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
48 unsigned long page_mask;
49 unsigned long page_size;
51 #define PAGE_ALIGN(buf) \
52 (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
55 int thread_number = 0;
60 unsigned long done_secs = 0;
62 static struct fio_mutex *startup_mutex;
63 static struct fio_mutex *writeout_mutex;
64 static volatile int fio_abort;
65 static int exit_value;
66 static timer_t ival_timer;
67 static pthread_t gtod_thread;
68 static struct flist_head *cgroup_list;
69 static char *cgroup_mnt;
71 struct io_log *agg_io_log[2];
73 #define TERMINATE_ALL (-1)
74 #define JOB_START_TIMEOUT (5 * 1000)
76 void td_set_runstate(struct thread_data *td, int runstate)
78 if (td->runstate == runstate)
81 dprint(FD_PROCESS, "pid=%d: runstate %d -> %d\n", (int) td->pid,
82 td->runstate, runstate);
83 td->runstate = runstate;
86 static void terminate_threads(int group_id)
88 struct thread_data *td;
91 dprint(FD_PROCESS, "terminate group_id=%d\n", group_id);
94 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
95 dprint(FD_PROCESS, "setting terminate on %s/%d\n",
96 td->o.name, (int) td->pid);
98 td->o.start_delay = 0;
101 * if the thread is running, just let it exit
103 if (td->runstate < TD_RUNNING)
104 kill(td->pid, SIGQUIT);
106 struct ioengine_ops *ops = td->io_ops;
108 if (ops && (ops->flags & FIO_SIGQUIT))
109 kill(td->pid, SIGQUIT);
115 static void status_timer_arm(void)
117 struct itimerspec value;
119 value.it_value.tv_sec = 0;
120 value.it_value.tv_nsec = DISK_UTIL_MSEC * 1000000;
121 value.it_interval.tv_sec = 0;
122 value.it_interval.tv_nsec = DISK_UTIL_MSEC * 1000000;
124 timer_settime(ival_timer, 0, &value, NULL);
127 static void ival_fn(union sigval sig)
131 print_thread_status();
136 * Happens on thread runs with ctrl-c, ignore our own SIGQUIT
138 static void sig_quit(int sig)
142 static void sig_int(int sig)
145 log_info("\nfio: terminating on signal %d\n", sig);
147 terminate_threads(TERMINATE_ALL);
151 static void posix_timer_teardown(void)
153 timer_delete(ival_timer);
156 static void posix_timer_setup(void)
160 memset(&evt, 0, sizeof(evt));
161 evt.sigev_notify = SIGEV_THREAD;
162 evt.sigev_notify_function = ival_fn;
164 if (timer_create(CLOCK_MONOTONIC, &evt, &ival_timer) < 0)
165 perror("timer_create");
168 static void set_sig_handlers(void)
170 struct sigaction act;
172 memset(&act, 0, sizeof(act));
173 act.sa_handler = sig_int;
174 act.sa_flags = SA_RESTART;
175 sigaction(SIGINT, &act, NULL);
177 memset(&act, 0, sizeof(act));
178 act.sa_handler = sig_quit;
179 act.sa_flags = SA_RESTART;
180 sigaction(SIGQUIT, &act, NULL);
184 * Check if we are above the minimum rate given.
186 static int __check_min_rate(struct thread_data *td, struct timeval *now,
189 unsigned long long bytes = 0;
190 unsigned long iops = 0;
193 unsigned int ratemin = 0;
194 unsigned int rate_iops = 0;
195 unsigned int rate_iops_min = 0;
197 assert(ddir_rw(ddir));
199 if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
203 * allow a 2 second settle period in the beginning
205 if (mtime_since(&td->start, now) < 2000)
208 iops += td->io_blocks[ddir];
209 bytes += td->this_io_bytes[ddir];
210 ratemin += td->o.ratemin[ddir];
211 rate_iops += td->o.rate_iops[ddir];
212 rate_iops_min += td->o.rate_iops_min[ddir];
215 * if rate blocks is set, sample is running
217 if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
218 spent = mtime_since(&td->lastrate[ddir], now);
219 if (spent < td->o.ratecycle)
222 if (td->o.rate[ddir]) {
224 * check bandwidth specified rate
226 if (bytes < td->rate_bytes[ddir]) {
227 log_err("%s: min rate %u not met\n", td->o.name,
231 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
232 if (rate < ratemin ||
233 bytes < td->rate_bytes[ddir]) {
234 log_err("%s: min rate %u not met, got"
235 " %luKB/sec\n", td->o.name,
242 * checks iops specified rate
244 if (iops < rate_iops) {
245 log_err("%s: min iops rate %u not met\n",
246 td->o.name, rate_iops);
249 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
250 if (rate < rate_iops_min ||
251 iops < td->rate_blocks[ddir]) {
252 log_err("%s: min iops rate %u not met,"
253 " got %lu\n", td->o.name,
254 rate_iops_min, rate);
260 td->rate_bytes[ddir] = bytes;
261 td->rate_blocks[ddir] = iops;
262 memcpy(&td->lastrate[ddir], now, sizeof(*now));
266 static int check_min_rate(struct thread_data *td, struct timeval *now,
267 unsigned long *bytes_done)
272 ret |= __check_min_rate(td, now, 0);
274 ret |= __check_min_rate(td, now, 1);
279 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
283 if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
290 * When job exits, we can cancel the in-flight IO if we are using async
291 * io. Attempt to do so.
293 static void cleanup_pending_aio(struct thread_data *td)
295 struct flist_head *entry, *n;
300 * get immediately available events, if any
302 r = io_u_queued_complete(td, 0, NULL);
307 * now cancel remaining active events
309 if (td->io_ops->cancel) {
310 flist_for_each_safe(entry, n, &td->io_u_busylist) {
311 io_u = flist_entry(entry, struct io_u, list);
314 * if the io_u isn't in flight, then that generally
315 * means someone leaked an io_u. complain but fix
316 * it up, so we don't stall here.
318 if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
319 log_err("fio: non-busy IO on busy list\n");
322 r = td->io_ops->cancel(td, io_u);
330 r = io_u_queued_complete(td, td->cur_depth, NULL);
334 * Helper to handle the final sync of a file. Works just like the normal
335 * io path, just does everything sync.
337 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
339 struct io_u *io_u = __get_io_u(td);
345 io_u->ddir = DDIR_SYNC;
348 if (td_io_prep(td, io_u)) {
354 ret = td_io_queue(td, io_u);
356 td_verror(td, io_u->error, "td_io_queue");
359 } else if (ret == FIO_Q_QUEUED) {
360 if (io_u_queued_complete(td, 1, NULL) < 0)
362 } else if (ret == FIO_Q_COMPLETED) {
364 td_verror(td, io_u->error, "td_io_queue");
368 if (io_u_sync_complete(td, io_u, NULL) < 0)
370 } else if (ret == FIO_Q_BUSY) {
371 if (td_io_commit(td))
379 static inline void update_tv_cache(struct thread_data *td)
381 if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
382 fio_gettime(&td->tv_cache, NULL);
385 static int break_on_this_error(struct thread_data *td, int *retptr)
389 if (ret < 0 || td->error) {
392 if (!td->o.continue_on_error)
400 if (td_non_fatal_error(err)) {
402 * Continue with the I/Os in case of
405 update_error_count(td, err);
409 } else if (td->o.fill_device && err == ENOSPC) {
411 * We expect to hit this error if
412 * fill_device option is set.
419 * Stop the I/O in case of a fatal
422 update_error_count(td, err);
431 * The main verify engine. Runs over the writes we previously submitted,
432 * reads the blocks back in, and checks the crc/md5 of the data.
434 static void do_verify(struct thread_data *td)
441 dprint(FD_VERIFY, "starting loop\n");
444 * sync io first and invalidate cache, to make sure we really
447 for_each_file(td, f, i) {
448 if (!fio_file_open(f))
450 if (fio_io_sync(td, f))
452 if (file_invalidate_cache(td, f))
459 td_set_runstate(td, TD_VERIFYING);
462 while (!td->terminate) {
467 if (runtime_exceeded(td, &td->tv_cache)) {
472 io_u = __get_io_u(td);
476 if (get_next_verify(td, io_u)) {
481 if (td_io_prep(td, io_u)) {
486 if (td->o.verify_async)
487 io_u->end_io = verify_io_u_async;
489 io_u->end_io = verify_io_u;
491 ret = td_io_queue(td, io_u);
493 case FIO_Q_COMPLETED:
496 clear_io_u(td, io_u);
497 } else if (io_u->resid) {
498 int bytes = io_u->xfer_buflen - io_u->resid;
499 struct fio_file *f = io_u->file;
505 td_verror(td, EIO, "full resid");
510 io_u->xfer_buflen = io_u->resid;
511 io_u->xfer_buf += bytes;
512 io_u->offset += bytes;
514 if (ddir_rw(io_u->ddir))
515 td->ts.short_io_u[io_u->ddir]++;
517 if (io_u->offset == f->real_file_size)
520 requeue_io_u(td, &io_u);
523 ret = io_u_sync_complete(td, io_u, NULL);
531 requeue_io_u(td, &io_u);
532 ret2 = td_io_commit(td);
538 td_verror(td, -ret, "td_io_queue");
542 if (break_on_this_error(td, &ret))
546 * if we can queue more, do so. but check if there are
547 * completed io_u's first.
549 full = queue_full(td) || ret == FIO_Q_BUSY;
550 if (full || !td->o.iodepth_batch_complete) {
551 min_events = min(td->o.iodepth_batch_complete,
553 if (full && !min_events)
558 * Reap required number of io units, if any,
559 * and do the verification on them through
560 * the callback handler
562 if (io_u_queued_complete(td, min_events, NULL) < 0) {
566 } while (full && (td->cur_depth > td->o.iodepth_low));
573 min_events = td->cur_depth;
576 ret = io_u_queued_complete(td, min_events, NULL);
578 cleanup_pending_aio(td);
580 td_set_runstate(td, TD_RUNNING);
582 dprint(FD_VERIFY, "exiting loop\n");
586 * Main IO worker function. It retrieves io_u's to process and queues
587 * and reaps them, checking for rate and errors along the way.
589 static void do_io(struct thread_data *td)
594 if (in_ramp_time(td))
595 td_set_runstate(td, TD_RAMP);
597 td_set_runstate(td, TD_RUNNING);
599 while ( (td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
600 (!flist_empty(&td->trim_list)) ||
601 ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) ) {
602 struct timeval comp_time;
603 unsigned long bytes_done[2] = { 0, 0 };
613 if (runtime_exceeded(td, &td->tv_cache)) {
623 * Add verification end_io handler, if asked to verify
624 * a previously written file.
626 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
628 if (td->o.verify_async)
629 io_u->end_io = verify_io_u_async;
631 io_u->end_io = verify_io_u;
632 td_set_runstate(td, TD_VERIFYING);
633 } else if (in_ramp_time(td))
634 td_set_runstate(td, TD_RAMP);
636 td_set_runstate(td, TD_RUNNING);
638 ret = td_io_queue(td, io_u);
640 case FIO_Q_COMPLETED:
643 clear_io_u(td, io_u);
644 } else if (io_u->resid) {
645 int bytes = io_u->xfer_buflen - io_u->resid;
646 struct fio_file *f = io_u->file;
652 td_verror(td, EIO, "full resid");
657 io_u->xfer_buflen = io_u->resid;
658 io_u->xfer_buf += bytes;
659 io_u->offset += bytes;
661 if (ddir_rw(io_u->ddir))
662 td->ts.short_io_u[io_u->ddir]++;
664 if (io_u->offset == f->real_file_size)
667 requeue_io_u(td, &io_u);
670 if (__should_check_rate(td, 0) ||
671 __should_check_rate(td, 1))
672 fio_gettime(&comp_time, NULL);
674 ret = io_u_sync_complete(td, io_u, bytes_done);
681 * if the engine doesn't have a commit hook,
682 * the io_u is really queued. if it does have such
683 * a hook, it has to call io_u_queued() itself.
685 if (td->io_ops->commit == NULL)
686 io_u_queued(td, io_u);
689 requeue_io_u(td, &io_u);
690 ret2 = td_io_commit(td);
700 if (break_on_this_error(td, &ret))
704 * See if we need to complete some commands
706 full = queue_full(td) || ret == FIO_Q_BUSY;
707 if (full || !td->o.iodepth_batch_complete) {
708 min_evts = min(td->o.iodepth_batch_complete,
710 if (full && !min_evts)
713 if (__should_check_rate(td, 0) ||
714 __should_check_rate(td, 1))
715 fio_gettime(&comp_time, NULL);
718 ret = io_u_queued_complete(td, min_evts, bytes_done);
722 } while (full && (td->cur_depth > td->o.iodepth_low));
727 if (!(bytes_done[0] + bytes_done[1]))
730 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
731 if (check_min_rate(td, &comp_time, bytes_done)) {
732 if (exitall_on_terminate)
733 terminate_threads(td->groupid);
734 td_verror(td, EIO, "check_min_rate");
739 if (td->o.thinktime) {
740 unsigned long long b;
742 b = td->io_blocks[0] + td->io_blocks[1];
743 if (!(b % td->o.thinktime_blocks)) {
746 if (td->o.thinktime_spin)
747 usec_spin(td->o.thinktime_spin);
749 left = td->o.thinktime - td->o.thinktime_spin;
751 usec_sleep(td, left);
756 if (td->trim_entries)
757 printf("trim entries %ld\n", td->trim_entries);
759 if (td->o.fill_device && td->error == ENOSPC) {
768 ret = io_u_queued_complete(td, i, NULL);
770 if (should_fsync(td) && td->o.end_fsync) {
771 td_set_runstate(td, TD_FSYNCING);
773 for_each_file(td, f, i) {
774 if (!fio_file_open(f))
780 cleanup_pending_aio(td);
783 * stop job if we failed doing any IO
785 if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
789 static void cleanup_io_u(struct thread_data *td)
791 struct flist_head *entry, *n;
794 flist_for_each_safe(entry, n, &td->io_u_freelist) {
795 io_u = flist_entry(entry, struct io_u, list);
797 flist_del(&io_u->list);
804 static int init_io_u(struct thread_data *td)
808 int cl_align, i, max_units;
811 max_units = td->o.iodepth;
812 max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
813 td->orig_buffer_size = (unsigned long long) max_bs
814 * (unsigned long long) max_units;
816 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
819 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
820 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
823 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
824 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
828 if (allocate_io_mem(td))
831 if (td->o.odirect || td->o.mem_align ||
832 (td->io_ops->flags & FIO_RAWIO))
833 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
837 cl_align = os_cache_line_size();
839 for (i = 0; i < max_units; i++) {
845 if (posix_memalign(&ptr, cl_align, sizeof(*io_u))) {
846 log_err("fio: posix_memalign=%s\n", strerror(errno));
851 memset(io_u, 0, sizeof(*io_u));
852 INIT_FLIST_HEAD(&io_u->list);
853 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
855 if (!(td->io_ops->flags & FIO_NOIO)) {
856 io_u->buf = p + max_bs * i;
857 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
859 if (td_write(td) && !td->o.refill_buffers)
860 io_u_fill_buffer(td, io_u, max_bs);
861 else if (td_write(td) && td->o.verify_pattern_bytes) {
863 * Fill the buffer with the pattern if we are
864 * going to be doing writes.
866 fill_pattern(td, io_u->buf, max_bs, io_u);
871 io_u->flags = IO_U_F_FREE;
872 flist_add(&io_u->list, &td->io_u_freelist);
878 static int switch_ioscheduler(struct thread_data *td)
880 char tmp[256], tmp2[128];
884 if (td->io_ops->flags & FIO_DISKLESSIO)
887 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
889 f = fopen(tmp, "r+");
891 if (errno == ENOENT) {
892 log_err("fio: os or kernel doesn't support IO scheduler"
896 td_verror(td, errno, "fopen iosched");
903 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
904 if (ferror(f) || ret != 1) {
905 td_verror(td, errno, "fwrite");
913 * Read back and check that the selected scheduler is now the default.
915 ret = fread(tmp, 1, sizeof(tmp), f);
916 if (ferror(f) || ret < 0) {
917 td_verror(td, errno, "fread");
922 sprintf(tmp2, "[%s]", td->o.ioscheduler);
923 if (!strstr(tmp, tmp2)) {
924 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
925 td_verror(td, EINVAL, "iosched_switch");
934 static int keep_running(struct thread_data *td)
936 unsigned long long io_done;
940 if (td->o.time_based)
947 io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
949 if (io_done < td->o.size)
955 static void reset_io_counters(struct thread_data *td)
957 td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
958 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
960 td->rate_bytes[0] = td->rate_bytes[1] = 0;
961 td->rate_blocks[0] = td->rate_blocks[1] = 0;
963 td->last_was_sync = 0;
966 * reset file done count if we are to start over
968 if (td->o.time_based || td->o.loops)
969 td->nr_done_files = 0;
972 * Set the same seed to get repeatable runs
974 td_fill_rand_seeds(td);
977 void reset_all_stats(struct thread_data *td)
982 reset_io_counters(td);
984 for (i = 0; i < 2; i++) {
986 td->io_blocks[i] = 0;
987 td->io_issues[i] = 0;
988 td->ts.total_io_u[i] = 0;
991 fio_gettime(&tv, NULL);
992 memcpy(&td->epoch, &tv, sizeof(tv));
993 memcpy(&td->start, &tv, sizeof(tv));
996 static void clear_io_state(struct thread_data *td)
1001 reset_io_counters(td);
1004 for_each_file(td, f, i)
1005 fio_file_clear_done(f);
1008 static int exec_string(const char *string)
1010 int ret, newlen = strlen(string) + 1 + 8;
1013 str = malloc(newlen);
1014 sprintf(str, "sh -c %s", string);
1018 log_err("fio: exec of cmd <%s> failed\n", str);
1025 * Entry point for the thread based jobs. The process based jobs end up
1026 * here as well, after a little setup.
1028 static void *thread_main(void *data)
1030 unsigned long long runtime[2], elapsed;
1031 struct thread_data *td = data;
1032 pthread_condattr_t attr;
1035 if (!td->o.use_thread)
1040 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1042 INIT_FLIST_HEAD(&td->io_u_freelist);
1043 INIT_FLIST_HEAD(&td->io_u_busylist);
1044 INIT_FLIST_HEAD(&td->io_u_requeues);
1045 INIT_FLIST_HEAD(&td->io_log_list);
1046 INIT_FLIST_HEAD(&td->io_hist_list);
1047 INIT_FLIST_HEAD(&td->verify_list);
1048 INIT_FLIST_HEAD(&td->trim_list);
1049 pthread_mutex_init(&td->io_u_lock, NULL);
1050 td->io_hist_tree = RB_ROOT;
1052 pthread_condattr_init(&attr);
1053 pthread_cond_init(&td->verify_cond, &attr);
1054 pthread_cond_init(&td->free_cond, &attr);
1056 td_set_runstate(td, TD_INITIALIZED);
1057 dprint(FD_MUTEX, "up startup_mutex\n");
1058 fio_mutex_up(startup_mutex);
1059 dprint(FD_MUTEX, "wait on td->mutex\n");
1060 fio_mutex_down(td->mutex);
1061 dprint(FD_MUTEX, "done waiting on td->mutex\n");
1064 * the ->mutex mutex is now no longer used, close it to avoid
1065 * eating a file descriptor
1067 fio_mutex_remove(td->mutex);
1069 if (td->o.uid != -1U && setuid(td->o.uid)) {
1070 td_verror(td, errno, "setuid");
1073 if (td->o.gid != -1U && setgid(td->o.gid)) {
1074 td_verror(td, errno, "setgid");
1079 * May alter parameters that init_io_u() will use, so we need to
1088 if (td->o.verify_async && verify_async_init(td))
1091 if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1092 td_verror(td, errno, "cpu_set_affinity");
1097 * If we have a gettimeofday() thread, make sure we exclude that
1098 * thread from this job
1100 if (td->o.gtod_cpu) {
1101 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1102 if (fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1103 td_verror(td, errno, "cpu_set_affinity");
1108 if (td->ioprio_set) {
1109 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1110 td_verror(td, errno, "ioprio_set");
1115 if (td->o.cgroup_weight && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1118 if (nice(td->o.nice) == -1) {
1119 td_verror(td, errno, "nice");
1123 if (td->o.ioscheduler && switch_ioscheduler(td))
1126 if (!td->o.create_serialize && setup_files(td))
1132 if (init_random_map(td))
1135 if (td->o.exec_prerun) {
1136 if (exec_string(td->o.exec_prerun))
1140 if (td->o.pre_read) {
1141 if (pre_read_files(td) < 0)
1145 fio_gettime(&td->epoch, NULL);
1146 getrusage(RUSAGE_SELF, &td->ts.ru_start);
1148 runtime[0] = runtime[1] = 0;
1150 while (keep_running(td)) {
1151 fio_gettime(&td->start, NULL);
1152 memcpy(&td->ts.stat_sample_time[0], &td->start,
1154 memcpy(&td->ts.stat_sample_time[1], &td->start,
1156 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1158 if (td->o.ratemin[0] || td->o.ratemin[1])
1159 memcpy(&td->lastrate, &td->ts.stat_sample_time,
1160 sizeof(td->lastrate));
1165 prune_io_piece_log(td);
1171 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1172 elapsed = utime_since_now(&td->start);
1173 runtime[DDIR_READ] += elapsed;
1175 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1176 elapsed = utime_since_now(&td->start);
1177 runtime[DDIR_WRITE] += elapsed;
1180 if (td->error || td->terminate)
1183 if (!td->o.do_verify ||
1184 td->o.verify == VERIFY_NONE ||
1185 (td->io_ops->flags & FIO_UNIDIR))
1190 fio_gettime(&td->start, NULL);
1194 runtime[DDIR_READ] += utime_since_now(&td->start);
1196 if (td->error || td->terminate)
1200 update_rusage_stat(td);
1201 td->ts.runtime[0] = (runtime[0] + 999) / 1000;
1202 td->ts.runtime[1] = (runtime[1] + 999) / 1000;
1203 td->ts.total_run_time = mtime_since_now(&td->epoch);
1204 td->ts.io_bytes[0] = td->io_bytes[0];
1205 td->ts.io_bytes[1] = td->io_bytes[1];
1207 fio_mutex_down(writeout_mutex);
1208 if (td->ts.bw_log) {
1209 if (td->o.bw_log_file) {
1210 finish_log_named(td, td->ts.bw_log,
1211 td->o.bw_log_file, "bw");
1213 finish_log(td, td->ts.bw_log, "bw");
1215 if (td->ts.lat_log) {
1216 if (td->o.lat_log_file) {
1217 finish_log_named(td, td->ts.lat_log,
1218 td->o.lat_log_file, "lat");
1220 finish_log(td, td->ts.lat_log, "lat");
1222 if (td->ts.slat_log) {
1223 if (td->o.lat_log_file) {
1224 finish_log_named(td, td->ts.slat_log,
1225 td->o.lat_log_file, "slat");
1227 finish_log(td, td->ts.slat_log, "slat");
1229 if (td->ts.clat_log) {
1230 if (td->o.lat_log_file) {
1231 finish_log_named(td, td->ts.clat_log,
1232 td->o.lat_log_file, "clat");
1234 finish_log(td, td->ts.clat_log, "clat");
1236 fio_mutex_up(writeout_mutex);
1237 if (td->o.exec_postrun)
1238 exec_string(td->o.exec_postrun);
1240 if (exitall_on_terminate)
1241 terminate_threads(td->groupid);
1245 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1248 if (td->o.verify_async)
1249 verify_async_exit(td);
1251 close_and_free_files(td);
1254 cgroup_shutdown(td, &cgroup_mnt);
1256 if (td->o.cpumask_set) {
1257 int ret = fio_cpuset_exit(&td->o.cpumask);
1259 td_verror(td, ret, "fio_cpuset_exit");
1263 * do this very late, it will log file closing as well
1265 if (td->o.write_iolog_file)
1266 write_iolog_close(td);
1268 options_mem_free(td);
1269 td_set_runstate(td, TD_EXITED);
1270 return (void *) (unsigned long) td->error;
1274 * We cannot pass the td data into a forked process, so attach the td and
1275 * pass it to the thread worker.
1277 static int fork_main(int shmid, int offset)
1279 struct thread_data *td;
1282 data = shmat(shmid, NULL, 0);
1283 if (data == (void *) -1) {
1290 td = data + offset * sizeof(struct thread_data);
1291 ret = thread_main(td);
1293 return (int) (unsigned long) ret;
1297 * Run over the job map and reap the threads that have exited, if any.
1299 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1301 struct thread_data *td;
1302 int i, cputhreads, realthreads, pending, status, ret;
1305 * reap exited threads (TD_EXITED -> TD_REAPED)
1307 realthreads = pending = cputhreads = 0;
1308 for_each_td(td, i) {
1312 * ->io_ops is NULL for a thread that has closed its
1315 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1324 if (td->runstate == TD_REAPED)
1326 if (td->o.use_thread) {
1327 if (td->runstate == TD_EXITED) {
1328 td_set_runstate(td, TD_REAPED);
1335 if (td->runstate == TD_EXITED)
1339 * check if someone quit or got killed in an unusual way
1341 ret = waitpid(td->pid, &status, flags);
1343 if (errno == ECHILD) {
1344 log_err("fio: pid=%d disappeared %d\n",
1345 (int) td->pid, td->runstate);
1346 td_set_runstate(td, TD_REAPED);
1350 } else if (ret == td->pid) {
1351 if (WIFSIGNALED(status)) {
1352 int sig = WTERMSIG(status);
1355 log_err("fio: pid=%d, got signal=%d\n",
1356 (int) td->pid, sig);
1357 td_set_runstate(td, TD_REAPED);
1360 if (WIFEXITED(status)) {
1361 if (WEXITSTATUS(status) && !td->error)
1362 td->error = WEXITSTATUS(status);
1364 td_set_runstate(td, TD_REAPED);
1370 * thread is not dead, continue
1376 (*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
1377 (*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
1384 done_secs += mtime_since_now(&td->epoch) / 1000;
1387 if (*nr_running == cputhreads && !pending && realthreads)
1388 terminate_threads(TERMINATE_ALL);
1391 static void *gtod_thread_main(void *data)
1393 fio_mutex_up(startup_mutex);
1396 * As long as we have jobs around, update the clock. It would be nice
1397 * to have some way of NOT hammering that CPU with gettimeofday(),
1398 * but I'm not sure what to use outside of a simple CPU nop to relax
1399 * it - we don't want to lose precision.
1409 static int fio_start_gtod_thread(void)
1411 pthread_attr_t attr;
1414 pthread_attr_init(&attr);
1415 pthread_attr_setstacksize(&attr, PTHREAD_STACK_MIN);
1416 ret = pthread_create(>od_thread, &attr, gtod_thread_main, NULL);
1417 pthread_attr_destroy(&attr);
1419 log_err("Can't create gtod thread: %s\n", strerror(ret));
1423 ret = pthread_detach(gtod_thread);
1425 log_err("Can't detatch gtod thread: %s\n", strerror(ret));
1429 dprint(FD_MUTEX, "wait on startup_mutex\n");
1430 fio_mutex_down(startup_mutex);
1431 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1436 * Main function for kicking off and reaping jobs, as needed.
1438 static void run_threads(void)
1440 struct thread_data *td;
1441 unsigned long spent;
1442 int i, todo, nr_running, m_rate, t_rate, nr_started;
1444 if (fio_pin_memory())
1447 if (fio_gtod_offload && fio_start_gtod_thread())
1450 if (!terse_output) {
1451 log_info("Starting ");
1453 log_info("%d thread%s", nr_thread,
1454 nr_thread > 1 ? "s" : "");
1458 log_info("%d process%s", nr_process,
1459 nr_process > 1 ? "es" : "");
1467 todo = thread_number;
1470 m_rate = t_rate = 0;
1472 for_each_td(td, i) {
1473 print_status_init(td->thread_number - 1);
1475 if (!td->o.create_serialize) {
1481 * do file setup here so it happens sequentially,
1482 * we don't want X number of threads getting their
1483 * client data interspersed on disk
1485 if (setup_files(td)) {
1488 log_err("fio: pid=%d, err=%d/%s\n",
1489 (int) td->pid, td->error, td->verror);
1490 td_set_runstate(td, TD_REAPED);
1497 * for sharing to work, each job must always open
1498 * its own files. so close them, if we opened them
1501 for_each_file(td, f, i) {
1502 if (fio_file_open(f))
1503 td_io_close_file(td, f);
1513 struct thread_data *map[MAX_JOBS];
1514 struct timeval this_start;
1515 int this_jobs = 0, left;
1518 * create threads (TD_NOT_CREATED -> TD_CREATED)
1520 for_each_td(td, i) {
1521 if (td->runstate != TD_NOT_CREATED)
1525 * never got a chance to start, killed by other
1526 * thread for some reason
1528 if (td->terminate) {
1533 if (td->o.start_delay) {
1534 spent = mtime_since_genesis();
1536 if (td->o.start_delay * 1000 > spent)
1540 if (td->o.stonewall && (nr_started || nr_running)) {
1541 dprint(FD_PROCESS, "%s: stonewall wait\n",
1547 * Set state to created. Thread will transition
1548 * to TD_INITIALIZED when it's done setting up.
1550 td_set_runstate(td, TD_CREATED);
1551 map[this_jobs++] = td;
1554 if (td->o.use_thread) {
1557 dprint(FD_PROCESS, "will pthread_create\n");
1558 ret = pthread_create(&td->thread, NULL,
1561 log_err("pthread_create: %s\n",
1566 ret = pthread_detach(td->thread);
1568 log_err("pthread_detach: %s",
1572 dprint(FD_PROCESS, "will fork\n");
1575 int ret = fork_main(shm_id, i);
1578 } else if (i == fio_debug_jobno)
1579 *fio_debug_jobp = pid;
1581 dprint(FD_MUTEX, "wait on startup_mutex\n");
1582 if (fio_mutex_down_timeout(startup_mutex, 10)) {
1583 log_err("fio: job startup hung? exiting.\n");
1584 terminate_threads(TERMINATE_ALL);
1589 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1593 * Wait for the started threads to transition to
1596 fio_gettime(&this_start, NULL);
1598 while (left && !fio_abort) {
1599 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1604 for (i = 0; i < this_jobs; i++) {
1608 if (td->runstate == TD_INITIALIZED) {
1611 } else if (td->runstate >= TD_EXITED) {
1615 nr_running++; /* work-around... */
1621 log_err("fio: %d jobs failed to start\n", left);
1622 for (i = 0; i < this_jobs; i++) {
1626 kill(td->pid, SIGTERM);
1632 * start created threads (TD_INITIALIZED -> TD_RUNNING).
1634 for_each_td(td, i) {
1635 if (td->runstate != TD_INITIALIZED)
1638 if (in_ramp_time(td))
1639 td_set_runstate(td, TD_RAMP);
1641 td_set_runstate(td, TD_RUNNING);
1644 m_rate += td->o.ratemin[0] + td->o.ratemin[1];
1645 t_rate += td->o.rate[0] + td->o.rate[1];
1647 fio_mutex_up(td->mutex);
1650 reap_threads(&nr_running, &t_rate, &m_rate);
1656 while (nr_running) {
1657 reap_threads(&nr_running, &t_rate, &m_rate);
1665 int main(int argc, char *argv[])
1670 init_rand(&__fio_rand_state);
1673 * We need locale for number printing, if it isn't set then just
1674 * go with the US format.
1676 if (!getenv("LC_NUMERIC"))
1677 setlocale(LC_NUMERIC, "en_US");
1679 ps = sysconf(_SC_PAGESIZE);
1681 log_err("Failed to get page size\n");
1688 fio_keywords_init();
1690 if (parse_options(argc, argv))
1693 if (exec_profile && load_profile(exec_profile))
1700 setup_log(&agg_io_log[DDIR_READ]);
1701 setup_log(&agg_io_log[DDIR_WRITE]);
1704 startup_mutex = fio_mutex_init(0);
1705 writeout_mutex = fio_mutex_init(1);
1709 posix_timer_setup();
1712 cgroup_list = smalloc(sizeof(*cgroup_list));
1713 INIT_FLIST_HEAD(cgroup_list);
1720 __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1721 __finish_log(agg_io_log[DDIR_WRITE],
1722 "agg-write_bw.log");
1726 cgroup_kill(cgroup_list);
1730 posix_timer_teardown();
1731 fio_mutex_remove(startup_mutex);
1732 fio_mutex_remove(writeout_mutex);