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
34 #include <sys/ioctl.h>
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;
52 struct io_log *agg_io_log[2];
54 #define TERMINATE_ALL (-1)
55 #define JOB_START_TIMEOUT (5 * 1000)
57 static inline void td_set_runstate(struct thread_data *td, int runstate)
59 td->runstate = runstate;
62 static void terminate_threads(int group_id, int forced_kill)
64 struct thread_data *td;
68 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
72 td_set_runstate(td, TD_EXITED);
77 static void sig_handler(int sig)
82 disk_util_timer_arm();
83 print_thread_status();
86 fprintf(stderr, "fio: got segfault, aborting\n");
87 terminate_threads(TERMINATE_ALL, 1);
91 printf("\nfio: terminating on signal %d\n", sig);
93 terminate_threads(TERMINATE_ALL, 0);
99 * Check if we are above the minimum rate given.
101 static int check_min_rate(struct thread_data *td, struct timeval *now)
108 * allow a 2 second settle period in the beginning
110 if (mtime_since(&td->start, now) < 2000)
114 * if rate blocks is set, sample is running
116 if (td->rate_bytes) {
117 spent = mtime_since(&td->lastrate, now);
118 if (spent < td->ratecycle)
121 rate = (td->this_io_bytes[ddir] - td->rate_bytes) / spent;
122 if (rate < td->ratemin) {
123 fprintf(f_out, "%s: min rate %u not met, got %luKiB/sec\n", td->name, td->ratemin, rate);
128 td->rate_bytes = td->this_io_bytes[ddir];
129 memcpy(&td->lastrate, now, sizeof(*now));
133 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
137 if (mtime_since(&td->epoch, t) >= td->timeout * 1000)
143 static struct fio_file *get_next_file(struct thread_data *td)
145 unsigned int old_next_file = td->next_file;
149 f = &td->files[td->next_file];
152 if (td->next_file >= td->nr_files)
159 } while (td->next_file != old_next_file);
165 * When job exits, we can cancel the in-flight IO if we are using async
166 * io. Attempt to do so.
168 static void cleanup_pending_aio(struct thread_data *td)
170 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
171 struct list_head *entry, *n;
172 struct io_completion_data icd;
177 * get immediately available events, if any
179 r = td_io_getevents(td, 0, td->cur_depth, &ts);
182 ios_completed(td, &icd);
186 * now cancel remaining active events
188 if (td->io_ops->cancel) {
189 list_for_each_safe(entry, n, &td->io_u_busylist) {
190 io_u = list_entry(entry, struct io_u, list);
192 r = td->io_ops->cancel(td, io_u);
199 r = td_io_getevents(td, td->cur_depth, td->cur_depth, NULL);
202 ios_completed(td, &icd);
208 * Helper to handle the final sync of a file. Works just like the normal
209 * io path, just does everything sync.
211 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
213 struct io_u *io_u = __get_io_u(td);
214 struct io_completion_data icd;
220 io_u->ddir = DDIR_SYNC;
223 if (td_io_prep(td, io_u)) {
228 ret = td_io_queue(td, io_u);
230 td_verror(td, io_u->error);
235 ret = td_io_getevents(td, 1, td->cur_depth, NULL);
242 ios_completed(td, &icd);
244 td_verror(td, icd.error);
252 * The main verify engine. Runs over the writes we previusly submitted,
253 * reads the blocks back in, and checks the crc/md5 of the data.
255 static void do_verify(struct thread_data *td)
257 struct io_u *io_u, *v_io_u = NULL;
258 struct io_completion_data icd;
263 * sync io first and invalidate cache, to make sure we really
266 for_each_file(td, f, i) {
268 file_invalidate_cache(td, f);
271 td_set_runstate(td, TD_VERIFYING);
277 io_u = __get_io_u(td);
281 if (runtime_exceeded(td, &io_u->start_time)) {
286 if (get_next_verify(td, io_u)) {
291 f = get_next_file(td);
297 if (td_io_prep(td, io_u)) {
302 ret = td_io_queue(td, io_u);
304 td_verror(td, io_u->error);
310 * we have one pending to verify, do that while
311 * we are doing io on the next one
313 if (do_io_u_verify(td, &v_io_u))
316 ret = td_io_getevents(td, 1, 1, NULL);
323 v_io_u = td->io_ops->event(td, 0);
326 fio_gettime(&icd.time, NULL);
327 io_completed(td, v_io_u, &icd);
330 td_verror(td, icd.error);
331 put_io_u(td, v_io_u);
337 * if we can't submit more io, we need to verify now
339 if (queue_full(td) && do_io_u_verify(td, &v_io_u))
344 do_io_u_verify(td, &v_io_u);
347 cleanup_pending_aio(td);
349 td_set_runstate(td, TD_RUNNING);
353 * Not really an io thread, all it does is burn CPU cycles in the specified
356 static void do_cpuio(struct thread_data *td)
359 int split = 100 / td->cpuload;
362 while (!td->terminate) {
363 fio_gettime(&e, NULL);
365 if (runtime_exceeded(td, &e))
371 usec_sleep(td, 10000);
378 * Main IO worker function. It retrieves io_u's to process and queues
379 * and reaps them, checking for rate and errors along the way.
381 static void do_io(struct thread_data *td)
383 struct io_completion_data icd;
389 td_set_runstate(td, TD_RUNNING);
391 while (td->this_io_bytes[td->ddir] < td->io_size) {
392 struct timespec *timeout;
399 f = get_next_file(td);
403 io_u = get_io_u(td, f);
407 memcpy(&s, &io_u->start_time, sizeof(s));
409 ret = td_io_queue(td, io_u);
411 td_verror(td, io_u->error);
416 add_slat_sample(td, io_u->ddir, mtime_since(&io_u->start_time, &io_u->issue_time));
418 if (td->cur_depth < td->iodepth) {
419 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
428 ret = td_io_getevents(td, min_evts, td->cur_depth, timeout);
436 ios_completed(td, &icd);
438 td_verror(td, icd.error);
443 * the rate is batched for now, it should work for batches
444 * of completions except the very first one which may look
447 usec = utime_since(&s, &icd.time);
449 rate_throttle(td, usec, icd.bytes_done[td->ddir], td->ddir);
451 if (check_min_rate(td, &icd.time)) {
452 if (exitall_on_terminate)
453 terminate_threads(td->groupid, 0);
454 td_verror(td, ENODATA);
458 if (runtime_exceeded(td, &icd.time))
462 unsigned long long b;
464 b = td->io_blocks[0] + td->io_blocks[1];
465 if (!(b % td->thinktime_blocks))
466 usec_sleep(td, td->thinktime);
472 cleanup_pending_aio(td);
474 if (should_fsync(td) && td->end_fsync) {
475 td_set_runstate(td, TD_FSYNCING);
476 for_each_file(td, f, i)
482 static void cleanup_io_u(struct thread_data *td)
484 struct list_head *entry, *n;
487 list_for_each_safe(entry, n, &td->io_u_freelist) {
488 io_u = list_entry(entry, struct io_u, list);
490 list_del(&io_u->list);
498 * "randomly" fill the buffer contents
500 static void fill_rand_buf(struct io_u *io_u, int max_bs)
502 int *ptr = io_u->buf;
504 while ((void *) ptr - io_u->buf < max_bs) {
505 *ptr = rand() * 0x9e370001;
510 static int init_io_u(struct thread_data *td)
517 if (td->io_ops->flags & FIO_CPUIO)
520 if (td->io_ops->flags & FIO_SYNCIO)
523 max_units = td->iodepth;
525 max_bs = max(td->max_bs[DDIR_READ], td->max_bs[DDIR_WRITE]);
526 td->orig_buffer_size = max_bs * max_units;
528 if (td->mem_type == MEM_SHMHUGE || td->mem_type == MEM_MMAPHUGE)
529 td->orig_buffer_size = (td->orig_buffer_size + td->hugepage_size - 1) & ~(td->hugepage_size - 1);
531 td->orig_buffer_size += page_mask;
533 if (allocate_io_mem(td))
536 p = ALIGN(td->orig_buffer);
537 for (i = 0; i < max_units; i++) {
538 io_u = malloc(sizeof(*io_u));
539 memset(io_u, 0, sizeof(*io_u));
540 INIT_LIST_HEAD(&io_u->list);
542 io_u->buf = p + max_bs * i;
543 if (td_write(td) || td_rw(td))
544 fill_rand_buf(io_u, max_bs);
547 list_add(&io_u->list, &td->io_u_freelist);
553 static int switch_ioscheduler(struct thread_data *td)
555 char tmp[256], tmp2[128];
559 if (td->io_ops->flags & FIO_CPUIO)
562 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
564 f = fopen(tmp, "r+");
566 td_verror(td, errno);
573 ret = fwrite(td->ioscheduler, strlen(td->ioscheduler), 1, f);
574 if (ferror(f) || ret != 1) {
575 td_verror(td, errno);
583 * Read back and check that the selected scheduler is now the default.
585 ret = fread(tmp, 1, sizeof(tmp), f);
586 if (ferror(f) || ret < 0) {
587 td_verror(td, errno);
592 sprintf(tmp2, "[%s]", td->ioscheduler);
593 if (!strstr(tmp, tmp2)) {
594 log_err("fio: io scheduler %s not found\n", td->ioscheduler);
595 td_verror(td, EINVAL);
604 static void clear_io_state(struct thread_data *td)
609 td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0;
610 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
613 for_each_file(td, f, i) {
615 if (td->io_ops->flags & FIO_SYNCIO)
616 lseek(f->fd, SEEK_SET, 0);
619 memset(f->file_map, 0, f->num_maps * sizeof(long));
624 * Entry point for the thread based jobs. The process based jobs end up
625 * here as well, after a little setup.
627 static void *thread_main(void *data)
629 unsigned long long runtime[2];
630 struct thread_data *td = data;
637 INIT_LIST_HEAD(&td->io_u_freelist);
638 INIT_LIST_HEAD(&td->io_u_busylist);
639 INIT_LIST_HEAD(&td->io_hist_list);
640 INIT_LIST_HEAD(&td->io_log_list);
645 if (fio_setaffinity(td) == -1) {
646 td_verror(td, errno);
654 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
655 td_verror(td, errno);
660 if (nice(td->nice) == -1) {
661 td_verror(td, errno);
665 if (init_random_state(td))
668 if (td->ioscheduler && switch_ioscheduler(td))
671 td_set_runstate(td, TD_INITIALIZED);
672 fio_sem_up(&startup_sem);
673 fio_sem_down(&td->mutex);
675 if (!td->create_serialize && setup_files(td))
681 * Do this late, as some IO engines would like to have the
682 * files setup prior to initializing structures.
688 system(td->exec_prerun);
690 fio_gettime(&td->epoch, NULL);
691 getrusage(RUSAGE_SELF, &td->ru_start);
693 runtime[0] = runtime[1] = 0;
694 while (td->loops--) {
695 fio_gettime(&td->start, NULL);
696 memcpy(&td->stat_sample_time, &td->start, sizeof(td->start));
699 memcpy(&td->lastrate, &td->stat_sample_time, sizeof(td->lastrate));
702 prune_io_piece_log(td);
704 if (td->io_ops->flags & FIO_CPUIO)
709 runtime[td->ddir] += utime_since_now(&td->start);
710 if (td_rw(td) && td->io_bytes[td->ddir ^ 1])
711 runtime[td->ddir ^ 1] = runtime[td->ddir];
713 if (td->error || td->terminate)
716 if (td->verify == VERIFY_NONE)
720 fio_gettime(&td->start, NULL);
724 runtime[DDIR_READ] += utime_since_now(&td->start);
726 if (td->error || td->terminate)
730 update_rusage_stat(td);
731 fio_gettime(&td->end_time, NULL);
732 td->runtime[0] = runtime[0] / 1000;
733 td->runtime[1] = runtime[1] / 1000;
736 finish_log(td, td->bw_log, "bw");
738 finish_log(td, td->slat_log, "slat");
740 finish_log(td, td->clat_log, "clat");
741 if (td->write_iolog_file)
742 write_iolog_close(td);
743 if (td->exec_postrun)
744 system(td->exec_postrun);
746 if (exitall_on_terminate)
747 terminate_threads(td->groupid, 0);
753 td_set_runstate(td, TD_EXITED);
759 * We cannot pass the td data into a forked process, so attach the td and
760 * pass it to the thread worker.
762 static void *fork_main(int shmid, int offset)
764 struct thread_data *td;
767 data = shmat(shmid, NULL, 0);
768 if (data == (void *) -1) {
773 td = data + offset * sizeof(struct thread_data);
780 * Run over the job map and reap the threads that have exited, if any.
782 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
784 struct thread_data *td;
785 int i, cputhreads, pending;
788 * reap exited threads (TD_EXITED -> TD_REAPED)
790 pending = cputhreads = 0;
793 * ->io_ops is NULL for a thread that has closed its
796 if (td->io_ops && td->io_ops->flags & FIO_CPUIO)
799 if (td->runstate != TD_EXITED) {
800 if (td->runstate < TD_RUNNING)
806 td_set_runstate(td, TD_REAPED);
808 if (td->use_thread) {
811 if (pthread_join(td->thread, (void *) &ret))
812 perror("thread_join");
814 waitpid(td->pid, NULL, 0);
817 (*m_rate) -= td->ratemin;
818 (*t_rate) -= td->rate;
821 if (*nr_running == cputhreads && !pending)
822 terminate_threads(TERMINATE_ALL, 0);
826 * Main function for kicking off and reaping jobs, as needed.
828 static void run_threads(void)
830 struct thread_data *td;
832 int i, todo, nr_running, m_rate, t_rate, nr_started;
834 if (fio_pin_memory())
838 printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
842 signal(SIGINT, sig_handler);
843 signal(SIGALRM, sig_handler);
845 todo = thread_number;
851 print_status_init(td->thread_number - 1);
853 if (!td->create_serialize) {
859 * do file setup here so it happens sequentially,
860 * we don't want X number of threads getting their
861 * client data interspersed on disk
863 if (setup_files(td)) {
864 td_set_runstate(td, TD_REAPED);
872 struct thread_data *map[MAX_JOBS];
873 struct timeval this_start;
874 int this_jobs = 0, left;
877 * create threads (TD_NOT_CREATED -> TD_CREATED)
880 if (td->runstate != TD_NOT_CREATED)
884 * never got a chance to start, killed by other
885 * thread for some reason
892 if (td->start_delay) {
893 spent = mtime_since_genesis();
895 if (td->start_delay * 1000 > spent)
899 if (td->stonewall && (nr_started || nr_running))
903 * Set state to created. Thread will transition
904 * to TD_INITIALIZED when it's done setting up.
906 td_set_runstate(td, TD_CREATED);
907 map[this_jobs++] = td;
908 fio_sem_init(&startup_sem, 1);
911 if (td->use_thread) {
912 if (pthread_create(&td->thread, NULL, thread_main, td)) {
913 perror("thread_create");
918 fio_sem_down(&startup_sem);
920 fork_main(shm_id, i);
927 * Wait for the started threads to transition to
930 fio_gettime(&this_start, NULL);
932 while (left && !fio_abort) {
933 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
938 for (i = 0; i < this_jobs; i++) {
942 if (td->runstate == TD_INITIALIZED) {
945 } else if (td->runstate >= TD_EXITED) {
949 nr_running++; /* work-around... */
955 log_err("fio: %d jobs failed to start\n", left);
956 for (i = 0; i < this_jobs; i++) {
960 kill(td->pid, SIGTERM);
966 * start created threads (TD_INITIALIZED -> TD_RUNNING).
969 if (td->runstate != TD_INITIALIZED)
972 td_set_runstate(td, TD_RUNNING);
975 m_rate += td->ratemin;
978 fio_sem_up(&td->mutex);
981 reap_threads(&nr_running, &t_rate, &m_rate);
988 reap_threads(&nr_running, &t_rate, &m_rate);
996 int main(int argc, char *argv[])
1000 if (parse_options(argc, argv))
1003 if (!thread_number) {
1004 log_err("Nothing to do\n");
1008 ps = sysconf(_SC_PAGESIZE);
1010 log_err("Failed to get page size\n");
1017 setup_log(&agg_io_log[DDIR_READ]);
1018 setup_log(&agg_io_log[DDIR_WRITE]);
1021 disk_util_timer_arm();
1028 __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1029 __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");