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
39 static unsigned long page_mask;
41 (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
44 int thread_number = 0;
48 static volatile int startup_sem;
49 static volatile int fio_abort;
51 struct io_log *agg_io_log[2];
53 #define TERMINATE_ALL (-1)
54 #define JOB_START_TIMEOUT (5 * 1000)
56 static inline void td_set_runstate(struct thread_data *td, int runstate)
58 td->runstate = runstate;
61 static void terminate_threads(int group_id, int forced_kill)
63 struct thread_data *td;
67 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
71 td_set_runstate(td, TD_EXITED);
76 static void sig_handler(int sig)
81 disk_util_timer_arm();
82 print_thread_status();
85 printf("\nfio: terminating on signal %d\n", sig);
87 terminate_threads(TERMINATE_ALL, 0);
93 * Check if we are above the minimum rate given.
95 static int check_min_rate(struct thread_data *td, struct timeval *now)
102 * allow a 2 second settle period in the beginning
104 if (mtime_since(&td->start, now) < 2000)
108 * if rate blocks is set, sample is running
110 if (td->rate_bytes) {
111 spent = mtime_since(&td->lastrate, now);
112 if (spent < td->ratecycle)
115 rate = (td->this_io_bytes[ddir] - td->rate_bytes) / spent;
116 if (rate < td->ratemin) {
117 fprintf(f_out, "%s: min rate %u not met, got %luKiB/sec\n", td->name, td->ratemin, rate);
122 td->rate_bytes = td->this_io_bytes[ddir];
123 memcpy(&td->lastrate, now, sizeof(*now));
127 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
131 if (mtime_since(&td->epoch, t) >= td->timeout * 1000)
137 static struct fio_file *get_next_file(struct thread_data *td)
139 unsigned int old_next_file = td->next_file;
143 f = &td->files[td->next_file];
146 if (td->next_file >= td->nr_files)
153 } while (td->next_file != old_next_file);
159 * When job exits, we can cancel the in-flight IO if we are using async
160 * io. Attempt to do so.
162 static void cleanup_pending_aio(struct thread_data *td)
164 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
165 struct list_head *entry, *n;
166 struct io_completion_data icd;
171 * get immediately available events, if any
173 r = td_io_getevents(td, 0, td->cur_depth, &ts);
176 ios_completed(td, &icd);
180 * now cancel remaining active events
182 if (td->io_ops->cancel) {
183 list_for_each_safe(entry, n, &td->io_u_busylist) {
184 io_u = list_entry(entry, struct io_u, list);
186 r = td->io_ops->cancel(td, io_u);
193 r = td_io_getevents(td, td->cur_depth, td->cur_depth, NULL);
196 ios_completed(td, &icd);
202 * Helper to handle the final sync of a file. Works just like the normal
203 * io path, just does everything sync.
205 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
207 struct io_u *io_u = __get_io_u(td);
208 struct io_completion_data icd;
214 io_u->ddir = DDIR_SYNC;
217 if (td_io_prep(td, io_u)) {
222 ret = td_io_queue(td, io_u);
224 td_verror(td, io_u->error);
229 ret = td_io_getevents(td, 1, td->cur_depth, NULL);
236 ios_completed(td, &icd);
238 td_verror(td, icd.error);
246 * The main verify engine. Runs over the writes we previusly submitted,
247 * reads the blocks back in, and checks the crc/md5 of the data.
249 static void do_verify(struct thread_data *td)
251 struct io_u *io_u, *v_io_u = NULL;
252 struct io_completion_data icd;
257 * sync io first and invalidate cache, to make sure we really
260 for_each_file(td, f, i) {
262 file_invalidate_cache(td, f);
265 td_set_runstate(td, TD_VERIFYING);
271 io_u = __get_io_u(td);
275 if (runtime_exceeded(td, &io_u->start_time)) {
280 if (get_next_verify(td, io_u)) {
285 f = get_next_file(td);
291 if (td_io_prep(td, io_u)) {
296 ret = td_io_queue(td, io_u);
298 td_verror(td, io_u->error);
304 * we have one pending to verify, do that while
305 * we are doing io on the next one
307 if (do_io_u_verify(td, &v_io_u))
310 ret = td_io_getevents(td, 1, 1, NULL);
317 v_io_u = td->io_ops->event(td, 0);
320 fio_gettime(&icd.time, NULL);
321 io_completed(td, v_io_u, &icd);
324 td_verror(td, icd.error);
325 put_io_u(td, v_io_u);
331 * if we can't submit more io, we need to verify now
333 if (queue_full(td) && do_io_u_verify(td, &v_io_u))
338 do_io_u_verify(td, &v_io_u);
341 cleanup_pending_aio(td);
343 td_set_runstate(td, TD_RUNNING);
347 * Not really an io thread, all it does is burn CPU cycles in the specified
350 static void do_cpuio(struct thread_data *td)
353 int split = 100 / td->cpuload;
356 while (!td->terminate) {
357 fio_gettime(&e, NULL);
359 if (runtime_exceeded(td, &e))
365 usec_sleep(td, 10000);
372 * Main IO worker function. It retrieves io_u's to process and queues
373 * and reaps them, checking for rate and errors along the way.
375 static void do_io(struct thread_data *td)
377 struct io_completion_data icd;
383 td_set_runstate(td, TD_RUNNING);
385 while (td->this_io_bytes[td->ddir] < td->io_size) {
386 struct timespec *timeout;
393 f = get_next_file(td);
397 io_u = get_io_u(td, f);
401 memcpy(&s, &io_u->start_time, sizeof(s));
404 ret = td_io_queue(td, io_u);
406 if (ret > 0 && (io_u->xfer_buflen != io_u->resid) &&
409 * short read/write. requeue.
411 io_u->xfer_buflen = io_u->resid;
412 io_u->xfer_buf += ret;
415 td_verror(td, io_u->error);
421 add_slat_sample(td, io_u->ddir, mtime_since(&io_u->start_time, &io_u->issue_time));
423 if (td->cur_depth < td->iodepth) {
424 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
433 ret = td_io_getevents(td, min_evts, td->cur_depth, timeout);
441 ios_completed(td, &icd);
443 td_verror(td, icd.error);
448 * the rate is batched for now, it should work for batches
449 * of completions except the very first one which may look
452 usec = utime_since(&s, &icd.time);
454 rate_throttle(td, usec, icd.bytes_done[td->ddir], td->ddir);
456 if (check_min_rate(td, &icd.time)) {
457 if (exitall_on_terminate)
458 terminate_threads(td->groupid, 0);
459 td_verror(td, ENODATA);
463 if (runtime_exceeded(td, &icd.time))
467 unsigned long long b;
469 b = td->io_blocks[0] + td->io_blocks[1];
470 if (!(b % td->thinktime_blocks))
471 usec_sleep(td, td->thinktime);
477 cleanup_pending_aio(td);
479 if (should_fsync(td) && td->end_fsync) {
480 td_set_runstate(td, TD_FSYNCING);
481 for_each_file(td, f, i)
487 static void cleanup_io_u(struct thread_data *td)
489 struct list_head *entry, *n;
492 list_for_each_safe(entry, n, &td->io_u_freelist) {
493 io_u = list_entry(entry, struct io_u, list);
495 list_del(&io_u->list);
503 * "randomly" fill the buffer contents
505 static void fill_rand_buf(struct io_u *io_u, int max_bs)
507 int *ptr = io_u->buf;
509 while ((void *) ptr - io_u->buf < max_bs) {
510 *ptr = rand() * 0x9e370001;
515 static int init_io_u(struct thread_data *td)
522 if (td->io_ops->flags & FIO_CPUIO)
525 if (td->io_ops->flags & FIO_SYNCIO)
528 max_units = td->iodepth;
530 max_bs = max(td->max_bs[DDIR_READ], td->max_bs[DDIR_WRITE]);
531 td->orig_buffer_size = max_bs * max_units;
533 if (td->mem_type == MEM_SHMHUGE || td->mem_type == MEM_MMAPHUGE)
534 td->orig_buffer_size = (td->orig_buffer_size + td->hugepage_size - 1) & ~(td->hugepage_size - 1);
536 td->orig_buffer_size += page_mask;
538 if (allocate_io_mem(td))
541 p = ALIGN(td->orig_buffer);
542 for (i = 0; i < max_units; i++) {
543 io_u = malloc(sizeof(*io_u));
544 memset(io_u, 0, sizeof(*io_u));
545 INIT_LIST_HEAD(&io_u->list);
547 io_u->buf = p + max_bs * i;
548 if (td_write(td) || td_rw(td))
549 fill_rand_buf(io_u, max_bs);
552 list_add(&io_u->list, &td->io_u_freelist);
558 static int switch_ioscheduler(struct thread_data *td)
560 char tmp[256], tmp2[128];
564 if (td->io_ops->flags & FIO_CPUIO)
567 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
569 f = fopen(tmp, "r+");
571 td_verror(td, errno);
578 ret = fwrite(td->ioscheduler, strlen(td->ioscheduler), 1, f);
579 if (ferror(f) || ret != 1) {
580 td_verror(td, errno);
588 * Read back and check that the selected scheduler is now the default.
590 ret = fread(tmp, 1, sizeof(tmp), f);
591 if (ferror(f) || ret < 0) {
592 td_verror(td, errno);
597 sprintf(tmp2, "[%s]", td->ioscheduler);
598 if (!strstr(tmp, tmp2)) {
599 log_err("fio: io scheduler %s not found\n", td->ioscheduler);
600 td_verror(td, EINVAL);
609 static void clear_io_state(struct thread_data *td)
614 td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0;
615 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
618 for_each_file(td, f, i) {
620 if (td->io_ops->flags & FIO_SYNCIO)
621 lseek(f->fd, SEEK_SET, 0);
624 memset(f->file_map, 0, f->num_maps * sizeof(long));
629 * Entry point for the thread based jobs. The process based jobs end up
630 * here as well, after a little setup.
632 static void *thread_main(void *data)
634 unsigned long long runtime[2];
635 struct thread_data *td = data;
642 INIT_LIST_HEAD(&td->io_u_freelist);
643 INIT_LIST_HEAD(&td->io_u_busylist);
644 INIT_LIST_HEAD(&td->io_hist_list);
645 INIT_LIST_HEAD(&td->io_log_list);
650 if (fio_setaffinity(td) == -1) {
651 td_verror(td, errno);
659 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
660 td_verror(td, errno);
665 if (nice(td->nice) == -1) {
666 td_verror(td, errno);
670 if (init_random_state(td))
673 if (td->ioscheduler && switch_ioscheduler(td))
676 td_set_runstate(td, TD_INITIALIZED);
677 fio_sem_up(&startup_sem);
678 fio_sem_down(&td->mutex);
680 if (!td->create_serialize && setup_files(td))
686 * Do this late, as some IO engines would like to have the
687 * files setup prior to initializing structures.
692 if (td->exec_prerun) {
693 if (system(td->exec_prerun) < 0)
697 fio_gettime(&td->epoch, NULL);
698 getrusage(RUSAGE_SELF, &td->ru_start);
700 runtime[0] = runtime[1] = 0;
701 while (td->loops--) {
702 fio_gettime(&td->start, NULL);
703 memcpy(&td->stat_sample_time, &td->start, sizeof(td->start));
706 memcpy(&td->lastrate, &td->stat_sample_time, sizeof(td->lastrate));
709 prune_io_piece_log(td);
711 if (td->io_ops->flags & FIO_CPUIO)
716 runtime[td->ddir] += utime_since_now(&td->start);
717 if (td_rw(td) && td->io_bytes[td->ddir ^ 1])
718 runtime[td->ddir ^ 1] = runtime[td->ddir];
720 if (td->error || td->terminate)
723 if (td->verify == VERIFY_NONE)
727 fio_gettime(&td->start, NULL);
731 runtime[DDIR_READ] += utime_since_now(&td->start);
733 if (td->error || td->terminate)
737 update_rusage_stat(td);
738 fio_gettime(&td->end_time, NULL);
739 td->runtime[0] = runtime[0] / 1000;
740 td->runtime[1] = runtime[1] / 1000;
743 finish_log(td, td->bw_log, "bw");
745 finish_log(td, td->slat_log, "slat");
747 finish_log(td, td->clat_log, "clat");
748 if (td->write_iolog_file)
749 write_iolog_close(td);
750 if (td->exec_postrun) {
751 if (system(td->exec_postrun) < 0)
752 log_err("fio: postrun %s failed\n", td->exec_postrun);
755 if (exitall_on_terminate)
756 terminate_threads(td->groupid, 0);
762 td_set_runstate(td, TD_EXITED);
768 * We cannot pass the td data into a forked process, so attach the td and
769 * pass it to the thread worker.
771 static void *fork_main(int shmid, int offset)
773 struct thread_data *td;
776 data = shmat(shmid, NULL, 0);
777 if (data == (void *) -1) {
782 td = data + offset * sizeof(struct thread_data);
789 * Run over the job map and reap the threads that have exited, if any.
791 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
793 struct thread_data *td;
794 int i, cputhreads, pending;
797 * reap exited threads (TD_EXITED -> TD_REAPED)
799 pending = cputhreads = 0;
802 * ->io_ops is NULL for a thread that has closed its
805 if (td->io_ops && td->io_ops->flags & FIO_CPUIO)
808 if (td->runstate != TD_EXITED) {
809 if (td->runstate < TD_RUNNING)
815 td_set_runstate(td, TD_REAPED);
817 if (td->use_thread) {
820 if (pthread_join(td->thread, (void *) &ret))
821 perror("thread_join");
823 waitpid(td->pid, NULL, 0);
826 (*m_rate) -= td->ratemin;
827 (*t_rate) -= td->rate;
830 if (*nr_running == cputhreads && !pending)
831 terminate_threads(TERMINATE_ALL, 0);
835 * Main function for kicking off and reaping jobs, as needed.
837 static void run_threads(void)
839 struct thread_data *td;
841 int i, todo, nr_running, m_rate, t_rate, nr_started;
843 if (fio_pin_memory())
847 printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
851 signal(SIGINT, sig_handler);
852 signal(SIGALRM, sig_handler);
854 todo = thread_number;
860 print_status_init(td->thread_number - 1);
862 if (!td->create_serialize) {
868 * do file setup here so it happens sequentially,
869 * we don't want X number of threads getting their
870 * client data interspersed on disk
872 if (setup_files(td)) {
873 td_set_runstate(td, TD_REAPED);
881 struct thread_data *map[MAX_JOBS];
882 struct timeval this_start;
883 int this_jobs = 0, left;
886 * create threads (TD_NOT_CREATED -> TD_CREATED)
889 if (td->runstate != TD_NOT_CREATED)
893 * never got a chance to start, killed by other
894 * thread for some reason
901 if (td->start_delay) {
902 spent = mtime_since_genesis();
904 if (td->start_delay * 1000 > spent)
908 if (td->stonewall && (nr_started || nr_running))
912 * Set state to created. Thread will transition
913 * to TD_INITIALIZED when it's done setting up.
915 td_set_runstate(td, TD_CREATED);
916 map[this_jobs++] = td;
917 fio_sem_init(&startup_sem, 1);
920 if (td->use_thread) {
921 if (pthread_create(&td->thread, NULL, thread_main, td)) {
922 perror("thread_create");
927 fio_sem_down(&startup_sem);
929 fork_main(shm_id, i);
936 * Wait for the started threads to transition to
939 fio_gettime(&this_start, NULL);
941 while (left && !fio_abort) {
942 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
947 for (i = 0; i < this_jobs; i++) {
951 if (td->runstate == TD_INITIALIZED) {
954 } else if (td->runstate >= TD_EXITED) {
958 nr_running++; /* work-around... */
964 log_err("fio: %d jobs failed to start\n", left);
965 for (i = 0; i < this_jobs; i++) {
969 kill(td->pid, SIGTERM);
975 * start created threads (TD_INITIALIZED -> TD_RUNNING).
978 if (td->runstate != TD_INITIALIZED)
981 td_set_runstate(td, TD_RUNNING);
984 m_rate += td->ratemin;
987 fio_sem_up(&td->mutex);
990 reap_threads(&nr_running, &t_rate, &m_rate);
997 reap_threads(&nr_running, &t_rate, &m_rate);
1005 int main(int argc, char *argv[])
1010 * We need locale for number printing, if it isn't set then just
1011 * go with the US format.
1013 if (!getenv("LC_NUMERIC"))
1014 setlocale(LC_NUMERIC, "en_US");
1016 if (parse_options(argc, argv))
1019 if (!thread_number) {
1020 log_err("Nothing to do\n");
1024 ps = sysconf(_SC_PAGESIZE);
1026 log_err("Failed to get page size\n");
1033 setup_log(&agg_io_log[DDIR_READ]);
1034 setup_log(&agg_io_log[DDIR_WRITE]);
1037 disk_util_timer_arm();
1044 __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1045 __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");