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
50 unsigned long page_mask;
51 unsigned long page_size;
53 #define PAGE_ALIGN(buf) \
54 (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
57 int thread_number = 0;
62 unsigned long done_secs = 0;
64 static struct fio_mutex *startup_mutex;
65 static struct fio_mutex *writeout_mutex;
66 static volatile int fio_abort;
67 static int exit_value;
68 static pthread_t gtod_thread;
69 static pthread_t disk_util_thread;
70 static struct flist_head *cgroup_list;
71 static char *cgroup_mnt;
73 struct io_log *agg_io_log[2];
75 #define TERMINATE_ALL (-1)
76 #define JOB_START_TIMEOUT (5 * 1000)
78 void td_set_runstate(struct thread_data *td, int runstate)
80 if (td->runstate == runstate)
83 dprint(FD_PROCESS, "pid=%d: runstate %d -> %d\n", (int) td->pid,
84 td->runstate, runstate);
85 td->runstate = runstate;
88 static void terminate_threads(int group_id)
90 struct thread_data *td;
93 dprint(FD_PROCESS, "terminate group_id=%d\n", group_id);
96 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
97 dprint(FD_PROCESS, "setting terminate on %s/%d\n",
98 td->o.name, (int) td->pid);
100 td->o.start_delay = 0;
103 * if the thread is running, just let it exit
105 if (td->runstate < TD_RAMP)
106 kill(td->pid, SIGTERM);
108 struct ioengine_ops *ops = td->io_ops;
110 if (ops && (ops->flags & FIO_SIGTERM))
111 kill(td->pid, SIGTERM);
117 static void sig_int(int sig)
120 log_info("\nfio: terminating on signal %d\n", sig);
123 terminate_threads(TERMINATE_ALL);
127 static void *disk_thread_main(void *data)
129 fio_mutex_up(startup_mutex);
132 usleep(DISK_UTIL_MSEC * 1000);
136 print_thread_status();
142 static int create_disk_util_thread(void)
146 ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
148 log_err("Can't create disk util thread: %s\n", strerror(ret));
152 ret = pthread_detach(disk_util_thread);
154 log_err("Can't detatch disk util thread: %s\n", strerror(ret));
158 dprint(FD_MUTEX, "wait on startup_mutex\n");
159 fio_mutex_down(startup_mutex);
160 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
164 static void set_sig_handlers(void)
166 struct sigaction act;
168 memset(&act, 0, sizeof(act));
169 act.sa_handler = sig_int;
170 act.sa_flags = SA_RESTART;
171 sigaction(SIGINT, &act, NULL);
173 memset(&act, 0, sizeof(act));
174 act.sa_handler = sig_int;
175 act.sa_flags = SA_RESTART;
176 sigaction(SIGTERM, &act, NULL);
180 * Check if we are above the minimum rate given.
182 static int __check_min_rate(struct thread_data *td, struct timeval *now,
185 unsigned long long bytes = 0;
186 unsigned long iops = 0;
189 unsigned int ratemin = 0;
190 unsigned int rate_iops = 0;
191 unsigned int rate_iops_min = 0;
193 assert(ddir_rw(ddir));
195 if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
199 * allow a 2 second settle period in the beginning
201 if (mtime_since(&td->start, now) < 2000)
204 iops += td->io_blocks[ddir];
205 bytes += td->this_io_bytes[ddir];
206 ratemin += td->o.ratemin[ddir];
207 rate_iops += td->o.rate_iops[ddir];
208 rate_iops_min += td->o.rate_iops_min[ddir];
211 * if rate blocks is set, sample is running
213 if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
214 spent = mtime_since(&td->lastrate[ddir], now);
215 if (spent < td->o.ratecycle)
218 if (td->o.rate[ddir]) {
220 * check bandwidth specified rate
222 if (bytes < td->rate_bytes[ddir]) {
223 log_err("%s: min rate %u not met\n", td->o.name,
227 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
228 if (rate < ratemin ||
229 bytes < td->rate_bytes[ddir]) {
230 log_err("%s: min rate %u not met, got"
231 " %luKB/sec\n", td->o.name,
238 * checks iops specified rate
240 if (iops < rate_iops) {
241 log_err("%s: min iops rate %u not met\n",
242 td->o.name, rate_iops);
245 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
246 if (rate < rate_iops_min ||
247 iops < td->rate_blocks[ddir]) {
248 log_err("%s: min iops rate %u not met,"
249 " got %lu\n", td->o.name,
250 rate_iops_min, rate);
256 td->rate_bytes[ddir] = bytes;
257 td->rate_blocks[ddir] = iops;
258 memcpy(&td->lastrate[ddir], now, sizeof(*now));
262 static int check_min_rate(struct thread_data *td, struct timeval *now,
263 unsigned long *bytes_done)
268 ret |= __check_min_rate(td, now, 0);
270 ret |= __check_min_rate(td, now, 1);
275 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
279 if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
286 * When job exits, we can cancel the in-flight IO if we are using async
287 * io. Attempt to do so.
289 static void cleanup_pending_aio(struct thread_data *td)
291 struct flist_head *entry, *n;
296 * get immediately available events, if any
298 r = io_u_queued_complete(td, 0, NULL);
303 * now cancel remaining active events
305 if (td->io_ops->cancel) {
306 flist_for_each_safe(entry, n, &td->io_u_busylist) {
307 io_u = flist_entry(entry, struct io_u, list);
310 * if the io_u isn't in flight, then that generally
311 * means someone leaked an io_u. complain but fix
312 * it up, so we don't stall here.
314 if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
315 log_err("fio: non-busy IO on busy list\n");
318 r = td->io_ops->cancel(td, io_u);
326 r = io_u_queued_complete(td, td->cur_depth, NULL);
330 * Helper to handle the final sync of a file. Works just like the normal
331 * io path, just does everything sync.
333 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
335 struct io_u *io_u = __get_io_u(td);
341 io_u->ddir = DDIR_SYNC;
344 if (td_io_prep(td, io_u)) {
350 ret = td_io_queue(td, io_u);
352 td_verror(td, io_u->error, "td_io_queue");
355 } else if (ret == FIO_Q_QUEUED) {
356 if (io_u_queued_complete(td, 1, NULL) < 0)
358 } else if (ret == FIO_Q_COMPLETED) {
360 td_verror(td, io_u->error, "td_io_queue");
364 if (io_u_sync_complete(td, io_u, NULL) < 0)
366 } else if (ret == FIO_Q_BUSY) {
367 if (td_io_commit(td))
375 static inline void __update_tv_cache(struct thread_data *td)
377 fio_gettime(&td->tv_cache, NULL);
380 static inline void update_tv_cache(struct thread_data *td)
382 if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
383 __update_tv_cache(td);
386 static int break_on_this_error(struct thread_data *td, int *retptr)
390 if (ret < 0 || td->error) {
393 if (!td->o.continue_on_error)
401 if (td_non_fatal_error(err)) {
403 * Continue with the I/Os in case of
406 update_error_count(td, err);
410 } else if (td->o.fill_device && err == ENOSPC) {
412 * We expect to hit this error if
413 * fill_device option is set.
420 * Stop the I/O in case of a fatal
423 update_error_count(td, err);
432 * The main verify engine. Runs over the writes we previously submitted,
433 * reads the blocks back in, and checks the crc/md5 of the data.
435 static void do_verify(struct thread_data *td)
442 dprint(FD_VERIFY, "starting loop\n");
445 * sync io first and invalidate cache, to make sure we really
448 for_each_file(td, f, i) {
449 if (!fio_file_open(f))
451 if (fio_io_sync(td, f))
453 if (file_invalidate_cache(td, f))
460 td_set_runstate(td, TD_VERIFYING);
463 while (!td->terminate) {
468 if (runtime_exceeded(td, &td->tv_cache)) {
469 __update_tv_cache(td);
470 if (runtime_exceeded(td, &td->tv_cache)) {
476 io_u = __get_io_u(td);
480 if (get_next_verify(td, io_u)) {
485 if (td_io_prep(td, io_u)) {
490 if (td->o.verify_async)
491 io_u->end_io = verify_io_u_async;
493 io_u->end_io = verify_io_u;
495 ret = td_io_queue(td, io_u);
497 case FIO_Q_COMPLETED:
500 clear_io_u(td, io_u);
501 } else if (io_u->resid) {
502 int bytes = io_u->xfer_buflen - io_u->resid;
508 td_verror(td, EIO, "full resid");
513 io_u->xfer_buflen = io_u->resid;
514 io_u->xfer_buf += bytes;
515 io_u->offset += bytes;
517 if (ddir_rw(io_u->ddir))
518 td->ts.short_io_u[io_u->ddir]++;
521 if (io_u->offset == f->real_file_size)
524 requeue_io_u(td, &io_u);
527 ret = io_u_sync_complete(td, io_u, NULL);
535 requeue_io_u(td, &io_u);
536 ret2 = td_io_commit(td);
542 td_verror(td, -ret, "td_io_queue");
546 if (break_on_this_error(td, &ret))
550 * if we can queue more, do so. but check if there are
551 * completed io_u's first.
553 full = queue_full(td) || ret == FIO_Q_BUSY;
554 if (full || !td->o.iodepth_batch_complete) {
555 min_events = min(td->o.iodepth_batch_complete,
557 if (full && !min_events)
562 * Reap required number of io units, if any,
563 * and do the verification on them through
564 * the callback handler
566 if (io_u_queued_complete(td, min_events, NULL) < 0) {
570 } while (full && (td->cur_depth > td->o.iodepth_low));
577 min_events = td->cur_depth;
580 ret = io_u_queued_complete(td, min_events, NULL);
582 cleanup_pending_aio(td);
584 td_set_runstate(td, TD_RUNNING);
586 dprint(FD_VERIFY, "exiting loop\n");
590 * Main IO worker function. It retrieves io_u's to process and queues
591 * and reaps them, checking for rate and errors along the way.
593 static void do_io(struct thread_data *td)
598 if (in_ramp_time(td))
599 td_set_runstate(td, TD_RAMP);
601 td_set_runstate(td, TD_RUNNING);
603 while ( (td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
604 (!flist_empty(&td->trim_list)) ||
605 ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) ) {
606 struct timeval comp_time;
607 unsigned long bytes_done[2] = { 0, 0 };
617 if (runtime_exceeded(td, &td->tv_cache)) {
618 __update_tv_cache(td);
619 if (runtime_exceeded(td, &td->tv_cache)) {
630 * Add verification end_io handler, if asked to verify
631 * a previously written file.
633 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
635 if (td->o.verify_async)
636 io_u->end_io = verify_io_u_async;
638 io_u->end_io = verify_io_u;
639 td_set_runstate(td, TD_VERIFYING);
640 } else if (in_ramp_time(td))
641 td_set_runstate(td, TD_RAMP);
643 td_set_runstate(td, TD_RUNNING);
645 ret = td_io_queue(td, io_u);
647 case FIO_Q_COMPLETED:
650 clear_io_u(td, io_u);
651 } else if (io_u->resid) {
652 int bytes = io_u->xfer_buflen - io_u->resid;
653 struct fio_file *f = io_u->file;
659 td_verror(td, EIO, "full resid");
664 io_u->xfer_buflen = io_u->resid;
665 io_u->xfer_buf += bytes;
666 io_u->offset += bytes;
668 if (ddir_rw(io_u->ddir))
669 td->ts.short_io_u[io_u->ddir]++;
671 if (io_u->offset == f->real_file_size)
674 requeue_io_u(td, &io_u);
677 if (__should_check_rate(td, 0) ||
678 __should_check_rate(td, 1))
679 fio_gettime(&comp_time, NULL);
681 ret = io_u_sync_complete(td, io_u, bytes_done);
688 * if the engine doesn't have a commit hook,
689 * the io_u is really queued. if it does have such
690 * a hook, it has to call io_u_queued() itself.
692 if (td->io_ops->commit == NULL)
693 io_u_queued(td, io_u);
696 requeue_io_u(td, &io_u);
697 ret2 = td_io_commit(td);
707 if (break_on_this_error(td, &ret))
711 * See if we need to complete some commands
713 full = queue_full(td) || ret == FIO_Q_BUSY;
714 if (full || !td->o.iodepth_batch_complete) {
715 min_evts = min(td->o.iodepth_batch_complete,
717 if (full && !min_evts)
720 if (__should_check_rate(td, 0) ||
721 __should_check_rate(td, 1))
722 fio_gettime(&comp_time, NULL);
725 ret = io_u_queued_complete(td, min_evts, bytes_done);
729 } while (full && (td->cur_depth > td->o.iodepth_low));
734 if (!(bytes_done[0] + bytes_done[1]))
737 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
738 if (check_min_rate(td, &comp_time, bytes_done)) {
739 if (exitall_on_terminate)
740 terminate_threads(td->groupid);
741 td_verror(td, EIO, "check_min_rate");
746 if (td->o.thinktime) {
747 unsigned long long b;
749 b = td->io_blocks[0] + td->io_blocks[1];
750 if (!(b % td->o.thinktime_blocks)) {
753 if (td->o.thinktime_spin)
754 usec_spin(td->o.thinktime_spin);
756 left = td->o.thinktime - td->o.thinktime_spin;
758 usec_sleep(td, left);
763 if (td->trim_entries)
764 printf("trim entries %ld\n", td->trim_entries);
766 if (td->o.fill_device && td->error == ENOSPC) {
775 ret = io_u_queued_complete(td, i, NULL);
776 if (td->o.fill_device && td->error == ENOSPC)
780 if (should_fsync(td) && td->o.end_fsync) {
781 td_set_runstate(td, TD_FSYNCING);
783 for_each_file(td, f, i) {
784 if (!fio_file_open(f))
790 cleanup_pending_aio(td);
793 * stop job if we failed doing any IO
795 if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
799 static void cleanup_io_u(struct thread_data *td)
801 struct flist_head *entry, *n;
804 flist_for_each_safe(entry, n, &td->io_u_freelist) {
805 io_u = flist_entry(entry, struct io_u, list);
807 flist_del(&io_u->list);
808 fio_memfree(io_u, sizeof(*io_u));
814 static int init_io_u(struct thread_data *td)
818 int cl_align, i, max_units;
821 max_units = td->o.iodepth;
822 max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
823 td->orig_buffer_size = (unsigned long long) max_bs
824 * (unsigned long long) max_units;
826 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
829 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
830 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
833 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
834 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
838 if (allocate_io_mem(td))
841 if (td->o.odirect || td->o.mem_align ||
842 (td->io_ops->flags & FIO_RAWIO))
843 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
847 cl_align = os_cache_line_size();
849 for (i = 0; i < max_units; i++) {
855 ptr = fio_memalign(cl_align, sizeof(*io_u));
857 log_err("fio: unable to allocate aligned memory\n");
862 memset(io_u, 0, sizeof(*io_u));
863 INIT_FLIST_HEAD(&io_u->list);
864 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
866 if (!(td->io_ops->flags & FIO_NOIO)) {
867 io_u->buf = p + max_bs * i;
868 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
870 if (td_write(td) && !td->o.refill_buffers)
871 io_u_fill_buffer(td, io_u, max_bs);
872 else if (td_write(td) && td->o.verify_pattern_bytes) {
874 * Fill the buffer with the pattern if we are
875 * going to be doing writes.
877 fill_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
882 io_u->flags = IO_U_F_FREE;
883 flist_add(&io_u->list, &td->io_u_freelist);
889 static int switch_ioscheduler(struct thread_data *td)
891 char tmp[256], tmp2[128];
895 if (td->io_ops->flags & FIO_DISKLESSIO)
898 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
900 f = fopen(tmp, "r+");
902 if (errno == ENOENT) {
903 log_err("fio: os or kernel doesn't support IO scheduler"
907 td_verror(td, errno, "fopen iosched");
914 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
915 if (ferror(f) || ret != 1) {
916 td_verror(td, errno, "fwrite");
924 * Read back and check that the selected scheduler is now the default.
926 ret = fread(tmp, 1, sizeof(tmp), f);
927 if (ferror(f) || ret < 0) {
928 td_verror(td, errno, "fread");
933 sprintf(tmp2, "[%s]", td->o.ioscheduler);
934 if (!strstr(tmp, tmp2)) {
935 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
936 td_verror(td, EINVAL, "iosched_switch");
945 static int keep_running(struct thread_data *td)
947 unsigned long long io_done;
951 if (td->o.time_based)
958 io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
960 if (io_done < td->o.size)
966 static void reset_io_counters(struct thread_data *td)
968 td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
969 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
971 td->rate_bytes[0] = td->rate_bytes[1] = 0;
972 td->rate_blocks[0] = td->rate_blocks[1] = 0;
974 td->last_was_sync = 0;
977 * reset file done count if we are to start over
979 if (td->o.time_based || td->o.loops)
980 td->nr_done_files = 0;
983 void reset_all_stats(struct thread_data *td)
988 reset_io_counters(td);
990 for (i = 0; i < 2; i++) {
992 td->io_blocks[i] = 0;
993 td->io_issues[i] = 0;
994 td->ts.total_io_u[i] = 0;
997 fio_gettime(&tv, NULL);
998 td->ts.runtime[0] = 0;
999 td->ts.runtime[1] = 0;
1000 memcpy(&td->epoch, &tv, sizeof(tv));
1001 memcpy(&td->start, &tv, sizeof(tv));
1004 static void clear_io_state(struct thread_data *td)
1009 reset_io_counters(td);
1012 for_each_file(td, f, i)
1013 fio_file_clear_done(f);
1016 * Set the same seed to get repeatable runs
1018 td_fill_rand_seeds(td);
1021 static int exec_string(const char *string)
1023 int ret, newlen = strlen(string) + 1 + 8;
1026 str = malloc(newlen);
1027 sprintf(str, "sh -c %s", string);
1031 log_err("fio: exec of cmd <%s> failed\n", str);
1038 * Entry point for the thread based jobs. The process based jobs end up
1039 * here as well, after a little setup.
1041 static void *thread_main(void *data)
1043 unsigned long long elapsed;
1044 struct thread_data *td = data;
1045 pthread_condattr_t attr;
1048 if (!td->o.use_thread)
1053 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1055 INIT_FLIST_HEAD(&td->io_u_freelist);
1056 INIT_FLIST_HEAD(&td->io_u_busylist);
1057 INIT_FLIST_HEAD(&td->io_u_requeues);
1058 INIT_FLIST_HEAD(&td->io_log_list);
1059 INIT_FLIST_HEAD(&td->io_hist_list);
1060 INIT_FLIST_HEAD(&td->verify_list);
1061 INIT_FLIST_HEAD(&td->trim_list);
1062 pthread_mutex_init(&td->io_u_lock, NULL);
1063 td->io_hist_tree = RB_ROOT;
1065 pthread_condattr_init(&attr);
1066 pthread_cond_init(&td->verify_cond, &attr);
1067 pthread_cond_init(&td->free_cond, &attr);
1069 td_set_runstate(td, TD_INITIALIZED);
1070 dprint(FD_MUTEX, "up startup_mutex\n");
1071 fio_mutex_up(startup_mutex);
1072 dprint(FD_MUTEX, "wait on td->mutex\n");
1073 fio_mutex_down(td->mutex);
1074 dprint(FD_MUTEX, "done waiting on td->mutex\n");
1077 * the ->mutex mutex is now no longer used, close it to avoid
1078 * eating a file descriptor
1080 fio_mutex_remove(td->mutex);
1083 * A new gid requires privilege, so we need to do this before setting
1086 if (td->o.gid != -1U && setgid(td->o.gid)) {
1087 td_verror(td, errno, "setgid");
1090 if (td->o.uid != -1U && setuid(td->o.uid)) {
1091 td_verror(td, errno, "setuid");
1096 * If we have a gettimeofday() thread, make sure we exclude that
1097 * thread from this job
1100 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1103 * Set affinity first, in case it has an impact on the memory
1106 if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1107 td_verror(td, errno, "cpu_set_affinity");
1112 * May alter parameters that init_io_u() will use, so we need to
1121 if (td->o.verify_async && verify_async_init(td))
1124 if (td->ioprio_set) {
1125 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1126 td_verror(td, errno, "ioprio_set");
1131 if (td->o.cgroup_weight && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1134 if (nice(td->o.nice) == -1) {
1135 td_verror(td, errno, "nice");
1139 if (td->o.ioscheduler && switch_ioscheduler(td))
1142 if (!td->o.create_serialize && setup_files(td))
1148 if (init_random_map(td))
1151 if (td->o.exec_prerun) {
1152 if (exec_string(td->o.exec_prerun))
1156 if (td->o.pre_read) {
1157 if (pre_read_files(td) < 0)
1161 fio_gettime(&td->epoch, NULL);
1162 getrusage(RUSAGE_SELF, &td->ts.ru_start);
1165 while (keep_running(td)) {
1166 fio_gettime(&td->start, NULL);
1167 memcpy(&td->ts.stat_sample_time[0], &td->start,
1169 memcpy(&td->ts.stat_sample_time[1], &td->start,
1171 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1173 if (td->o.ratemin[0] || td->o.ratemin[1])
1174 memcpy(&td->lastrate, &td->ts.stat_sample_time,
1175 sizeof(td->lastrate));
1180 prune_io_piece_log(td);
1186 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1187 elapsed = utime_since_now(&td->start);
1188 td->ts.runtime[DDIR_READ] += elapsed;
1190 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1191 elapsed = utime_since_now(&td->start);
1192 td->ts.runtime[DDIR_WRITE] += elapsed;
1195 if (td->error || td->terminate)
1198 if (!td->o.do_verify ||
1199 td->o.verify == VERIFY_NONE ||
1200 (td->io_ops->flags & FIO_UNIDIR))
1205 fio_gettime(&td->start, NULL);
1209 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1211 if (td->error || td->terminate)
1215 update_rusage_stat(td);
1216 td->ts.runtime[0] = (td->ts.runtime[0] + 999) / 1000;
1217 td->ts.runtime[1] = (td->ts.runtime[1] + 999) / 1000;
1218 td->ts.total_run_time = mtime_since_now(&td->epoch);
1219 td->ts.io_bytes[0] = td->io_bytes[0];
1220 td->ts.io_bytes[1] = td->io_bytes[1];
1222 fio_mutex_down(writeout_mutex);
1223 if (td->ts.bw_log) {
1224 if (td->o.bw_log_file) {
1225 finish_log_named(td, td->ts.bw_log,
1226 td->o.bw_log_file, "bw");
1228 finish_log(td, td->ts.bw_log, "bw");
1230 if (td->ts.lat_log) {
1231 if (td->o.lat_log_file) {
1232 finish_log_named(td, td->ts.lat_log,
1233 td->o.lat_log_file, "lat");
1235 finish_log(td, td->ts.lat_log, "lat");
1237 if (td->ts.slat_log) {
1238 if (td->o.lat_log_file) {
1239 finish_log_named(td, td->ts.slat_log,
1240 td->o.lat_log_file, "slat");
1242 finish_log(td, td->ts.slat_log, "slat");
1244 if (td->ts.clat_log) {
1245 if (td->o.lat_log_file) {
1246 finish_log_named(td, td->ts.clat_log,
1247 td->o.lat_log_file, "clat");
1249 finish_log(td, td->ts.clat_log, "clat");
1251 fio_mutex_up(writeout_mutex);
1252 if (td->o.exec_postrun)
1253 exec_string(td->o.exec_postrun);
1255 if (exitall_on_terminate)
1256 terminate_threads(td->groupid);
1260 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1263 if (td->o.verify_async)
1264 verify_async_exit(td);
1266 close_and_free_files(td);
1269 cgroup_shutdown(td, &cgroup_mnt);
1271 if (td->o.cpumask_set) {
1272 int ret = fio_cpuset_exit(&td->o.cpumask);
1274 td_verror(td, ret, "fio_cpuset_exit");
1278 * do this very late, it will log file closing as well
1280 if (td->o.write_iolog_file)
1281 write_iolog_close(td);
1283 options_mem_free(td);
1284 td_set_runstate(td, TD_EXITED);
1285 return (void *) (unsigned long) td->error;
1289 * We cannot pass the td data into a forked process, so attach the td and
1290 * pass it to the thread worker.
1292 static int fork_main(int shmid, int offset)
1294 struct thread_data *td;
1297 data = shmat(shmid, NULL, 0);
1298 if (data == (void *) -1) {
1305 td = data + offset * sizeof(struct thread_data);
1306 ret = thread_main(td);
1308 return (int) (unsigned long) ret;
1312 * Run over the job map and reap the threads that have exited, if any.
1314 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1316 struct thread_data *td;
1317 int i, cputhreads, realthreads, pending, status, ret;
1320 * reap exited threads (TD_EXITED -> TD_REAPED)
1322 realthreads = pending = cputhreads = 0;
1323 for_each_td(td, i) {
1327 * ->io_ops is NULL for a thread that has closed its
1330 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1339 if (td->runstate == TD_REAPED)
1341 if (td->o.use_thread) {
1342 if (td->runstate == TD_EXITED) {
1343 td_set_runstate(td, TD_REAPED);
1350 if (td->runstate == TD_EXITED)
1354 * check if someone quit or got killed in an unusual way
1356 ret = waitpid(td->pid, &status, flags);
1358 if (errno == ECHILD) {
1359 log_err("fio: pid=%d disappeared %d\n",
1360 (int) td->pid, td->runstate);
1361 td_set_runstate(td, TD_REAPED);
1365 } else if (ret == td->pid) {
1366 if (WIFSIGNALED(status)) {
1367 int sig = WTERMSIG(status);
1370 log_err("fio: pid=%d, got signal=%d\n",
1371 (int) td->pid, sig);
1372 td_set_runstate(td, TD_REAPED);
1375 if (WIFEXITED(status)) {
1376 if (WEXITSTATUS(status) && !td->error)
1377 td->error = WEXITSTATUS(status);
1379 td_set_runstate(td, TD_REAPED);
1385 * thread is not dead, continue
1391 (*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
1392 (*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
1399 done_secs += mtime_since_now(&td->epoch) / 1000;
1402 if (*nr_running == cputhreads && !pending && realthreads)
1403 terminate_threads(TERMINATE_ALL);
1406 static void *gtod_thread_main(void *data)
1408 fio_mutex_up(startup_mutex);
1411 * As long as we have jobs around, update the clock. It would be nice
1412 * to have some way of NOT hammering that CPU with gettimeofday(),
1413 * but I'm not sure what to use outside of a simple CPU nop to relax
1414 * it - we don't want to lose precision.
1424 static int fio_start_gtod_thread(void)
1426 pthread_attr_t attr;
1429 pthread_attr_init(&attr);
1430 pthread_attr_setstacksize(&attr, PTHREAD_STACK_MIN);
1431 ret = pthread_create(>od_thread, &attr, gtod_thread_main, NULL);
1432 pthread_attr_destroy(&attr);
1434 log_err("Can't create gtod thread: %s\n", strerror(ret));
1438 ret = pthread_detach(gtod_thread);
1440 log_err("Can't detatch gtod thread: %s\n", strerror(ret));
1444 dprint(FD_MUTEX, "wait on startup_mutex\n");
1445 fio_mutex_down(startup_mutex);
1446 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1451 * Main function for kicking off and reaping jobs, as needed.
1453 static void run_threads(void)
1455 struct thread_data *td;
1456 unsigned long spent;
1457 int i, todo, nr_running, m_rate, t_rate, nr_started;
1459 if (fio_pin_memory())
1462 if (fio_gtod_offload && fio_start_gtod_thread())
1465 if (!terse_output) {
1466 log_info("Starting ");
1468 log_info("%d thread%s", nr_thread,
1469 nr_thread > 1 ? "s" : "");
1473 log_info("%d process%s", nr_process,
1474 nr_process > 1 ? "es" : "");
1482 todo = thread_number;
1485 m_rate = t_rate = 0;
1487 for_each_td(td, i) {
1488 print_status_init(td->thread_number - 1);
1490 if (!td->o.create_serialize) {
1496 * do file setup here so it happens sequentially,
1497 * we don't want X number of threads getting their
1498 * client data interspersed on disk
1500 if (setup_files(td)) {
1503 log_err("fio: pid=%d, err=%d/%s\n",
1504 (int) td->pid, td->error, td->verror);
1505 td_set_runstate(td, TD_REAPED);
1512 * for sharing to work, each job must always open
1513 * its own files. so close them, if we opened them
1516 for_each_file(td, f, j) {
1517 if (fio_file_open(f))
1518 td_io_close_file(td, f);
1528 struct thread_data *map[REAL_MAX_JOBS];
1529 struct timeval this_start;
1530 int this_jobs = 0, left;
1533 * create threads (TD_NOT_CREATED -> TD_CREATED)
1535 for_each_td(td, i) {
1536 if (td->runstate != TD_NOT_CREATED)
1540 * never got a chance to start, killed by other
1541 * thread for some reason
1543 if (td->terminate) {
1548 if (td->o.start_delay) {
1549 spent = mtime_since_genesis();
1551 if (td->o.start_delay * 1000 > spent)
1555 if (td->o.stonewall && (nr_started || nr_running)) {
1556 dprint(FD_PROCESS, "%s: stonewall wait\n",
1562 * Set state to created. Thread will transition
1563 * to TD_INITIALIZED when it's done setting up.
1565 td_set_runstate(td, TD_CREATED);
1566 map[this_jobs++] = td;
1569 if (td->o.use_thread) {
1572 dprint(FD_PROCESS, "will pthread_create\n");
1573 ret = pthread_create(&td->thread, NULL,
1576 log_err("pthread_create: %s\n",
1581 ret = pthread_detach(td->thread);
1583 log_err("pthread_detach: %s",
1587 dprint(FD_PROCESS, "will fork\n");
1590 int ret = fork_main(shm_id, i);
1593 } else if (i == fio_debug_jobno)
1594 *fio_debug_jobp = pid;
1596 dprint(FD_MUTEX, "wait on startup_mutex\n");
1597 if (fio_mutex_down_timeout(startup_mutex, 10)) {
1598 log_err("fio: job startup hung? exiting.\n");
1599 terminate_threads(TERMINATE_ALL);
1604 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1608 * Wait for the started threads to transition to
1611 fio_gettime(&this_start, NULL);
1613 while (left && !fio_abort) {
1614 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1619 for (i = 0; i < this_jobs; i++) {
1623 if (td->runstate == TD_INITIALIZED) {
1626 } else if (td->runstate >= TD_EXITED) {
1630 nr_running++; /* work-around... */
1636 log_err("fio: %d jobs failed to start\n", left);
1637 for (i = 0; i < this_jobs; i++) {
1641 kill(td->pid, SIGTERM);
1647 * start created threads (TD_INITIALIZED -> TD_RUNNING).
1649 for_each_td(td, i) {
1650 if (td->runstate != TD_INITIALIZED)
1653 if (in_ramp_time(td))
1654 td_set_runstate(td, TD_RAMP);
1656 td_set_runstate(td, TD_RUNNING);
1659 m_rate += td->o.ratemin[0] + td->o.ratemin[1];
1660 t_rate += td->o.rate[0] + td->o.rate[1];
1662 fio_mutex_up(td->mutex);
1665 reap_threads(&nr_running, &t_rate, &m_rate);
1671 while (nr_running) {
1672 reap_threads(&nr_running, &t_rate, &m_rate);
1680 int main(int argc, char *argv[])
1685 init_rand(&__fio_rand_state);
1688 * We need locale for number printing, if it isn't set then just
1689 * go with the US format.
1691 if (!getenv("LC_NUMERIC"))
1692 setlocale(LC_NUMERIC, "en_US");
1694 ps = sysconf(_SC_PAGESIZE);
1696 log_err("Failed to get page size\n");
1703 fio_keywords_init();
1705 if (parse_options(argc, argv))
1708 if (exec_profile && load_profile(exec_profile))
1715 setup_log(&agg_io_log[DDIR_READ]);
1716 setup_log(&agg_io_log[DDIR_WRITE]);
1719 startup_mutex = fio_mutex_init(0);
1720 if (startup_mutex == NULL)
1722 writeout_mutex = fio_mutex_init(1);
1723 if (writeout_mutex == NULL)
1727 create_disk_util_thread();
1729 cgroup_list = smalloc(sizeof(*cgroup_list));
1730 INIT_FLIST_HEAD(cgroup_list);
1737 __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1738 __finish_log(agg_io_log[DDIR_WRITE],
1739 "agg-write_bw.log");
1743 cgroup_kill(cgroup_list);
1747 fio_mutex_remove(startup_mutex);
1748 fio_mutex_remove(writeout_mutex);