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
44 unsigned long page_mask;
45 unsigned long page_size;
47 #define PAGE_ALIGN(buf) \
48 (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
51 int thread_number = 0;
56 unsigned long done_secs = 0;
58 static struct fio_mutex *startup_mutex;
59 static struct fio_mutex *writeout_mutex;
60 static volatile int fio_abort;
61 static int exit_value;
62 static struct itimerval itimer;
63 static pthread_t gtod_thread;
64 static struct flist_head *cgroup_list;
66 struct io_log *agg_io_log[2];
68 #define TERMINATE_ALL (-1)
69 #define JOB_START_TIMEOUT (5 * 1000)
71 void td_set_runstate(struct thread_data *td, int runstate)
73 if (td->runstate == runstate)
76 dprint(FD_PROCESS, "pid=%d: runstate %d -> %d\n", (int) td->pid,
77 td->runstate, runstate);
78 td->runstate = runstate;
81 static void terminate_threads(int group_id)
83 struct thread_data *td;
86 dprint(FD_PROCESS, "terminate group_id=%d\n", group_id);
89 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
90 dprint(FD_PROCESS, "setting terminate on %s/%d\n",
91 td->o.name, (int) td->pid);
93 td->o.start_delay = 0;
96 * if the thread is running, just let it exit
98 if (td->runstate < TD_RUNNING)
99 kill(td->pid, SIGQUIT);
101 struct ioengine_ops *ops = td->io_ops;
103 if (ops && (ops->flags & FIO_SIGQUIT))
104 kill(td->pid, SIGQUIT);
110 static void status_timer_arm(void)
112 itimer.it_value.tv_sec = 0;
113 itimer.it_value.tv_usec = DISK_UTIL_MSEC * 1000;
114 setitimer(ITIMER_REAL, &itimer, NULL);
117 static void sig_alrm(int fio_unused sig)
121 print_thread_status();
127 * Happens on thread runs with ctrl-c, ignore our own SIGQUIT
129 static void sig_quit(int sig)
133 static void sig_int(int sig)
136 printf("\nfio: terminating on signal %d\n", sig);
138 terminate_threads(TERMINATE_ALL);
142 static void sig_ill(int fio_unused sig)
147 log_err("fio: illegal instruction. your cpu does not support "
148 "the sse4.2 instruction for crc32c\n");
149 terminate_threads(TERMINATE_ALL);
153 static void set_sig_handlers(void)
155 struct sigaction act;
157 memset(&act, 0, sizeof(act));
158 act.sa_handler = sig_alrm;
159 act.sa_flags = SA_RESTART;
160 sigaction(SIGALRM, &act, NULL);
162 memset(&act, 0, sizeof(act));
163 act.sa_handler = sig_int;
164 act.sa_flags = SA_RESTART;
165 sigaction(SIGINT, &act, NULL);
167 memset(&act, 0, sizeof(act));
168 act.sa_handler = sig_ill;
169 act.sa_flags = SA_RESTART;
170 sigaction(SIGILL, &act, NULL);
172 memset(&act, 0, sizeof(act));
173 act.sa_handler = sig_quit;
174 act.sa_flags = SA_RESTART;
175 sigaction(SIGQUIT, &act, NULL);
179 * Check if we are above the minimum rate given.
181 static int __check_min_rate(struct thread_data *td, struct timeval *now,
184 unsigned long long bytes = 0;
185 unsigned long iops = 0;
188 unsigned int ratemin = 0;
189 unsigned int rate_iops = 0;
190 unsigned int rate_iops_min = 0;
192 if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
196 * allow a 2 second settle period in the beginning
198 if (mtime_since(&td->start, now) < 2000)
201 iops += td->io_blocks[ddir];
202 bytes += td->this_io_bytes[ddir];
203 ratemin += td->o.ratemin[ddir];
204 rate_iops += td->o.rate_iops[ddir];
205 rate_iops_min += td->o.rate_iops_min[ddir];
208 * if rate blocks is set, sample is running
210 if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
211 spent = mtime_since(&td->lastrate[ddir], now);
212 if (spent < td->o.ratecycle)
215 if (td->o.rate[ddir]) {
217 * check bandwidth specified rate
219 if (bytes < td->rate_bytes[ddir]) {
220 log_err("%s: min rate %u not met\n", td->o.name,
224 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
225 if (rate < ratemin ||
226 bytes < td->rate_bytes[ddir]) {
227 log_err("%s: min rate %u not met, got"
228 " %luKB/sec\n", td->o.name,
235 * checks iops specified rate
237 if (iops < rate_iops) {
238 log_err("%s: min iops rate %u not met\n",
239 td->o.name, rate_iops);
242 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
243 if (rate < rate_iops_min ||
244 iops < td->rate_blocks[ddir]) {
245 log_err("%s: min iops rate %u not met,"
246 " got %lu\n", td->o.name,
247 rate_iops_min, rate);
253 td->rate_bytes[ddir] = bytes;
254 td->rate_blocks[ddir] = iops;
255 memcpy(&td->lastrate[ddir], now, sizeof(*now));
259 static int check_min_rate(struct thread_data *td, struct timeval *now,
260 unsigned long *bytes_done)
265 ret |= __check_min_rate(td, now, 0);
267 ret |= __check_min_rate(td, now, 1);
272 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
276 if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
283 * When job exits, we can cancel the in-flight IO if we are using async
284 * io. Attempt to do so.
286 static void cleanup_pending_aio(struct thread_data *td)
288 struct flist_head *entry, *n;
293 * get immediately available events, if any
295 r = io_u_queued_complete(td, 0, NULL);
300 * now cancel remaining active events
302 if (td->io_ops->cancel) {
303 flist_for_each_safe(entry, n, &td->io_u_busylist) {
304 io_u = flist_entry(entry, struct io_u, list);
307 * if the io_u isn't in flight, then that generally
308 * means someone leaked an io_u. complain but fix
309 * it up, so we don't stall here.
311 if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
312 log_err("fio: non-busy IO on busy list\n");
315 r = td->io_ops->cancel(td, io_u);
323 r = io_u_queued_complete(td, td->cur_depth, NULL);
327 * Helper to handle the final sync of a file. Works just like the normal
328 * io path, just does everything sync.
330 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
332 struct io_u *io_u = __get_io_u(td);
338 io_u->ddir = DDIR_SYNC;
341 if (td_io_prep(td, io_u)) {
347 ret = td_io_queue(td, io_u);
349 td_verror(td, io_u->error, "td_io_queue");
352 } else if (ret == FIO_Q_QUEUED) {
353 if (io_u_queued_complete(td, 1, NULL) < 0)
355 } else if (ret == FIO_Q_COMPLETED) {
357 td_verror(td, io_u->error, "td_io_queue");
361 if (io_u_sync_complete(td, io_u, NULL) < 0)
363 } else if (ret == FIO_Q_BUSY) {
364 if (td_io_commit(td))
372 static inline void update_tv_cache(struct thread_data *td)
374 if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
375 fio_gettime(&td->tv_cache, NULL);
378 static int break_on_this_error(struct thread_data *td, int *retptr)
382 if (ret < 0 || td->error) {
385 if (!td->o.continue_on_error)
393 if (td_non_fatal_error(err)) {
395 * Continue with the I/Os in case of
398 update_error_count(td, err);
402 } else if (td->o.fill_device && err == ENOSPC) {
404 * We expect to hit this error if
405 * fill_device option is set.
412 * Stop the I/O in case of a fatal
415 update_error_count(td, err);
424 * The main verify engine. Runs over the writes we previously submitted,
425 * reads the blocks back in, and checks the crc/md5 of the data.
427 static void do_verify(struct thread_data *td)
435 * sync io first and invalidate cache, to make sure we really
438 for_each_file(td, f, i) {
439 if (!fio_file_open(f))
441 if (fio_io_sync(td, f))
443 if (file_invalidate_cache(td, f))
450 td_set_runstate(td, TD_VERIFYING);
453 while (!td->terminate) {
458 if (runtime_exceeded(td, &td->tv_cache)) {
463 io_u = __get_io_u(td);
467 if (get_next_verify(td, io_u)) {
472 if (td_io_prep(td, io_u)) {
477 if (td->o.verify_async)
478 io_u->end_io = verify_io_u_async;
480 io_u->end_io = verify_io_u;
482 ret = td_io_queue(td, io_u);
484 case FIO_Q_COMPLETED:
487 clear_io_u(td, io_u);
488 } else if (io_u->resid) {
489 int bytes = io_u->xfer_buflen - io_u->resid;
490 struct fio_file *f = io_u->file;
496 td_verror(td, EIO, "full resid");
501 io_u->xfer_buflen = io_u->resid;
502 io_u->xfer_buf += bytes;
503 io_u->offset += bytes;
505 td->ts.short_io_u[io_u->ddir]++;
507 if (io_u->offset == f->real_file_size)
510 requeue_io_u(td, &io_u);
513 ret = io_u_sync_complete(td, io_u, NULL);
521 requeue_io_u(td, &io_u);
522 ret2 = td_io_commit(td);
528 td_verror(td, -ret, "td_io_queue");
532 if (break_on_this_error(td, &ret))
536 * if we can queue more, do so. but check if there are
537 * completed io_u's first.
539 full = queue_full(td) || ret == FIO_Q_BUSY;
540 if (full || !td->o.iodepth_batch_complete) {
541 min_events = min(td->o.iodepth_batch_complete,
543 if (full && !min_events)
548 * Reap required number of io units, if any,
549 * and do the verification on them through
550 * the callback handler
552 if (io_u_queued_complete(td, min_events, NULL) < 0) {
556 } while (full && (td->cur_depth > td->o.iodepth_low));
563 min_events = td->cur_depth;
566 ret = io_u_queued_complete(td, min_events, NULL);
568 cleanup_pending_aio(td);
570 td_set_runstate(td, TD_RUNNING);
574 * Main IO worker function. It retrieves io_u's to process and queues
575 * and reaps them, checking for rate and errors along the way.
577 static void do_io(struct thread_data *td)
582 if (in_ramp_time(td))
583 td_set_runstate(td, TD_RAMP);
585 td_set_runstate(td, TD_RUNNING);
587 while ( (td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
588 ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) ) {
589 struct timeval comp_time;
590 unsigned long bytes_done[2] = { 0, 0 };
600 if (runtime_exceeded(td, &td->tv_cache)) {
610 * Add verification end_io handler, if asked to verify
611 * a previously written file.
613 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
615 if (td->o.verify_async)
616 io_u->end_io = verify_io_u_async;
618 io_u->end_io = verify_io_u;
619 td_set_runstate(td, TD_VERIFYING);
620 } else if (in_ramp_time(td))
621 td_set_runstate(td, TD_RAMP);
623 td_set_runstate(td, TD_RUNNING);
625 ret = td_io_queue(td, io_u);
627 case FIO_Q_COMPLETED:
630 clear_io_u(td, io_u);
631 } else if (io_u->resid) {
632 int bytes = io_u->xfer_buflen - io_u->resid;
633 struct fio_file *f = io_u->file;
639 td_verror(td, EIO, "full resid");
644 io_u->xfer_buflen = io_u->resid;
645 io_u->xfer_buf += bytes;
646 io_u->offset += bytes;
648 td->ts.short_io_u[io_u->ddir]++;
650 if (io_u->offset == f->real_file_size)
653 requeue_io_u(td, &io_u);
656 if (__should_check_rate(td, 0) ||
657 __should_check_rate(td, 1))
658 fio_gettime(&comp_time, NULL);
660 ret = io_u_sync_complete(td, io_u, bytes_done);
667 * if the engine doesn't have a commit hook,
668 * the io_u is really queued. if it does have such
669 * a hook, it has to call io_u_queued() itself.
671 if (td->io_ops->commit == NULL)
672 io_u_queued(td, io_u);
675 requeue_io_u(td, &io_u);
676 ret2 = td_io_commit(td);
686 if (break_on_this_error(td, &ret))
690 * See if we need to complete some commands
692 full = queue_full(td) || ret == FIO_Q_BUSY;
693 if (full || !td->o.iodepth_batch_complete) {
694 min_evts = min(td->o.iodepth_batch_complete,
696 if (full && !min_evts)
699 if (__should_check_rate(td, 0) ||
700 __should_check_rate(td, 1))
701 fio_gettime(&comp_time, NULL);
704 ret = io_u_queued_complete(td, min_evts, bytes_done);
708 } while (full && (td->cur_depth > td->o.iodepth_low));
713 if (!(bytes_done[0] + bytes_done[1]))
716 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
717 if (check_min_rate(td, &comp_time, bytes_done)) {
718 if (exitall_on_terminate)
719 terminate_threads(td->groupid);
720 td_verror(td, EIO, "check_min_rate");
725 if (td->o.thinktime) {
726 unsigned long long b;
728 b = td->io_blocks[0] + td->io_blocks[1];
729 if (!(b % td->o.thinktime_blocks)) {
732 if (td->o.thinktime_spin)
733 usec_spin(td->o.thinktime_spin);
735 left = td->o.thinktime - td->o.thinktime_spin;
737 usec_sleep(td, left);
742 if (td->o.fill_device && td->error == ENOSPC) {
751 ret = io_u_queued_complete(td, i, NULL);
753 if (should_fsync(td) && td->o.end_fsync) {
754 td_set_runstate(td, TD_FSYNCING);
756 for_each_file(td, f, i) {
757 if (!fio_file_open(f))
763 cleanup_pending_aio(td);
766 * stop job if we failed doing any IO
768 if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
772 static void cleanup_io_u(struct thread_data *td)
774 struct flist_head *entry, *n;
777 flist_for_each_safe(entry, n, &td->io_u_freelist) {
778 io_u = flist_entry(entry, struct io_u, list);
780 flist_del(&io_u->list);
787 static int init_io_u(struct thread_data *td)
791 int cl_align, i, max_units;
794 max_units = td->o.iodepth;
795 max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
796 td->orig_buffer_size = (unsigned long long) max_bs
797 * (unsigned long long) max_units;
799 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
802 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
803 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
806 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
807 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
811 if (allocate_io_mem(td))
814 if (td->o.odirect || td->o.mem_align)
815 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
819 cl_align = os_cache_line_size();
821 for (i = 0; i < max_units; i++) {
827 if (posix_memalign(&ptr, cl_align, sizeof(*io_u))) {
828 log_err("fio: posix_memalign=%s\n", strerror(errno));
833 memset(io_u, 0, sizeof(*io_u));
834 INIT_FLIST_HEAD(&io_u->list);
835 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
837 if (!(td->io_ops->flags & FIO_NOIO)) {
838 io_u->buf = p + max_bs * i;
839 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
841 if (td_write(td) && !td->o.refill_buffers)
842 io_u_fill_buffer(td, io_u, max_bs);
846 io_u->flags = IO_U_F_FREE;
847 flist_add(&io_u->list, &td->io_u_freelist);
853 static int switch_ioscheduler(struct thread_data *td)
855 char tmp[256], tmp2[128];
859 if (td->io_ops->flags & FIO_DISKLESSIO)
862 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
864 f = fopen(tmp, "r+");
866 if (errno == ENOENT) {
867 log_err("fio: os or kernel doesn't support IO scheduler"
871 td_verror(td, errno, "fopen iosched");
878 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
879 if (ferror(f) || ret != 1) {
880 td_verror(td, errno, "fwrite");
888 * Read back and check that the selected scheduler is now the default.
890 ret = fread(tmp, 1, sizeof(tmp), f);
891 if (ferror(f) || ret < 0) {
892 td_verror(td, errno, "fread");
897 sprintf(tmp2, "[%s]", td->o.ioscheduler);
898 if (!strstr(tmp, tmp2)) {
899 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
900 td_verror(td, EINVAL, "iosched_switch");
909 static int keep_running(struct thread_data *td)
911 unsigned long long io_done;
915 if (td->o.time_based)
922 io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
924 if (io_done < td->o.size)
930 static void reset_io_counters(struct thread_data *td)
932 td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
933 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
935 td->rate_bytes[0] = td->rate_bytes[1] = 0;
936 td->rate_blocks[0] = td->rate_blocks[1] = 0;
938 td->last_was_sync = 0;
941 * reset file done count if we are to start over
943 if (td->o.time_based || td->o.loops)
944 td->nr_done_files = 0;
947 * Set the same seed to get repeatable runs
949 td_fill_rand_seeds(td);
952 void reset_all_stats(struct thread_data *td)
957 reset_io_counters(td);
959 for (i = 0; i < 2; i++) {
961 td->io_blocks[i] = 0;
962 td->io_issues[i] = 0;
963 td->ts.total_io_u[i] = 0;
966 fio_gettime(&tv, NULL);
967 memcpy(&td->epoch, &tv, sizeof(tv));
968 memcpy(&td->start, &tv, sizeof(tv));
971 static void clear_io_state(struct thread_data *td)
976 reset_io_counters(td);
979 for_each_file(td, f, i)
980 fio_file_clear_done(f);
983 static int exec_string(const char *string)
985 int ret, newlen = strlen(string) + 1 + 8;
988 str = malloc(newlen);
989 sprintf(str, "sh -c %s", string);
993 log_err("fio: exec of cmd <%s> failed\n", str);
1000 * Entry point for the thread based jobs. The process based jobs end up
1001 * here as well, after a little setup.
1003 static void *thread_main(void *data)
1005 unsigned long long runtime[2], elapsed;
1006 struct thread_data *td = data;
1007 pthread_condattr_t attr;
1010 if (!td->o.use_thread)
1015 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1017 INIT_FLIST_HEAD(&td->io_u_freelist);
1018 INIT_FLIST_HEAD(&td->io_u_busylist);
1019 INIT_FLIST_HEAD(&td->io_u_requeues);
1020 INIT_FLIST_HEAD(&td->io_log_list);
1021 INIT_FLIST_HEAD(&td->io_hist_list);
1022 INIT_FLIST_HEAD(&td->verify_list);
1023 pthread_mutex_init(&td->io_u_lock, NULL);
1024 td->io_hist_tree = RB_ROOT;
1026 pthread_condattr_init(&attr);
1027 pthread_cond_init(&td->verify_cond, &attr);
1028 pthread_cond_init(&td->free_cond, &attr);
1030 td_set_runstate(td, TD_INITIALIZED);
1031 dprint(FD_MUTEX, "up startup_mutex\n");
1032 fio_mutex_up(startup_mutex);
1033 dprint(FD_MUTEX, "wait on td->mutex\n");
1034 fio_mutex_down(td->mutex);
1035 dprint(FD_MUTEX, "done waiting on td->mutex\n");
1038 * the ->mutex mutex is now no longer used, close it to avoid
1039 * eating a file descriptor
1041 fio_mutex_remove(td->mutex);
1044 * May alter parameters that init_io_u() will use, so we need to
1053 if (td->o.verify_async && verify_async_init(td))
1056 if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1057 td_verror(td, errno, "cpu_set_affinity");
1062 * If we have a gettimeofday() thread, make sure we exclude that
1063 * thread from this job
1065 if (td->o.gtod_cpu) {
1066 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1067 if (fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1068 td_verror(td, errno, "cpu_set_affinity");
1073 if (td->ioprio_set) {
1074 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1075 td_verror(td, errno, "ioprio_set");
1080 if (td->o.cgroup_weight && cgroup_setup(td, cgroup_list))
1083 if (nice(td->o.nice) == -1) {
1084 td_verror(td, errno, "nice");
1088 if (td->o.ioscheduler && switch_ioscheduler(td))
1091 if (!td->o.create_serialize && setup_files(td))
1097 if (init_random_map(td))
1100 if (td->o.exec_prerun) {
1101 if (exec_string(td->o.exec_prerun))
1105 if (td->o.pre_read) {
1106 if (pre_read_files(td) < 0)
1110 fio_gettime(&td->epoch, NULL);
1111 getrusage(RUSAGE_SELF, &td->ts.ru_start);
1113 runtime[0] = runtime[1] = 0;
1115 while (keep_running(td)) {
1116 fio_gettime(&td->start, NULL);
1117 memcpy(&td->ts.stat_sample_time[0], &td->start,
1119 memcpy(&td->ts.stat_sample_time[1], &td->start,
1121 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1123 if (td->o.ratemin[0] || td->o.ratemin[1])
1124 memcpy(&td->lastrate, &td->ts.stat_sample_time,
1125 sizeof(td->lastrate));
1130 prune_io_piece_log(td);
1136 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1137 elapsed = utime_since_now(&td->start);
1138 runtime[DDIR_READ] += elapsed;
1140 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1141 elapsed = utime_since_now(&td->start);
1142 runtime[DDIR_WRITE] += elapsed;
1145 if (td->error || td->terminate)
1148 if (!td->o.do_verify ||
1149 td->o.verify == VERIFY_NONE ||
1150 (td->io_ops->flags & FIO_UNIDIR))
1155 fio_gettime(&td->start, NULL);
1159 runtime[DDIR_READ] += utime_since_now(&td->start);
1161 if (td->error || td->terminate)
1165 update_rusage_stat(td);
1166 td->ts.runtime[0] = (runtime[0] + 999) / 1000;
1167 td->ts.runtime[1] = (runtime[1] + 999) / 1000;
1168 td->ts.total_run_time = mtime_since_now(&td->epoch);
1169 td->ts.io_bytes[0] = td->io_bytes[0];
1170 td->ts.io_bytes[1] = td->io_bytes[1];
1172 fio_mutex_down(writeout_mutex);
1173 if (td->ts.bw_log) {
1174 if (td->o.bw_log_file) {
1175 finish_log_named(td, td->ts.bw_log,
1176 td->o.bw_log_file, "bw");
1178 finish_log(td, td->ts.bw_log, "bw");
1180 if (td->ts.slat_log) {
1181 if (td->o.lat_log_file) {
1182 finish_log_named(td, td->ts.slat_log,
1183 td->o.lat_log_file, "slat");
1185 finish_log(td, td->ts.slat_log, "slat");
1187 if (td->ts.clat_log) {
1188 if (td->o.lat_log_file) {
1189 finish_log_named(td, td->ts.clat_log,
1190 td->o.lat_log_file, "clat");
1192 finish_log(td, td->ts.clat_log, "clat");
1194 fio_mutex_up(writeout_mutex);
1195 if (td->o.exec_postrun)
1196 exec_string(td->o.exec_postrun);
1198 if (exitall_on_terminate)
1199 terminate_threads(td->groupid);
1203 printf("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1206 if (td->o.verify_async)
1207 verify_async_exit(td);
1209 close_and_free_files(td);
1212 cgroup_shutdown(td);
1214 if (td->o.cpumask_set) {
1215 int ret = fio_cpuset_exit(&td->o.cpumask);
1217 td_verror(td, ret, "fio_cpuset_exit");
1221 * do this very late, it will log file closing as well
1223 if (td->o.write_iolog_file)
1224 write_iolog_close(td);
1226 options_mem_free(td);
1227 td_set_runstate(td, TD_EXITED);
1228 return (void *) (unsigned long) td->error;
1232 * We cannot pass the td data into a forked process, so attach the td and
1233 * pass it to the thread worker.
1235 static int fork_main(int shmid, int offset)
1237 struct thread_data *td;
1240 data = shmat(shmid, NULL, 0);
1241 if (data == (void *) -1) {
1248 td = data + offset * sizeof(struct thread_data);
1249 ret = thread_main(td);
1251 return (int) (unsigned long) ret;
1255 * Run over the job map and reap the threads that have exited, if any.
1257 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1259 struct thread_data *td;
1260 int i, cputhreads, realthreads, pending, status, ret;
1263 * reap exited threads (TD_EXITED -> TD_REAPED)
1265 realthreads = pending = cputhreads = 0;
1266 for_each_td(td, i) {
1270 * ->io_ops is NULL for a thread that has closed its
1273 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1282 if (td->runstate == TD_REAPED)
1284 if (td->o.use_thread) {
1285 if (td->runstate == TD_EXITED) {
1286 td_set_runstate(td, TD_REAPED);
1293 if (td->runstate == TD_EXITED)
1297 * check if someone quit or got killed in an unusual way
1299 ret = waitpid(td->pid, &status, flags);
1301 if (errno == ECHILD) {
1302 log_err("fio: pid=%d disappeared %d\n",
1303 (int) td->pid, td->runstate);
1304 td_set_runstate(td, TD_REAPED);
1308 } else if (ret == td->pid) {
1309 if (WIFSIGNALED(status)) {
1310 int sig = WTERMSIG(status);
1313 log_err("fio: pid=%d, got signal=%d\n",
1314 (int) td->pid, sig);
1315 td_set_runstate(td, TD_REAPED);
1318 if (WIFEXITED(status)) {
1319 if (WEXITSTATUS(status) && !td->error)
1320 td->error = WEXITSTATUS(status);
1322 td_set_runstate(td, TD_REAPED);
1328 * thread is not dead, continue
1334 (*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
1335 (*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
1342 done_secs += mtime_since_now(&td->epoch) / 1000;
1345 if (*nr_running == cputhreads && !pending && realthreads)
1346 terminate_threads(TERMINATE_ALL);
1349 static void *gtod_thread_main(void *data)
1351 fio_mutex_up(startup_mutex);
1354 * As long as we have jobs around, update the clock. It would be nice
1355 * to have some way of NOT hammering that CPU with gettimeofday(),
1356 * but I'm not sure what to use outside of a simple CPU nop to relax
1357 * it - we don't want to lose precision.
1367 static int fio_start_gtod_thread(void)
1371 ret = pthread_create(>od_thread, NULL, gtod_thread_main, NULL);
1373 log_err("Can't create gtod thread: %s\n", strerror(ret));
1377 ret = pthread_detach(gtod_thread);
1379 log_err("Can't detatch gtod thread: %s\n", strerror(ret));
1383 dprint(FD_MUTEX, "wait on startup_mutex\n");
1384 fio_mutex_down(startup_mutex);
1385 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1390 * Main function for kicking off and reaping jobs, as needed.
1392 static void run_threads(void)
1394 struct thread_data *td;
1395 unsigned long spent;
1396 int i, todo, nr_running, m_rate, t_rate, nr_started;
1398 if (fio_pin_memory())
1401 if (fio_gtod_offload && fio_start_gtod_thread())
1404 if (!terse_output) {
1405 printf("Starting ");
1407 printf("%d thread%s", nr_thread,
1408 nr_thread > 1 ? "s" : "");
1412 printf("%d process%s", nr_process,
1413 nr_process > 1 ? "es" : "");
1421 todo = thread_number;
1424 m_rate = t_rate = 0;
1426 for_each_td(td, i) {
1427 print_status_init(td->thread_number - 1);
1429 if (!td->o.create_serialize) {
1435 * do file setup here so it happens sequentially,
1436 * we don't want X number of threads getting their
1437 * client data interspersed on disk
1439 if (setup_files(td)) {
1442 log_err("fio: pid=%d, err=%d/%s\n",
1443 (int) td->pid, td->error, td->verror);
1444 td_set_runstate(td, TD_REAPED);
1451 * for sharing to work, each job must always open
1452 * its own files. so close them, if we opened them
1455 for_each_file(td, f, i) {
1456 if (fio_file_open(f))
1457 td_io_close_file(td, f);
1467 struct thread_data *map[MAX_JOBS];
1468 struct timeval this_start;
1469 int this_jobs = 0, left;
1472 * create threads (TD_NOT_CREATED -> TD_CREATED)
1474 for_each_td(td, i) {
1475 if (td->runstate != TD_NOT_CREATED)
1479 * never got a chance to start, killed by other
1480 * thread for some reason
1482 if (td->terminate) {
1487 if (td->o.start_delay) {
1488 spent = mtime_since_genesis();
1490 if (td->o.start_delay * 1000 > spent)
1494 if (td->o.stonewall && (nr_started || nr_running)) {
1495 dprint(FD_PROCESS, "%s: stonewall wait\n",
1501 * Set state to created. Thread will transition
1502 * to TD_INITIALIZED when it's done setting up.
1504 td_set_runstate(td, TD_CREATED);
1505 map[this_jobs++] = td;
1508 if (td->o.use_thread) {
1511 dprint(FD_PROCESS, "will pthread_create\n");
1512 ret = pthread_create(&td->thread, NULL,
1515 log_err("pthread_create: %s\n",
1520 ret = pthread_detach(td->thread);
1522 log_err("pthread_detach: %s",
1526 dprint(FD_PROCESS, "will fork\n");
1529 int ret = fork_main(shm_id, i);
1532 } else if (i == fio_debug_jobno)
1533 *fio_debug_jobp = pid;
1535 dprint(FD_MUTEX, "wait on startup_mutex\n");
1536 if (fio_mutex_down_timeout(startup_mutex, 10)) {
1537 log_err("fio: job startup hung? exiting.\n");
1538 terminate_threads(TERMINATE_ALL);
1543 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1547 * Wait for the started threads to transition to
1550 fio_gettime(&this_start, NULL);
1552 while (left && !fio_abort) {
1553 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1558 for (i = 0; i < this_jobs; i++) {
1562 if (td->runstate == TD_INITIALIZED) {
1565 } else if (td->runstate >= TD_EXITED) {
1569 nr_running++; /* work-around... */
1575 log_err("fio: %d jobs failed to start\n", left);
1576 for (i = 0; i < this_jobs; i++) {
1580 kill(td->pid, SIGTERM);
1586 * start created threads (TD_INITIALIZED -> TD_RUNNING).
1588 for_each_td(td, i) {
1589 if (td->runstate != TD_INITIALIZED)
1592 if (in_ramp_time(td))
1593 td_set_runstate(td, TD_RAMP);
1595 td_set_runstate(td, TD_RUNNING);
1598 m_rate += td->o.ratemin[0] + td->o.ratemin[1];
1599 t_rate += td->o.rate[0] + td->o.rate[1];
1601 fio_mutex_up(td->mutex);
1604 reap_threads(&nr_running, &t_rate, &m_rate);
1610 while (nr_running) {
1611 reap_threads(&nr_running, &t_rate, &m_rate);
1619 int main(int argc, char *argv[])
1626 * We need locale for number printing, if it isn't set then just
1627 * go with the US format.
1629 if (!getenv("LC_NUMERIC"))
1630 setlocale(LC_NUMERIC, "en_US");
1632 ps = sysconf(_SC_PAGESIZE);
1634 log_err("Failed to get page size\n");
1641 fio_keywords_init();
1643 if (parse_options(argc, argv))
1650 setup_log(&agg_io_log[DDIR_READ]);
1651 setup_log(&agg_io_log[DDIR_WRITE]);
1654 startup_mutex = fio_mutex_init(0);
1655 writeout_mutex = fio_mutex_init(1);
1661 cgroup_list = smalloc(sizeof(*cgroup_list));
1662 INIT_FLIST_HEAD(cgroup_list);
1669 __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1670 __finish_log(agg_io_log[DDIR_WRITE],
1671 "agg-write_bw.log");
1675 cgroup_kill(cgroup_list);
1678 fio_mutex_remove(startup_mutex);
1679 fio_mutex_remove(writeout_mutex);