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)
139 static struct fio_file *get_next_file(struct thread_data *td)
141 unsigned int old_next_file = td->next_file;
145 f = &td->files[td->next_file];
148 if (td->next_file >= td->nr_files)
155 } while (td->next_file != old_next_file);
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 timespec ts = { .tv_sec = 0, .tv_nsec = 0};
167 struct list_head *entry, *n;
168 struct io_completion_data icd;
173 * get immediately available events, if any
175 r = td_io_getevents(td, 0, td->cur_depth, &ts);
178 ios_completed(td, &icd);
182 * now cancel remaining active events
184 if (td->io_ops->cancel) {
185 list_for_each_safe(entry, n, &td->io_u_busylist) {
186 io_u = list_entry(entry, struct io_u, list);
188 r = td->io_ops->cancel(td, io_u);
195 r = td_io_getevents(td, td->cur_depth, td->cur_depth, NULL);
198 ios_completed(td, &icd);
204 * Helper to handle the final sync of a file. Works just like the normal
205 * io path, just does everything sync.
207 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
209 struct io_u *io_u = __get_io_u(td);
210 struct io_completion_data icd;
216 io_u->ddir = DDIR_SYNC;
219 if (td_io_prep(td, io_u)) {
224 ret = td_io_queue(td, io_u);
226 td_verror(td, io_u->error);
229 } else if (ret == FIO_Q_QUEUED) {
230 ret = td_io_getevents(td, 1, td->cur_depth, NULL);
237 ios_completed(td, &icd);
239 td_verror(td, icd.error);
242 } else if (ret == FIO_Q_COMPLETED) {
244 td_verror(td, io_u->error);
249 io_completed(td, io_u, &icd);
257 * The main verify engine. Runs over the writes we previusly submitted,
258 * reads the blocks back in, and checks the crc/md5 of the data.
260 static void do_verify(struct thread_data *td)
267 * sync io first and invalidate cache, to make sure we really
270 for_each_file(td, f, i) {
272 file_invalidate_cache(td, f);
275 td_set_runstate(td, TD_VERIFYING);
278 while (!td->terminate) {
279 io_u = __get_io_u(td);
283 if (runtime_exceeded(td, &io_u->start_time))
286 if (get_next_verify(td, io_u))
289 if (td_io_prep(td, io_u))
293 ret = td_io_queue(td, io_u);
296 case FIO_Q_COMPLETED:
299 if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
300 int bytes = io_u->xfer_buflen - io_u->resid;
302 io_u->xfer_buflen = io_u->resid;
303 io_u->xfer_buf += bytes;
306 if (do_io_u_verify(td, &io_u)) {
320 * We get here for a queued request, in the future we
321 * want to later make this take full advantage of
322 * keeping IO in flight while verifying others.
324 ret = td_io_getevents(td, 1, 1, NULL);
329 io_u = td->io_ops->event(td, 0);
331 if (do_io_u_verify(td, &io_u))
339 cleanup_pending_aio(td);
341 td_set_runstate(td, TD_RUNNING);
345 * Not really an io thread, all it does is burn CPU cycles in the specified
348 static void do_cpuio(struct thread_data *td)
351 int split = 100 / td->cpuload;
354 while (!td->terminate) {
355 fio_gettime(&e, NULL);
357 if (runtime_exceeded(td, &e))
363 usec_sleep(td, 10000);
370 * Main IO worker function. It retrieves io_u's to process and queues
371 * and reaps them, checking for rate and errors along the way.
373 static void do_io(struct thread_data *td)
375 struct io_completion_data icd;
381 td_set_runstate(td, TD_RUNNING);
383 while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->io_size) {
384 struct timespec *timeout;
391 f = get_next_file(td);
395 io_u = get_io_u(td, f);
399 memcpy(&s, &io_u->start_time, sizeof(s));
401 ret = td_io_queue(td, io_u);
404 case FIO_Q_COMPLETED:
409 if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
410 int bytes = io_u->xfer_buflen - io_u->resid;
412 io_u->xfer_buflen = io_u->resid;
413 io_u->xfer_buf += bytes;
417 io_completed(td, io_u, &icd);
431 add_slat_sample(td, io_u->ddir, mtime_since(&io_u->start_time, &io_u->issue_time));
433 if (ret == FIO_Q_QUEUED) {
434 if (td->cur_depth < td->iodepth) {
446 ret = td_io_getevents(td, min_evts, td->cur_depth, timeout);
454 ios_completed(td, &icd);
456 td_verror(td, icd.error);
462 * the rate is batched for now, it should work for batches
463 * of completions except the very first one which may look
466 usec = utime_since(&s, &icd.time);
468 rate_throttle(td, usec, icd.bytes_done[td->ddir], td->ddir);
470 if (check_min_rate(td, &icd.time)) {
471 if (exitall_on_terminate)
472 terminate_threads(td->groupid, 0);
473 td_verror(td, ENODATA);
477 if (runtime_exceeded(td, &icd.time))
481 unsigned long long b;
483 b = td->io_blocks[0] + td->io_blocks[1];
484 if (!(b % td->thinktime_blocks)) {
487 if (td->thinktime_spin)
488 __usec_sleep(td->thinktime_spin);
490 left = td->thinktime - td->thinktime_spin;
492 usec_sleep(td, left);
499 cleanup_pending_aio(td);
501 if (should_fsync(td) && td->end_fsync) {
502 td_set_runstate(td, TD_FSYNCING);
503 for_each_file(td, f, i)
509 static void cleanup_io_u(struct thread_data *td)
511 struct list_head *entry, *n;
514 list_for_each_safe(entry, n, &td->io_u_freelist) {
515 io_u = list_entry(entry, struct io_u, list);
517 list_del(&io_u->list);
525 * "randomly" fill the buffer contents
527 static void fill_rand_buf(struct io_u *io_u, int max_bs)
529 int *ptr = io_u->buf;
531 while ((void *) ptr - io_u->buf < max_bs) {
532 *ptr = rand() * 0x9e370001;
537 static int init_io_u(struct thread_data *td)
544 if (td->io_ops->flags & FIO_CPUIO)
547 if (td->io_ops->flags & FIO_SYNCIO)
550 max_units = td->iodepth;
552 max_bs = max(td->max_bs[DDIR_READ], td->max_bs[DDIR_WRITE]);
553 td->orig_buffer_size = max_bs * max_units;
555 if (td->mem_type == MEM_SHMHUGE || td->mem_type == MEM_MMAPHUGE)
556 td->orig_buffer_size = (td->orig_buffer_size + td->hugepage_size - 1) & ~(td->hugepage_size - 1);
558 td->orig_buffer_size += page_mask;
560 if (allocate_io_mem(td))
563 p = ALIGN(td->orig_buffer);
564 for (i = 0; i < max_units; i++) {
565 io_u = malloc(sizeof(*io_u));
566 memset(io_u, 0, sizeof(*io_u));
567 INIT_LIST_HEAD(&io_u->list);
569 io_u->buf = p + max_bs * i;
570 if (td_write(td) || td_rw(td))
571 fill_rand_buf(io_u, max_bs);
574 list_add(&io_u->list, &td->io_u_freelist);
580 static int switch_ioscheduler(struct thread_data *td)
582 char tmp[256], tmp2[128];
586 if (td->io_ops->flags & FIO_CPUIO)
589 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
591 f = fopen(tmp, "r+");
593 td_verror(td, errno);
600 ret = fwrite(td->ioscheduler, strlen(td->ioscheduler), 1, f);
601 if (ferror(f) || ret != 1) {
602 td_verror(td, errno);
610 * Read back and check that the selected scheduler is now the default.
612 ret = fread(tmp, 1, sizeof(tmp), f);
613 if (ferror(f) || ret < 0) {
614 td_verror(td, errno);
619 sprintf(tmp2, "[%s]", td->ioscheduler);
620 if (!strstr(tmp, tmp2)) {
621 log_err("fio: io scheduler %s not found\n", td->ioscheduler);
622 td_verror(td, EINVAL);
631 static void clear_io_state(struct thread_data *td)
636 td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0;
637 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
640 for_each_file(td, f, i) {
642 if (td->io_ops->flags & FIO_SYNCIO)
643 lseek(f->fd, SEEK_SET, 0);
646 memset(f->file_map, 0, f->num_maps * sizeof(long));
651 * Entry point for the thread based jobs. The process based jobs end up
652 * here as well, after a little setup.
654 static void *thread_main(void *data)
656 unsigned long long runtime[2];
657 struct thread_data *td = data;
664 INIT_LIST_HEAD(&td->io_u_freelist);
665 INIT_LIST_HEAD(&td->io_u_busylist);
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 getrusage(RUSAGE_SELF, &td->ru_start);
722 runtime[0] = runtime[1] = 0;
723 while (td->loops--) {
724 fio_gettime(&td->start, NULL);
725 memcpy(&td->stat_sample_time, &td->start, sizeof(td->start));
728 memcpy(&td->lastrate, &td->stat_sample_time, sizeof(td->lastrate));
731 prune_io_piece_log(td);
733 if (td->io_ops->flags & FIO_CPUIO)
738 runtime[td->ddir] += utime_since_now(&td->start);
739 if (td_rw(td) && td->io_bytes[td->ddir ^ 1])
740 runtime[td->ddir ^ 1] = runtime[td->ddir];
742 if (td->error || td->terminate)
745 if (td->verify == VERIFY_NONE)
749 fio_gettime(&td->start, NULL);
753 runtime[DDIR_READ] += utime_since_now(&td->start);
755 if (td->error || td->terminate)
759 update_rusage_stat(td);
760 fio_gettime(&td->end_time, NULL);
761 td->runtime[0] = runtime[0] / 1000;
762 td->runtime[1] = runtime[1] / 1000;
765 finish_log(td, td->bw_log, "bw");
767 finish_log(td, td->slat_log, "slat");
769 finish_log(td, td->clat_log, "clat");
770 if (td->write_iolog_file)
771 write_iolog_close(td);
772 if (td->exec_postrun) {
773 if (system(td->exec_postrun) < 0)
774 log_err("fio: postrun %s failed\n", td->exec_postrun);
777 if (exitall_on_terminate)
778 terminate_threads(td->groupid, 0);
782 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
786 td_set_runstate(td, TD_EXITED);
787 return (void *) td->error;
791 * We cannot pass the td data into a forked process, so attach the td and
792 * pass it to the thread worker.
794 static int fork_main(int shmid, int offset)
796 struct thread_data *td;
799 data = shmat(shmid, NULL, 0);
800 if (data == (void *) -1) {
807 td = data + offset * sizeof(struct thread_data);
808 ret = thread_main(td);
814 * Run over the job map and reap the threads that have exited, if any.
816 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
818 struct thread_data *td;
819 int i, cputhreads, pending, status, ret;
822 * reap exited threads (TD_EXITED -> TD_REAPED)
824 pending = cputhreads = 0;
827 * ->io_ops is NULL for a thread that has closed its
830 if (td->io_ops && td->io_ops->flags & FIO_CPUIO)
833 if (td->runstate < TD_EXITED) {
835 * check if someone quit or got killed in an unusual way
837 ret = waitpid(td->pid, &status, WNOHANG);
840 else if ((ret == td->pid) && WIFSIGNALED(status)) {
841 int sig = WTERMSIG(status);
843 log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
844 td_set_runstate(td, TD_REAPED);
849 if (td->runstate != TD_EXITED) {
850 if (td->runstate < TD_RUNNING)
859 td_set_runstate(td, TD_REAPED);
861 if (td->use_thread) {
864 if (pthread_join(td->thread, (void *) &ret))
865 perror("thread_join");
869 ret = waitpid(td->pid, &status, 0);
872 else if (WIFEXITED(status) && WEXITSTATUS(status)) {
880 (*m_rate) -= td->ratemin;
881 (*t_rate) -= td->rate;
884 if (*nr_running == cputhreads && !pending)
885 terminate_threads(TERMINATE_ALL, 0);
889 * Main function for kicking off and reaping jobs, as needed.
891 static void run_threads(void)
893 struct thread_data *td;
895 int i, todo, nr_running, m_rate, t_rate, nr_started;
897 if (fio_pin_memory())
901 printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
905 signal(SIGINT, sig_handler);
906 signal(SIGALRM, sig_handler);
908 todo = thread_number;
914 print_status_init(td->thread_number - 1);
916 if (!td->create_serialize) {
922 * do file setup here so it happens sequentially,
923 * we don't want X number of threads getting their
924 * client data interspersed on disk
926 if (setup_files(td)) {
929 log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
930 td_set_runstate(td, TD_REAPED);
938 struct thread_data *map[MAX_JOBS];
939 struct timeval this_start;
940 int this_jobs = 0, left;
943 * create threads (TD_NOT_CREATED -> TD_CREATED)
946 if (td->runstate != TD_NOT_CREATED)
950 * never got a chance to start, killed by other
951 * thread for some reason
958 if (td->start_delay) {
959 spent = mtime_since_genesis();
961 if (td->start_delay * 1000 > spent)
965 if (td->stonewall && (nr_started || nr_running))
969 * Set state to created. Thread will transition
970 * to TD_INITIALIZED when it's done setting up.
972 td_set_runstate(td, TD_CREATED);
973 map[this_jobs++] = td;
974 fio_sem_init(&startup_sem, 1);
977 if (td->use_thread) {
978 if (pthread_create(&td->thread, NULL, thread_main, td)) {
979 perror("thread_create");
984 fio_sem_down(&startup_sem);
986 int ret = fork_main(shm_id, i);
994 * Wait for the started threads to transition to
997 fio_gettime(&this_start, NULL);
999 while (left && !fio_abort) {
1000 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1005 for (i = 0; i < this_jobs; i++) {
1009 if (td->runstate == TD_INITIALIZED) {
1012 } else if (td->runstate >= TD_EXITED) {
1016 nr_running++; /* work-around... */
1022 log_err("fio: %d jobs failed to start\n", left);
1023 for (i = 0; i < this_jobs; i++) {
1027 kill(td->pid, SIGTERM);
1033 * start created threads (TD_INITIALIZED -> TD_RUNNING).
1035 for_each_td(td, i) {
1036 if (td->runstate != TD_INITIALIZED)
1039 td_set_runstate(td, TD_RUNNING);
1042 m_rate += td->ratemin;
1045 fio_sem_up(&td->mutex);
1048 reap_threads(&nr_running, &t_rate, &m_rate);
1054 while (nr_running) {
1055 reap_threads(&nr_running, &t_rate, &m_rate);
1063 int main(int argc, char *argv[])
1068 * We need locale for number printing, if it isn't set then just
1069 * go with the US format.
1071 if (!getenv("LC_NUMERIC"))
1072 setlocale(LC_NUMERIC, "en_US");
1074 if (parse_options(argc, argv))
1077 if (!thread_number) {
1078 log_err("Nothing to do\n");
1082 ps = sysconf(_SC_PAGESIZE);
1084 log_err("Failed to get page size\n");
1091 setup_log(&agg_io_log[DDIR_READ]);
1092 setup_log(&agg_io_log[DDIR_WRITE]);
1095 disk_util_timer_arm();
1102 __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1103 __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");