
261 * sync io first and invalidate cache, to make sure we really
264 for_each_file(td, f, i) {
265 if (fio_io_sync(td, f))
267 if (file_invalidate_cache(td, f))
274 td_set_runstate(td, TD_VERIFYING);
277 while (!td->terminate) {
280 io_u = __get_io_u(td);
284 if (runtime_exceeded(td, &io_u->start_time)) {
289 if (get_next_verify(td, io_u)) {
294 if (td_io_prep(td, io_u)) {
299 io_u->end_io = verify_io_u;
301 ret = td_io_queue(td, io_u);
303 case FIO_Q_COMPLETED:
306 else if (io_u->resid) {
307 int bytes = io_u->xfer_buflen - io_u->resid;
313 td_verror(td, ENODATA, "full resid");
317 io_u->xfer_buflen = io_u->resid;
318 io_u->xfer_buf += bytes;
319 requeue_io_u(td, &io_u);
321 ret = io_u_sync_complete(td, io_u);
329 requeue_io_u(td, &io_u);
330 ret2 = td_io_commit(td);
336 td_verror(td, -ret, "td_io_queue");
340 if (ret < 0 || td->error)
344 * if we can queue more, do so. but check if there are
345 * completed io_u's first.
348 if (queue_full(td) || ret == FIO_Q_BUSY) {
351 if (td->cur_depth > td->iodepth_low)
352 min_events = td->cur_depth - td->iodepth_low;
356 * Reap required number of io units, if any, and do the
357 * verification on them through the callback handler
359 if (io_u_queued_complete(td, min_events) < 0)
364 min_events = td->cur_depth;
367 ret = io_u_queued_complete(td, min_events);
369 cleanup_pending_aio(td);
371 td_set_runstate(td, TD_RUNNING);
375 * Main IO worker function. It retrieves io_u's to process and queues
376 * and reaps them, checking for rate and errors along the way.
378 static void do_io(struct thread_data *td)
384 td_set_runstate(td, TD_RUNNING);
386 while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->io_size) {
387 struct timeval comp_time;
400 memcpy(&s, &io_u->start_time, sizeof(s));
402 if (runtime_exceeded(td, &s)) {
407 ret = td_io_queue(td, io_u);
409 case FIO_Q_COMPLETED:
412 else if (io_u->resid) {
413 int bytes = io_u->xfer_buflen - io_u->resid;
419 td_verror(td, ENODATA, "full resid");
424 io_u->xfer_buflen = io_u->resid;
425 io_u->xfer_buf += bytes;
426 requeue_io_u(td, &io_u);
428 fio_gettime(&comp_time, NULL);
429 bytes_done = io_u_sync_complete(td, io_u);
436 * if the engine doesn't have a commit hook,
437 * the io_u is really queued. if it does have such
438 * a hook, it has to call io_u_queued() itself.
440 if (td->io_ops->commit == NULL)
441 io_u_queued(td, io_u);
444 requeue_io_u(td, &io_u);
445 ret2 = td_io_commit(td);
455 if (ret < 0 || td->error)
459 * See if we need to complete some commands
461 if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
463 if (queue_full(td) || ret == FIO_Q_BUSY) {
466 if (td->cur_depth > td->iodepth_low)
467 min_evts = td->cur_depth - td->iodepth_low;
470 fio_gettime(&comp_time, NULL);
471 bytes_done = io_u_queued_complete(td, min_evts);
480 * the rate is batched for now, it should work for batches
481 * of completions except the very first one which may look
484 usec = utime_since(&s, &comp_time);
486 rate_throttle(td, usec, bytes_done);
488 if (check_min_rate(td, &comp_time)) {
489 if (exitall_on_terminate)
490 terminate_threads(td->groupid);
491 td_verror(td, ENODATA, "check_min_rate");
496 unsigned long long b;
498 b = td->io_blocks[0] + td->io_blocks[1];
499 if (!(b % td->thinktime_blocks)) {
502 if (td->thinktime_spin)
503 __usec_sleep(td->thinktime_spin);
505 left = td->thinktime - td->thinktime_spin;
507 usec_sleep(td, left);
517 ret = io_u_queued_complete(td, i);
519 if (should_fsync(td) && td->end_fsync) {
520 td_set_runstate(td, TD_FSYNCING);
521 for_each_file(td, f, i)
525 cleanup_pending_aio(td);
528 static void cleanup_io_u(struct thread_data *td)
530 struct list_head *entry, *n;
533 list_for_each_safe(entry, n, &td->io_u_freelist) {
534 io_u = list_entry(entry, struct io_u, list);
536 list_del(&io_u->list);
544 * "randomly" fill the buffer contents
546 static void fill_rand_buf(struct io_u *io_u, int max_bs)
548 int *ptr = io_u->buf;
550 while ((void *) ptr - io_u->buf < max_bs) {
551 *ptr = rand() * 0x9e370001;
556 static int init_io_u(struct thread_data *td)
563 if (td->io_ops->flags & FIO_SYNCIO)
566 max_units = td->iodepth;
568 max_bs = max(td->max_bs[DDIR_READ], td->max_bs[DDIR_WRITE]);
569 td->orig_buffer_size = max_bs * max_units;
571 if (td->mem_type == MEM_SHMHUGE || td->mem_type == MEM_MMAPHUGE)
572 td->orig_buffer_size = (td->orig_buffer_size + td->hugepage_size - 1) & ~(td->hugepage_size - 1);
574 td->orig_buffer_size += page_mask;
576 if (allocate_io_mem(td))
579 p = ALIGN(td->orig_buffer);
580 for (i = 0; i < max_units; i++) {
581 io_u = malloc(sizeof(*io_u));
582 memset(io_u, 0, sizeof(*io_u));
583 INIT_LIST_HEAD(&io_u->list);
585 io_u->buf = p + max_bs * i;
586 if (td_write(td) || td_rw(td))
587 fill_rand_buf(io_u, max_bs);
590 io_u->flags = IO_U_F_FREE;
591 list_add(&io_u->list, &td->io_u_freelist);
599 static int switch_ioscheduler(struct thread_data *td)
601 char tmp[256], tmp2[128];
605 if (td->io_ops->flags & FIO_DISKLESSIO)
608 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
610 f = fopen(tmp, "r+");
612 td_verror(td, errno, "fopen");
619 ret = fwrite(td->ioscheduler, strlen(td->ioscheduler), 1, f);
620 if (ferror(f) || ret != 1) {
621 td_verror(td, errno, "fwrite");
629 * Read back and check that the selected scheduler is now the default.
631 ret = fread(tmp, 1, sizeof(tmp), f);
632 if (ferror(f) || ret < 0) {
633 td_verror(td, errno, "fread");
638 sprintf(tmp2, "[%s]", td->ioscheduler);
639 if (!strstr(tmp, tmp2)) {
640 log_err("fio: io scheduler %s not found\n", td->ioscheduler);
641 td_verror(td, EINVAL, "iosched_switch");
650 static int clear_io_state(struct thread_data *td)
655 td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
656 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
660 td->last_was_sync = 0;
662 for_each_file(td, f, i)
663 td_io_close_file(td, f);
666 for_each_file(td, f, i) {
667 ret = td_io_open_file(td, f);
676 * Entry point for the thread based jobs. The process based jobs end up
677 * here as well, after a little setup.
679 static void *thread_main(void *data)
681 unsigned long long runtime[2];
682 struct thread_data *td = data;
690 INIT_LIST_HEAD(&td->io_u_freelist);
691 INIT_LIST_HEAD(&td->io_u_busylist);
692 INIT_LIST_HEAD(&td->io_u_requeues);
693 INIT_LIST_HEAD(&td->io_hist_list);
694 INIT_LIST_HEAD(&td->io_log_list);
699 if (fio_setaffinity(td) == -1) {
700 td_verror(td, errno, "cpu_set_affinity");
708 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
709 td_verror(td, errno, "ioprio_set");
714 if (nice(td->nice) == -1) {
715 td_verror(td, errno, "nice");
719 if (init_random_state(td))
722 if (td->ioscheduler && switch_ioscheduler(td))
725 td_set_runstate(td, TD_INITIALIZED);
726 fio_sem_up(startup_sem);
727 fio_sem_down(td->mutex);
730 * the ->mutex semaphore is now no longer used, close it to avoid
731 * eating a file descriptor
733 fio_sem_remove(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;
755 while (td->loops--) {
756 fio_gettime(&td->start, NULL);
757 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
760 memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
762 if (clear_state && clear_io_state(td))
765 prune_io_piece_log(td);
771 if (td_read(td) && td->io_bytes[DDIR_READ])
772 runtime[DDIR_READ] += utime_since_now(&td->start);
773 if (td_write(td) && td->io_bytes[DDIR_WRITE])
774 runtime[DDIR_WRITE] += utime_since_now(&td->start);
776 if (td->error || td->terminate)
779 if (td->verify == VERIFY_NONE)
782 if (clear_io_state(td))
785 fio_gettime(&td->start, NULL);
789 runtime[DDIR_READ] += utime_since_now(&td->start);
791 if (td->error || td->terminate)
795 update_rusage_stat(td);
796 td->ts.runtime[0] = runtime[0] / 1000;
797 td->ts.runtime[1] = runtime[1] / 1000;
798 td->ts.total_run_time = mtime_since_now(&td->epoch);
799 td->ts.io_bytes[0] = td->io_bytes[0];
800 td->ts.io_bytes[1] = td->io_bytes[1];
803 finish_log(td, td->ts.bw_log, "bw");
805 finish_log(td, td->ts.slat_log, "slat");
807 finish_log(td, td->ts.clat_log, "clat");
808 if (td->write_iolog_file)
809 write_iolog_close(td);
810 if (td->exec_postrun) {
811 if (system(td->exec_postrun) < 0)
812 log_err("fio: postrun %s failed\n", td->exec_postrun);
815 if (exitall_on_terminate)
816 terminate_threads(td->groupid);
820 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
824 td_set_runstate(td, TD_EXITED);
825 return (void *) (unsigned long) td->error;
827 fio_sem_up(startup_sem);
832 * We cannot pass the td data into a forked process, so attach the td and
833 * pass it to the thread worker.
835 static int fork_main(int shmid, int offset)
837 struct thread_data *td;
840 data = shmat(shmid, NULL, 0);
841 if (data == (void *) -1) {
848 td = data + offset * sizeof(struct thread_data);
849 ret = thread_main(td);
851 return (int) (unsigned long) ret;
855 * Run over the job map and reap the threads that have exited, if any.
857 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
859 struct thread_data *td;
860 int i, cputhreads, pending, status, ret;
863 * reap exited threads (TD_EXITED -> TD_REAPED)
865 pending = cputhreads = 0;
870 * ->io_ops is NULL for a thread that has closed its
873 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
876 if (!td->pid || td->runstate == TD_REAPED)
878 if (td->use_thread) {
879 if (td->runstate == TD_EXITED) {
880 td_set_runstate(td, TD_REAPED);
887 if (td->runstate == TD_EXITED)
891 * check if someone quit or got killed in an unusual way
893 ret = waitpid(td->pid, &status, flags);
895 if (errno == ECHILD) {
896 log_err("fio: pid=%d disappeared %d\n", td->pid, td->runstate);
897 td_set_runstate(td, TD_REAPED);
901 } else if (ret == td->pid) {
902 if (WIFSIGNALED(status)) {
903 int sig = WTERMSIG(status);
906 log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
907 td_set_runstate(td, TD_REAPED);
910 if (WIFEXITED(status)) {
911 if (WEXITSTATUS(status) && !td->error)
912 td->error = WEXITSTATUS(status);
914 td_set_runstate(td, TD_REAPED);
920 * thread is not dead, continue
924 if (td->use_thread) {
927 if (pthread_join(td->thread, (void *) &ret))
928 perror("pthread_join");
932 (*m_rate) -= td->ratemin;
933 (*t_rate) -= td->rate;
939 if (*nr_running == cputhreads && !pending)
940 terminate_threads(TERMINATE_ALL);
944 * Main function for kicking off and reaping jobs, as needed.
946 static void run_threads(void)
948 struct thread_data *td;
950 int i, todo, nr_running, m_rate, t_rate, nr_started;
952 if (fio_pin_memory())
958 printf("%d thread%s", nr_thread, nr_thread > 1 ? "s" : "");
962 printf("%d process%s", nr_process, nr_process > 1 ? "es" : "");
968 signal(SIGINT, sig_handler);
969 signal(SIGALRM, sig_handler);
971 todo = thread_number;
977 print_status_init(td->thread_number - 1);
979 if (!td->create_serialize) {
985 * do file setup here so it happens sequentially,
986 * we don't want X number of threads getting their
987 * client data interspersed on disk
989 if (setup_files(td)) {
992 log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
993 td_set_runstate(td, TD_REAPED);
1003 struct thread_data *map[MAX_JOBS];
1004 struct timeval this_start;
1005 int this_jobs = 0, left;
1008 * create threads (TD_NOT_CREATED -> TD_CREATED)
1010 for_each_td(td, i) {
1011 if (td->runstate != TD_NOT_CREATED)
1015 * never got a chance to start, killed by other
1016 * thread for some reason
1018 if (td->terminate) {
1023 if (td->start_delay) {
1024 spent = mtime_since_genesis();
1026 if (td->start_delay * 1000 > spent)
1030 if (td->stonewall && (nr_started || nr_running))
1034 * Set state to created. Thread will transition
1035 * to TD_INITIALIZED when it's done setting up.
1037 td_set_runstate(td, TD_CREATED);
1038 map[this_jobs++] = td;
1041 if (td->use_thread) {
1042 if (pthread_create(&td->thread, NULL, thread_main, td)) {
1043 perror("thread_create");
1049 int ret = fork_main(shm_id, i);
1054 fio_sem_down(startup_sem);
1058 * Wait for the started threads to transition to
1061 fio_gettime(&this_start, NULL);
1063 while (left && !fio_abort) {
1064 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1069 for (i = 0; i < this_jobs; i++) {
1073 if (td->runstate == TD_INITIALIZED) {
1076 } else if (td->runstate >= TD_EXITED) {
1080 nr_running++; /* work-around... */
1086 log_err("fio: %d jobs failed to start\n", left);
1087 for (i = 0; i < this_jobs; i++) {
1091 kill(td->pid, SIGTERM);
1097 * start created threads (TD_INITIALIZED -> TD_RUNNING).
1099 for_each_td(td, i) {
1100 if (td->runstate != TD_INITIALIZED)
1103 td_set_runstate(td, TD_RUNNING);
1106 m_rate += td->ratemin;
1109 fio_sem_up(td->mutex);
1112 reap_threads(&nr_running, &t_rate, &m_rate);
1118 while (nr_running) {
1119 reap_threads(&nr_running, &t_rate, &m_rate);
1127 int main(int argc, char *argv[])
1132 * We need locale for number printing, if it isn't set then just
1133 * go with the US format.
1135 if (!getenv("LC_NUMERIC"))
1136 setlocale(LC_NUMERIC, "en_US");
1138 if (parse_options(argc, argv))
1141 if (!thread_number) {
1142 log_err("Nothing to do\n");
1146 ps = sysconf(_SC_PAGESIZE);
1148 log_err("Failed to get page size\n");
1155 setup_log(&agg_io_log[DDIR_READ]);
1156 setup_log(&agg_io_log[DDIR_WRITE]);
1159 startup_sem = fio_sem_init(0);
1163 disk_util_timer_arm();
1170 __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1171 __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");
1175 fio_sem_remove(startup_sem);