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)
99 unsigned long long bytes = 0;
104 * allow a 2 second settle period in the beginning
106 if (mtime_since(&td->start, now) < 2000)
110 bytes += td->this_io_bytes[DDIR_READ];
112 bytes += td->this_io_bytes[DDIR_WRITE];
115 * if rate blocks is set, sample is running
117 if (td->rate_bytes) {
118 spent = mtime_since(&td->lastrate, now);
119 if (spent < td->ratecycle)
122 if (bytes < td->rate_bytes) {
123 fprintf(f_out, "%s: min rate %u not met\n", td->name, td->ratemin);
126 rate = (bytes - td->rate_bytes) / spent;
127 if (rate < td->ratemin || bytes < td->rate_bytes) {
128 fprintf(f_out, "%s: min rate %u not met, got %luKiB/sec\n", td->name, td->ratemin, rate);
134 td->rate_bytes = bytes;
135 memcpy(&td->lastrate, now, sizeof(*now));
139 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
143 if (mtime_since(&td->epoch, t) >= td->timeout * 1000)
150 * When job exits, we can cancel the in-flight IO if we are using async
151 * io. Attempt to do so.
153 static void cleanup_pending_aio(struct thread_data *td)
155 struct list_head *entry, *n;
160 * get immediately available events, if any
162 r = io_u_queued_complete(td, 0);
167 * now cancel remaining active events
169 if (td->io_ops->cancel) {
170 list_for_each_safe(entry, n, &td->io_u_busylist) {
171 io_u = list_entry(entry, struct io_u, list);
174 * if the io_u isn't in flight, then that generally
175 * means someone leaked an io_u. complain but fix
176 * it up, so we don't stall here.
178 if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
179 log_err("fio: non-busy IO on busy list\n");
182 r = td->io_ops->cancel(td, io_u);
190 r = io_u_queued_complete(td, td->cur_depth);
194 * Helper to handle the final sync of a file. Works just like the normal
195 * io path, just does everything sync.
197 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
199 struct io_u *io_u = __get_io_u(td);
205 io_u->ddir = DDIR_SYNC;
208 if (td_io_prep(td, io_u)) {
214 ret = td_io_queue(td, io_u);
216 td_verror(td, io_u->error, "td_io_queue");
219 } else if (ret == FIO_Q_QUEUED) {
220 if (io_u_queued_complete(td, 1) < 0)
222 } else if (ret == FIO_Q_COMPLETED) {
224 td_verror(td, io_u->error, "td_io_queue");
228 if (io_u_sync_complete(td, io_u) < 0)
230 } else if (ret == FIO_Q_BUSY) {
231 if (td_io_commit(td))
240 * The main verify engine. Runs over the writes we previusly submitted,
241 * reads the blocks back in, and checks the crc/md5 of the data.
243 static void do_verify(struct thread_data *td)
247 int ret, i, min_events;
250 * sync io first and invalidate cache, to make sure we really
253 for_each_file(td, f, i) {
254 if (fio_io_sync(td, f))
256 if (file_invalidate_cache(td, f))
263 td_set_runstate(td, TD_VERIFYING);
266 while (!td->terminate) {
267 io_u = __get_io_u(td);
271 if (runtime_exceeded(td, &io_u->start_time)) {
276 if (get_next_verify(td, io_u)) {
281 if (td_io_prep(td, io_u)) {
286 io_u->end_io = verify_io_u;
288 ret = td_io_queue(td, io_u);
291 case FIO_Q_COMPLETED:
294 if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
295 int bytes = io_u->xfer_buflen - io_u->resid;
297 io_u->xfer_buflen = io_u->resid;
298 io_u->xfer_buf += bytes;
301 ret = io_u_sync_complete(td, io_u);
308 requeue_io_u(td, &io_u);
309 ret = td_io_commit(td);
313 td_verror(td, -ret, "td_io_queue");
317 if (ret < 0 || td->error)
321 * if we can queue more, do so. but check if there are
322 * completed io_u's first.
325 if (queue_full(td) || ret == FIO_Q_BUSY) {
328 if (td->cur_depth > td->iodepth_low)
329 min_events = td->cur_depth - td->iodepth_low;
333 * Reap required number of io units, if any, and do the
334 * verification on them through the callback handler
336 if (io_u_queued_complete(td, min_events) < 0)
341 min_events = td->cur_depth;
344 ret = io_u_queued_complete(td, min_events);
346 cleanup_pending_aio(td);
348 td_set_runstate(td, TD_RUNNING);
352 * Not really an io thread, all it does is burn CPU cycles in the specified
355 static void do_cpuio(struct thread_data *td)
358 int split = 100 / td->cpuload;
361 while (!td->terminate) {
362 fio_gettime(&e, NULL);
364 if (runtime_exceeded(td, &e))
370 usec_sleep(td, 10000);
377 * Main IO worker function. It retrieves io_u's to process and queues
378 * and reaps them, checking for rate and errors along the way.
380 static void do_io(struct thread_data *td)
386 td_set_runstate(td, TD_RUNNING);
388 while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->io_size) {
389 struct timeval comp_time;
401 memcpy(&s, &io_u->start_time, sizeof(s));
403 if (runtime_exceeded(td, &s)) {
408 ret = td_io_queue(td, io_u);
411 case FIO_Q_COMPLETED:
416 if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
417 int bytes = io_u->xfer_buflen - io_u->resid;
419 io_u->xfer_buflen = io_u->resid;
420 io_u->xfer_buf += bytes;
423 fio_gettime(&comp_time, NULL);
424 bytes_done = io_u_sync_complete(td, io_u);
430 * if the engine doesn't have a commit hook,
431 * the io_u is really queued. if it does have such
432 * a hook, it has to call io_u_queued() itself.
434 if (td->io_ops->commit == NULL)
435 io_u_queued(td, io_u);
438 requeue_io_u(td, &io_u);
439 ret = td_io_commit(td);
447 if (ret < 0 || td->error)
451 * See if we need to complete some commands
453 if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
455 if (queue_full(td) || ret == FIO_Q_BUSY) {
458 if (td->cur_depth > td->iodepth_low)
459 min_evts = td->cur_depth - td->iodepth_low;
462 fio_gettime(&comp_time, NULL);
463 bytes_done = io_u_queued_complete(td, min_evts);
472 * the rate is batched for now, it should work for batches
473 * of completions except the very first one which may look
476 usec = utime_since(&s, &comp_time);
478 rate_throttle(td, usec, bytes_done);
480 if (check_min_rate(td, &comp_time)) {
481 if (exitall_on_terminate)
482 terminate_threads(td->groupid, 0);
483 td_verror(td, ENODATA, "check_min_rate");
488 unsigned long long b;
490 b = td->io_blocks[0] + td->io_blocks[1];
491 if (!(b % td->thinktime_blocks)) {
494 if (td->thinktime_spin)
495 __usec_sleep(td->thinktime_spin);
497 left = td->thinktime - td->thinktime_spin;
499 usec_sleep(td, left);
509 ret = io_u_queued_complete(td, i);
511 if (should_fsync(td) && td->end_fsync) {
512 td_set_runstate(td, TD_FSYNCING);
513 for_each_file(td, f, i)
517 cleanup_pending_aio(td);
520 static void cleanup_io_u(struct thread_data *td)
522 struct list_head *entry, *n;
525 list_for_each_safe(entry, n, &td->io_u_freelist) {
526 io_u = list_entry(entry, struct io_u, list);
528 list_del(&io_u->list);
536 * "randomly" fill the buffer contents
538 static void fill_rand_buf(struct io_u *io_u, int max_bs)
540 int *ptr = io_u->buf;
542 while ((void *) ptr - io_u->buf < max_bs) {
543 *ptr = rand() * 0x9e370001;
548 static int init_io_u(struct thread_data *td)
555 if (td->io_ops->flags & FIO_CPUIO)
558 if (td->io_ops->flags & FIO_SYNCIO)
561 max_units = td->iodepth;
563 max_bs = max(td->max_bs[DDIR_READ], td->max_bs[DDIR_WRITE]);
564 td->orig_buffer_size = max_bs * max_units;
566 if (td->mem_type == MEM_SHMHUGE || td->mem_type == MEM_MMAPHUGE)
567 td->orig_buffer_size = (td->orig_buffer_size + td->hugepage_size - 1) & ~(td->hugepage_size - 1);
569 td->orig_buffer_size += page_mask;
571 if (allocate_io_mem(td))
574 p = ALIGN(td->orig_buffer);
575 for (i = 0; i < max_units; i++) {
576 io_u = malloc(sizeof(*io_u));
577 memset(io_u, 0, sizeof(*io_u));
578 INIT_LIST_HEAD(&io_u->list);
580 io_u->buf = p + max_bs * i;
581 if (td_write(td) || td_rw(td))
582 fill_rand_buf(io_u, max_bs);
585 io_u->flags = IO_U_F_FREE;
586 list_add(&io_u->list, &td->io_u_freelist);
594 static int switch_ioscheduler(struct thread_data *td)
596 char tmp[256], tmp2[128];
600 if (td->io_ops->flags & FIO_CPUIO)
603 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
605 f = fopen(tmp, "r+");
607 td_verror(td, errno, "fopen");
614 ret = fwrite(td->ioscheduler, strlen(td->ioscheduler), 1, f);
615 if (ferror(f) || ret != 1) {
616 td_verror(td, errno, "fwrite");
624 * Read back and check that the selected scheduler is now the default.
626 ret = fread(tmp, 1, sizeof(tmp), f);
627 if (ferror(f) || ret < 0) {
628 td_verror(td, errno, "fread");
633 sprintf(tmp2, "[%s]", td->ioscheduler);
634 if (!strstr(tmp, tmp2)) {
635 log_err("fio: io scheduler %s not found\n", td->ioscheduler);
636 td_verror(td, EINVAL, "iosched_switch");
645 static void clear_io_state(struct thread_data *td)
650 td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
651 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
654 td->last_was_sync = 0;
656 for_each_file(td, f, i) {
657 f->last_completed_pos = 0;
660 if (td->io_ops->flags & FIO_SYNCIO)
661 lseek(f->fd, SEEK_SET, 0);
664 memset(f->file_map, 0, f->num_maps * sizeof(long));
669 * Entry point for the thread based jobs. The process based jobs end up
670 * here as well, after a little setup.
672 static void *thread_main(void *data)
674 unsigned long long runtime[2];
675 struct thread_data *td = data;
682 INIT_LIST_HEAD(&td->io_u_freelist);
683 INIT_LIST_HEAD(&td->io_u_busylist);
684 INIT_LIST_HEAD(&td->io_u_requeues);
685 INIT_LIST_HEAD(&td->io_hist_list);
686 INIT_LIST_HEAD(&td->io_log_list);
691 if (fio_setaffinity(td) == -1) {
692 td_verror(td, errno, "cpu_set_affinity");
700 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
701 td_verror(td, errno, "ioprio_set");
706 if (nice(td->nice) == -1) {
707 td_verror(td, errno, "nice");
711 if (init_random_state(td))
714 if (td->ioscheduler && switch_ioscheduler(td))
717 td_set_runstate(td, TD_INITIALIZED);
718 fio_sem_up(&startup_sem);
719 fio_sem_down(&td->mutex);
721 if (!td->create_serialize && setup_files(td))
727 * Do this late, as some IO engines would like to have the
728 * files setup prior to initializing structures.
733 if (td->exec_prerun) {
734 if (system(td->exec_prerun) < 0)
738 fio_gettime(&td->epoch, NULL);
739 memcpy(&td->timeout_end, &td->epoch, sizeof(td->epoch));
740 getrusage(RUSAGE_SELF, &td->ts.ru_start);
742 runtime[0] = runtime[1] = 0;
743 while (td->loops--) {
744 fio_gettime(&td->start, NULL);
745 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
748 memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
751 prune_io_piece_log(td);
753 if (td->io_ops->flags & FIO_CPUIO)
758 if (td_read(td) && td->io_bytes[DDIR_READ])
759 runtime[DDIR_READ] += utime_since_now(&td->start);
760 if (td_write(td) && td->io_bytes[DDIR_WRITE])
761 runtime[DDIR_WRITE] += utime_since_now(&td->start);
763 if (td->error || td->terminate)
766 if (td->verify == VERIFY_NONE)
770 fio_gettime(&td->start, NULL);
774 runtime[DDIR_READ] += utime_since_now(&td->start);
776 if (td->error || td->terminate)
780 update_rusage_stat(td);
781 fio_gettime(&td->end_time, NULL);
782 td->runtime[0] = runtime[0] / 1000;
783 td->runtime[1] = runtime[1] / 1000;
786 finish_log(td, td->ts.bw_log, "bw");
788 finish_log(td, td->ts.slat_log, "slat");
790 finish_log(td, td->ts.clat_log, "clat");
791 if (td->write_iolog_file)
792 write_iolog_close(td);
793 if (td->exec_postrun) {
794 if (system(td->exec_postrun) < 0)
795 log_err("fio: postrun %s failed\n", td->exec_postrun);
798 if (exitall_on_terminate)
799 terminate_threads(td->groupid, 0);
803 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
807 td_set_runstate(td, TD_EXITED);
808 return (void *) (unsigned long) td->error;
812 * We cannot pass the td data into a forked process, so attach the td and
813 * pass it to the thread worker.
815 static int fork_main(int shmid, int offset)
817 struct thread_data *td;
820 data = shmat(shmid, NULL, 0);
821 if (data == (void *) -1) {
828 td = data + offset * sizeof(struct thread_data);
829 ret = thread_main(td);
831 return (int) (unsigned long) ret;
835 * Run over the job map and reap the threads that have exited, if any.
837 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
839 struct thread_data *td;
840 int i, cputhreads, pending, status, ret;
843 * reap exited threads (TD_EXITED -> TD_REAPED)
845 pending = cputhreads = 0;
850 * ->io_ops is NULL for a thread that has closed its
853 if (td->io_ops && td->io_ops->flags & FIO_CPUIO)
856 if (!td->pid || td->runstate == TD_REAPED)
858 if (td->use_thread) {
859 if (td->runstate == TD_EXITED) {
860 td_set_runstate(td, TD_REAPED);
867 if (td->runstate == TD_EXITED)
871 * check if someone quit or got killed in an unusual way
873 ret = waitpid(td->pid, &status, flags);
875 if (errno == ECHILD) {
876 log_err("fio: pid=%d disappeared %d\n", td->pid, td->runstate);
877 td_set_runstate(td, TD_REAPED);
881 } else if (ret == td->pid) {
882 if (WIFSIGNALED(status)) {
883 int sig = WTERMSIG(status);
885 log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
886 td_set_runstate(td, TD_REAPED);
889 if (WIFEXITED(status)) {
890 if (WEXITSTATUS(status) && !td->error)
891 td->error = WEXITSTATUS(status);
893 td_set_runstate(td, TD_REAPED);
899 * thread is not dead, continue
903 if (td->use_thread) {
906 if (pthread_join(td->thread, (void *) &ret))
907 perror("pthread_join");
911 (*m_rate) -= td->ratemin;
912 (*t_rate) -= td->rate;
918 if (*nr_running == cputhreads && !pending)
919 terminate_threads(TERMINATE_ALL, 0);
923 * Main function for kicking off and reaping jobs, as needed.
925 static void run_threads(void)
927 struct thread_data *td;
929 int i, todo, nr_running, m_rate, t_rate, nr_started;
931 if (fio_pin_memory())
935 printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
939 signal(SIGINT, sig_handler);
940 signal(SIGALRM, sig_handler);
942 todo = thread_number;
948 print_status_init(td->thread_number - 1);
950 if (!td->create_serialize) {
956 * do file setup here so it happens sequentially,
957 * we don't want X number of threads getting their
958 * client data interspersed on disk
960 if (setup_files(td)) {
963 log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
964 td_set_runstate(td, TD_REAPED);
974 struct thread_data *map[MAX_JOBS];
975 struct timeval this_start;
976 int this_jobs = 0, left;
979 * create threads (TD_NOT_CREATED -> TD_CREATED)
982 if (td->runstate != TD_NOT_CREATED)
986 * never got a chance to start, killed by other
987 * thread for some reason
994 if (td->start_delay) {
995 spent = mtime_since_genesis();
997 if (td->start_delay * 1000 > spent)
1001 if (td->stonewall && (nr_started || nr_running))
1005 * Set state to created. Thread will transition
1006 * to TD_INITIALIZED when it's done setting up.
1008 td_set_runstate(td, TD_CREATED);
1009 map[this_jobs++] = td;
1010 fio_sem_init(&startup_sem, 1);
1013 if (td->use_thread) {
1014 if (pthread_create(&td->thread, NULL, thread_main, td)) {
1015 perror("thread_create");
1020 fio_sem_down(&startup_sem);
1022 int ret = fork_main(shm_id, i);
1030 * Wait for the started threads to transition to
1033 fio_gettime(&this_start, NULL);
1035 while (left && !fio_abort) {
1036 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1041 for (i = 0; i < this_jobs; i++) {
1045 if (td->runstate == TD_INITIALIZED) {
1048 } else if (td->runstate >= TD_EXITED) {
1052 nr_running++; /* work-around... */
1058 log_err("fio: %d jobs failed to start\n", left);
1059 for (i = 0; i < this_jobs; i++) {
1063 kill(td->pid, SIGTERM);
1069 * start created threads (TD_INITIALIZED -> TD_RUNNING).
1071 for_each_td(td, i) {
1072 if (td->runstate != TD_INITIALIZED)
1075 td_set_runstate(td, TD_RUNNING);
1078 m_rate += td->ratemin;
1081 fio_sem_up(&td->mutex);
1084 reap_threads(&nr_running, &t_rate, &m_rate);
1090 while (nr_running) {
1091 reap_threads(&nr_running, &t_rate, &m_rate);
1099 int main(int argc, char *argv[])
1104 * We need locale for number printing, if it isn't set then just
1105 * go with the US format.
1107 if (!getenv("LC_NUMERIC"))
1108 setlocale(LC_NUMERIC, "en_US");
1110 if (parse_options(argc, argv))
1113 if (!thread_number) {
1114 log_err("Nothing to do\n");
1118 ps = sysconf(_SC_PAGESIZE);
1120 log_err("Failed to get page size\n");
1127 setup_log(&agg_io_log[DDIR_READ]);
1128 setup_log(&agg_io_log[DDIR_WRITE]);
1133 disk_util_timer_arm();
1140 __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1141 __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");