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));
410 ret = td_io_queue(td, io_u);
412 if (ret > 0 && (io_u->xfer_buflen != io_u->resid) &&
415 * short read/write. requeue.
417 io_u->xfer_buflen = io_u->resid;
418 io_u->xfer_buf += ret;
421 td_verror(td, io_u->error);
427 add_slat_sample(td, io_u->ddir, mtime_since(&io_u->start_time, &io_u->issue_time));
429 if (td->cur_depth < td->iodepth) {
430 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
439 ret = td_io_getevents(td, min_evts, td->cur_depth, timeout);
447 ios_completed(td, &icd);
449 td_verror(td, icd.error);
454 * the rate is batched for now, it should work for batches
455 * of completions except the very first one which may look
458 usec = utime_since(&s, &icd.time);
460 rate_throttle(td, usec, icd.bytes_done[td->ddir], td->ddir);
462 if (check_min_rate(td, &icd.time)) {
463 if (exitall_on_terminate)
464 terminate_threads(td->groupid, 0);
465 td_verror(td, ENODATA);
469 if (runtime_exceeded(td, &icd.time))
473 unsigned long long b;
475 b = td->io_blocks[0] + td->io_blocks[1];
476 if (!(b % td->thinktime_blocks))
477 usec_sleep(td, td->thinktime);
483 cleanup_pending_aio(td);
485 if (should_fsync(td) && td->end_fsync) {
486 td_set_runstate(td, TD_FSYNCING);
487 for_each_file(td, f, i)
493 static void cleanup_io_u(struct thread_data *td)
495 struct list_head *entry, *n;
498 list_for_each_safe(entry, n, &td->io_u_freelist) {
499 io_u = list_entry(entry, struct io_u, list);
501 list_del(&io_u->list);
509 * "randomly" fill the buffer contents
511 static void fill_rand_buf(struct io_u *io_u, int max_bs)
513 int *ptr = io_u->buf;
515 while ((void *) ptr - io_u->buf < max_bs) {
516 *ptr = rand() * 0x9e370001;
521 static int init_io_u(struct thread_data *td)
528 if (td->io_ops->flags & FIO_CPUIO)
531 if (td->io_ops->flags & FIO_SYNCIO)
534 max_units = td->iodepth;
536 max_bs = max(td->max_bs[DDIR_READ], td->max_bs[DDIR_WRITE]);
537 td->orig_buffer_size = max_bs * max_units;
539 if (td->mem_type == MEM_SHMHUGE || td->mem_type == MEM_MMAPHUGE)
540 td->orig_buffer_size = (td->orig_buffer_size + td->hugepage_size - 1) & ~(td->hugepage_size - 1);
542 td->orig_buffer_size += page_mask;
544 if (allocate_io_mem(td))
547 p = ALIGN(td->orig_buffer);
548 for (i = 0; i < max_units; i++) {
549 io_u = malloc(sizeof(*io_u));
550 memset(io_u, 0, sizeof(*io_u));
551 INIT_LIST_HEAD(&io_u->list);
553 io_u->buf = p + max_bs * i;
554 if (td_write(td) || td_rw(td))
555 fill_rand_buf(io_u, max_bs);
558 list_add(&io_u->list, &td->io_u_freelist);
564 static int switch_ioscheduler(struct thread_data *td)
566 char tmp[256], tmp2[128];
570 if (td->io_ops->flags & FIO_CPUIO)
573 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
575 f = fopen(tmp, "r+");
577 td_verror(td, errno);
584 ret = fwrite(td->ioscheduler, strlen(td->ioscheduler), 1, f);
585 if (ferror(f) || ret != 1) {
586 td_verror(td, errno);
594 * Read back and check that the selected scheduler is now the default.
596 ret = fread(tmp, 1, sizeof(tmp), f);
597 if (ferror(f) || ret < 0) {
598 td_verror(td, errno);
603 sprintf(tmp2, "[%s]", td->ioscheduler);
604 if (!strstr(tmp, tmp2)) {
605 log_err("fio: io scheduler %s not found\n", td->ioscheduler);
606 td_verror(td, EINVAL);
615 static void clear_io_state(struct thread_data *td)
620 td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0;
621 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
624 for_each_file(td, f, i) {
626 if (td->io_ops->flags & FIO_SYNCIO)
627 lseek(f->fd, SEEK_SET, 0);
630 memset(f->file_map, 0, f->num_maps * sizeof(long));
635 * Entry point for the thread based jobs. The process based jobs end up
636 * here as well, after a little setup.
638 static void *thread_main(void *data)
640 unsigned long long runtime[2];
641 struct thread_data *td = data;
648 INIT_LIST_HEAD(&td->io_u_freelist);
649 INIT_LIST_HEAD(&td->io_u_busylist);
650 INIT_LIST_HEAD(&td->io_hist_list);
651 INIT_LIST_HEAD(&td->io_log_list);
656 if (fio_setaffinity(td) == -1) {
657 td_verror(td, errno);
665 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
666 td_verror(td, errno);
671 if (nice(td->nice) == -1) {
672 td_verror(td, errno);
676 if (init_random_state(td))
679 if (td->ioscheduler && switch_ioscheduler(td))
682 td_set_runstate(td, TD_INITIALIZED);
683 fio_sem_up(&startup_sem);
684 fio_sem_down(&td->mutex);
686 if (!td->create_serialize && setup_files(td))
692 * Do this late, as some IO engines would like to have the
693 * files setup prior to initializing structures.
699 system(td->exec_prerun);
701 fio_gettime(&td->epoch, NULL);
702 getrusage(RUSAGE_SELF, &td->ru_start);
704 runtime[0] = runtime[1] = 0;
705 while (td->loops--) {
706 fio_gettime(&td->start, NULL);
707 memcpy(&td->stat_sample_time, &td->start, sizeof(td->start));
710 memcpy(&td->lastrate, &td->stat_sample_time, sizeof(td->lastrate));
713 prune_io_piece_log(td);
715 if (td->io_ops->flags & FIO_CPUIO)
720 runtime[td->ddir] += utime_since_now(&td->start);
721 if (td_rw(td) && td->io_bytes[td->ddir ^ 1])
722 runtime[td->ddir ^ 1] = runtime[td->ddir];
724 if (td->error || td->terminate)
727 if (td->verify == VERIFY_NONE)
731 fio_gettime(&td->start, NULL);
735 runtime[DDIR_READ] += utime_since_now(&td->start);
737 if (td->error || td->terminate)
741 update_rusage_stat(td);
742 fio_gettime(&td->end_time, NULL);
743 td->runtime[0] = runtime[0] / 1000;
744 td->runtime[1] = runtime[1] / 1000;
747 finish_log(td, td->bw_log, "bw");
749 finish_log(td, td->slat_log, "slat");
751 finish_log(td, td->clat_log, "clat");
752 if (td->write_iolog_file)
753 write_iolog_close(td);
754 if (td->exec_postrun)
755 system(td->exec_postrun);
757 if (exitall_on_terminate)
758 terminate_threads(td->groupid, 0);
764 td_set_runstate(td, TD_EXITED);
770 * We cannot pass the td data into a forked process, so attach the td and
771 * pass it to the thread worker.
773 static void *fork_main(int shmid, int offset)
775 struct thread_data *td;
778 data = shmat(shmid, NULL, 0);
779 if (data == (void *) -1) {
784 td = data + offset * sizeof(struct thread_data);
791 * Run over the job map and reap the threads that have exited, if any.
793 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
795 struct thread_data *td;
796 int i, cputhreads, pending;
799 * reap exited threads (TD_EXITED -> TD_REAPED)
801 pending = cputhreads = 0;
804 * ->io_ops is NULL for a thread that has closed its
807 if (td->io_ops && td->io_ops->flags & FIO_CPUIO)
810 if (td->runstate != TD_EXITED) {
811 if (td->runstate < TD_RUNNING)
817 td_set_runstate(td, TD_REAPED);
819 if (td->use_thread) {
822 if (pthread_join(td->thread, (void *) &ret))
823 perror("thread_join");
825 waitpid(td->pid, NULL, 0);
828 (*m_rate) -= td->ratemin;
829 (*t_rate) -= td->rate;
832 if (*nr_running == cputhreads && !pending)
833 terminate_threads(TERMINATE_ALL, 0);
837 * Main function for kicking off and reaping jobs, as needed.
839 static void run_threads(void)
841 struct thread_data *td;
843 int i, todo, nr_running, m_rate, t_rate, nr_started;
845 if (fio_pin_memory())
849 printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
853 signal(SIGINT, sig_handler);
854 signal(SIGALRM, sig_handler);
856 todo = thread_number;
862 print_status_init(td->thread_number - 1);
864 if (!td->create_serialize) {
870 * do file setup here so it happens sequentially,
871 * we don't want X number of threads getting their
872 * client data interspersed on disk
874 if (setup_files(td)) {
875 td_set_runstate(td, TD_REAPED);
883 struct thread_data *map[MAX_JOBS];
884 struct timeval this_start;
885 int this_jobs = 0, left;
888 * create threads (TD_NOT_CREATED -> TD_CREATED)
891 if (td->runstate != TD_NOT_CREATED)
895 * never got a chance to start, killed by other
896 * thread for some reason
903 if (td->start_delay) {
904 spent = mtime_since_genesis();
906 if (td->start_delay * 1000 > spent)
910 if (td->stonewall && (nr_started || nr_running))
914 * Set state to created. Thread will transition
915 * to TD_INITIALIZED when it's done setting up.
917 td_set_runstate(td, TD_CREATED);
918 map[this_jobs++] = td;
919 fio_sem_init(&startup_sem, 1);
922 if (td->use_thread) {
923 if (pthread_create(&td->thread, NULL, thread_main, td)) {
924 perror("thread_create");
929 fio_sem_down(&startup_sem);
931 fork_main(shm_id, i);
938 * Wait for the started threads to transition to
941 fio_gettime(&this_start, NULL);
943 while (left && !fio_abort) {
944 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
949 for (i = 0; i < this_jobs; i++) {
953 if (td->runstate == TD_INITIALIZED) {
956 } else if (td->runstate >= TD_EXITED) {
960 nr_running++; /* work-around... */
966 log_err("fio: %d jobs failed to start\n", left);
967 for (i = 0; i < this_jobs; i++) {
971 kill(td->pid, SIGTERM);
977 * start created threads (TD_INITIALIZED -> TD_RUNNING).
980 if (td->runstate != TD_INITIALIZED)
983 td_set_runstate(td, TD_RUNNING);
986 m_rate += td->ratemin;
989 fio_sem_up(&td->mutex);
992 reap_threads(&nr_running, &t_rate, &m_rate);
999 reap_threads(&nr_running, &t_rate, &m_rate);
1007 int main(int argc, char *argv[])
1011 if (parse_options(argc, argv))
1014 if (!thread_number) {
1015 log_err("Nothing to do\n");
1019 ps = sysconf(_SC_PAGESIZE);
1021 log_err("Failed to get page size\n");
1028 setup_log(&agg_io_log[DDIR_READ]);
1029 setup_log(&agg_io_log[DDIR_WRITE]);
1032 disk_util_timer_arm();
1039 __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1040 __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");