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
46 unsigned long page_mask;
47 unsigned long page_size;
49 #define PAGE_ALIGN(buf) \
50 (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
53 int thread_number = 0;
58 unsigned long done_secs = 0;
60 static struct fio_mutex *startup_mutex;
61 static struct fio_mutex *writeout_mutex;
62 static volatile int fio_abort;
63 static int exit_value;
64 static struct itimerval itimer;
65 static pthread_t gtod_thread;
66 static struct flist_head *cgroup_list;
67 static char *cgroup_mnt;
69 struct io_log *agg_io_log[2];
71 #define TERMINATE_ALL (-1)
72 #define JOB_START_TIMEOUT (5 * 1000)
74 void td_set_runstate(struct thread_data *td, int runstate)
76 if (td->runstate == runstate)
79 dprint(FD_PROCESS, "pid=%d: runstate %d -> %d\n", (int) td->pid,
80 td->runstate, runstate);
81 td->runstate = runstate;
84 static void terminate_threads(int group_id)
86 struct thread_data *td;
89 dprint(FD_PROCESS, "terminate group_id=%d\n", group_id);
92 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
93 dprint(FD_PROCESS, "setting terminate on %s/%d\n",
94 td->o.name, (int) td->pid);
96 td->o.start_delay = 0;
99 * if the thread is running, just let it exit
101 if (td->runstate < TD_RUNNING)
102 kill(td->pid, SIGQUIT);
104 struct ioengine_ops *ops = td->io_ops;
106 if (ops && (ops->flags & FIO_SIGQUIT))
107 kill(td->pid, SIGQUIT);
113 static void status_timer_arm(void)
115 itimer.it_value.tv_sec = 0;
116 itimer.it_value.tv_usec = DISK_UTIL_MSEC * 1000;
117 setitimer(ITIMER_REAL, &itimer, NULL);
120 static void sig_alrm(int fio_unused sig)
124 print_thread_status();
130 * Happens on thread runs with ctrl-c, ignore our own SIGQUIT
132 static void sig_quit(int sig)
136 static void sig_int(int sig)
139 log_info("\nfio: terminating on signal %d\n", sig);
141 terminate_threads(TERMINATE_ALL);
145 static void sig_ill(int fio_unused sig)
150 log_err("fio: illegal instruction. your cpu does not support "
151 "the sse4.2 instruction for crc32c\n");
152 terminate_threads(TERMINATE_ALL);
156 static void set_sig_handlers(void)
158 struct sigaction act;
160 memset(&act, 0, sizeof(act));
161 act.sa_handler = sig_alrm;
162 act.sa_flags = SA_RESTART;
163 sigaction(SIGALRM, &act, NULL);
165 memset(&act, 0, sizeof(act));
166 act.sa_handler = sig_int;
167 act.sa_flags = SA_RESTART;
168 sigaction(SIGINT, &act, NULL);
170 memset(&act, 0, sizeof(act));
171 act.sa_handler = sig_ill;
172 act.sa_flags = SA_RESTART;
173 sigaction(SIGILL, &act, NULL);
175 memset(&act, 0, sizeof(act));
176 act.sa_handler = sig_quit;
177 act.sa_flags = SA_RESTART;
178 sigaction(SIGQUIT, &act, NULL);
182 * Check if we are above the minimum rate given.
184 static int __check_min_rate(struct thread_data *td, struct timeval *now,
187 unsigned long long bytes = 0;
188 unsigned long iops = 0;
191 unsigned int ratemin = 0;
192 unsigned int rate_iops = 0;
193 unsigned int rate_iops_min = 0;
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 if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
378 fio_gettime(&td->tv_cache, NULL);
381 static int break_on_this_error(struct thread_data *td, int *retptr)
385 if (ret < 0 || td->error) {
388 if (!td->o.continue_on_error)
396 if (td_non_fatal_error(err)) {
398 * Continue with the I/Os in case of
401 update_error_count(td, err);
405 } else if (td->o.fill_device && err == ENOSPC) {
407 * We expect to hit this error if
408 * fill_device option is set.
415 * Stop the I/O in case of a fatal
418 update_error_count(td, err);
427 * The main verify engine. Runs over the writes we previously submitted,
428 * reads the blocks back in, and checks the crc/md5 of the data.
430 static void do_verify(struct thread_data *td)
437 dprint(FD_VERIFY, "starting loop\n");
440 * sync io first and invalidate cache, to make sure we really
443 for_each_file(td, f, i) {
444 if (!fio_file_open(f))
446 if (fio_io_sync(td, f))
448 if (file_invalidate_cache(td, f))
455 td_set_runstate(td, TD_VERIFYING);
458 while (!td->terminate) {
463 if (runtime_exceeded(td, &td->tv_cache)) {
468 io_u = __get_io_u(td);
472 if (get_next_verify(td, io_u)) {
477 if (td_io_prep(td, io_u)) {
482 if (td->o.verify_async)
483 io_u->end_io = verify_io_u_async;
485 io_u->end_io = verify_io_u;
487 ret = td_io_queue(td, io_u);
489 case FIO_Q_COMPLETED:
492 clear_io_u(td, io_u);
493 } else if (io_u->resid) {
494 int bytes = io_u->xfer_buflen - io_u->resid;
495 struct fio_file *f = io_u->file;
501 td_verror(td, EIO, "full resid");
506 io_u->xfer_buflen = io_u->resid;
507 io_u->xfer_buf += bytes;
508 io_u->offset += bytes;
510 td->ts.short_io_u[io_u->ddir]++;
512 if (io_u->offset == f->real_file_size)
515 requeue_io_u(td, &io_u);
518 ret = io_u_sync_complete(td, io_u, NULL);
526 requeue_io_u(td, &io_u);
527 ret2 = td_io_commit(td);
533 td_verror(td, -ret, "td_io_queue");
537 if (break_on_this_error(td, &ret))
541 * if we can queue more, do so. but check if there are
542 * completed io_u's first.
544 full = queue_full(td) || ret == FIO_Q_BUSY;
545 if (full || !td->o.iodepth_batch_complete) {
546 min_events = min(td->o.iodepth_batch_complete,
548 if (full && !min_events)
553 * Reap required number of io units, if any,
554 * and do the verification on them through
555 * the callback handler
557 if (io_u_queued_complete(td, min_events, NULL) < 0) {
561 } while (full && (td->cur_depth > td->o.iodepth_low));
568 min_events = td->cur_depth;
571 ret = io_u_queued_complete(td, min_events, NULL);
573 cleanup_pending_aio(td);
575 td_set_runstate(td, TD_RUNNING);
577 dprint(FD_VERIFY, "exiting loop\n");
581 * Main IO worker function. It retrieves io_u's to process and queues
582 * and reaps them, checking for rate and errors along the way.
584 static void do_io(struct thread_data *td)
589 if (in_ramp_time(td))
590 td_set_runstate(td, TD_RAMP);
592 td_set_runstate(td, TD_RUNNING);
594 while ( (td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
595 ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) ) {
596 struct timeval comp_time;
597 unsigned long bytes_done[2] = { 0, 0 };
607 if (runtime_exceeded(td, &td->tv_cache)) {
617 * Add verification end_io handler, if asked to verify
618 * a previously written file.
620 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
622 if (td->o.verify_async)
623 io_u->end_io = verify_io_u_async;
625 io_u->end_io = verify_io_u;
626 td_set_runstate(td, TD_VERIFYING);
627 } else if (in_ramp_time(td))
628 td_set_runstate(td, TD_RAMP);
630 td_set_runstate(td, TD_RUNNING);
632 ret = td_io_queue(td, io_u);
634 case FIO_Q_COMPLETED:
637 clear_io_u(td, io_u);
638 } else if (io_u->resid) {
639 int bytes = io_u->xfer_buflen - io_u->resid;
640 struct fio_file *f = io_u->file;
646 td_verror(td, EIO, "full resid");
651 io_u->xfer_buflen = io_u->resid;
652 io_u->xfer_buf += bytes;
653 io_u->offset += bytes;
655 td->ts.short_io_u[io_u->ddir]++;
657 if (io_u->offset == f->real_file_size)
660 requeue_io_u(td, &io_u);
663 if (__should_check_rate(td, 0) ||
664 __should_check_rate(td, 1))
665 fio_gettime(&comp_time, NULL);
667 ret = io_u_sync_complete(td, io_u, bytes_done);
674 * if the engine doesn't have a commit hook,
675 * the io_u is really queued. if it does have such
676 * a hook, it has to call io_u_queued() itself.
678 if (td->io_ops->commit == NULL)
679 io_u_queued(td, io_u);
682 requeue_io_u(td, &io_u);
683 ret2 = td_io_commit(td);
693 if (break_on_this_error(td, &ret))
697 * See if we need to complete some commands
699 full = queue_full(td) || ret == FIO_Q_BUSY;
700 if (full || !td->o.iodepth_batch_complete) {
701 min_evts = min(td->o.iodepth_batch_complete,
703 if (full && !min_evts)
706 if (__should_check_rate(td, 0) ||
707 __should_check_rate(td, 1))
708 fio_gettime(&comp_time, NULL);
711 ret = io_u_queued_complete(td, min_evts, bytes_done);
715 } while (full && (td->cur_depth > td->o.iodepth_low));
720 if (!(bytes_done[0] + bytes_done[1]))
723 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
724 if (check_min_rate(td, &comp_time, bytes_done)) {
725 if (exitall_on_terminate)
726 terminate_threads(td->groupid);
727 td_verror(td, EIO, "check_min_rate");
732 if (td->o.thinktime) {
733 unsigned long long b;
735 b = td->io_blocks[0] + td->io_blocks[1];
736 if (!(b % td->o.thinktime_blocks)) {
739 if (td->o.thinktime_spin)
740 usec_spin(td->o.thinktime_spin);
742 left = td->o.thinktime - td->o.thinktime_spin;
744 usec_sleep(td, left);
749 if (td->o.fill_device && td->error == ENOSPC) {
758 ret = io_u_queued_complete(td, i, NULL);
760 if (should_fsync(td) && td->o.end_fsync) {
761 td_set_runstate(td, TD_FSYNCING);
763 for_each_file(td, f, i) {
764 if (!fio_file_open(f))
770 cleanup_pending_aio(td);
773 * stop job if we failed doing any IO
775 if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
779 static void cleanup_io_u(struct thread_data *td)
781 struct flist_head *entry, *n;
784 flist_for_each_safe(entry, n, &td->io_u_freelist) {
785 io_u = flist_entry(entry, struct io_u, list);
787 flist_del(&io_u->list);
794 static int init_io_u(struct thread_data *td)
798 int cl_align, i, max_units;
801 max_units = td->o.iodepth;
802 max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
803 td->orig_buffer_size = (unsigned long long) max_bs
804 * (unsigned long long) max_units;
806 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
809 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
810 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
813 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
814 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
818 if (allocate_io_mem(td))
821 if (td->o.odirect || td->o.mem_align)
822 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
826 cl_align = os_cache_line_size();
828 for (i = 0; i < max_units; i++) {
834 if (posix_memalign(&ptr, cl_align, sizeof(*io_u))) {
835 log_err("fio: posix_memalign=%s\n", strerror(errno));
840 memset(io_u, 0, sizeof(*io_u));
841 INIT_FLIST_HEAD(&io_u->list);
842 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
844 if (!(td->io_ops->flags & FIO_NOIO)) {
845 io_u->buf = p + max_bs * i;
846 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
848 if (td_write(td) && !td->o.refill_buffers)
849 io_u_fill_buffer(td, io_u, max_bs);
853 io_u->flags = IO_U_F_FREE;
854 flist_add(&io_u->list, &td->io_u_freelist);
860 static int switch_ioscheduler(struct thread_data *td)
862 char tmp[256], tmp2[128];
866 if (td->io_ops->flags & FIO_DISKLESSIO)
869 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
871 f = fopen(tmp, "r+");
873 if (errno == ENOENT) {
874 log_err("fio: os or kernel doesn't support IO scheduler"
878 td_verror(td, errno, "fopen iosched");
885 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
886 if (ferror(f) || ret != 1) {
887 td_verror(td, errno, "fwrite");
895 * Read back and check that the selected scheduler is now the default.
897 ret = fread(tmp, 1, sizeof(tmp), f);
898 if (ferror(f) || ret < 0) {
899 td_verror(td, errno, "fread");
904 sprintf(tmp2, "[%s]", td->o.ioscheduler);
905 if (!strstr(tmp, tmp2)) {
906 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
907 td_verror(td, EINVAL, "iosched_switch");
916 static int keep_running(struct thread_data *td)
918 unsigned long long io_done;
922 if (td->o.time_based)
929 io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
931 if (io_done < td->o.size)
937 static void reset_io_counters(struct thread_data *td)
939 td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
940 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
942 td->rate_bytes[0] = td->rate_bytes[1] = 0;
943 td->rate_blocks[0] = td->rate_blocks[1] = 0;
945 td->last_was_sync = 0;
948 * reset file done count if we are to start over
950 if (td->o.time_based || td->o.loops)
951 td->nr_done_files = 0;
954 * Set the same seed to get repeatable runs
956 td_fill_rand_seeds(td);
959 void reset_all_stats(struct thread_data *td)
964 reset_io_counters(td);
966 for (i = 0; i < 2; i++) {
968 td->io_blocks[i] = 0;
969 td->io_issues[i] = 0;
970 td->ts.total_io_u[i] = 0;
973 fio_gettime(&tv, NULL);
974 memcpy(&td->epoch, &tv, sizeof(tv));
975 memcpy(&td->start, &tv, sizeof(tv));
978 static void clear_io_state(struct thread_data *td)
983 reset_io_counters(td);
986 for_each_file(td, f, i)
987 fio_file_clear_done(f);
990 static int exec_string(const char *string)
992 int ret, newlen = strlen(string) + 1 + 8;
995 str = malloc(newlen);
996 sprintf(str, "sh -c %s", string);
1000 log_err("fio: exec of cmd <%s> failed\n", str);
1007 * Entry point for the thread based jobs. The process based jobs end up
1008 * here as well, after a little setup.
1010 static void *thread_main(void *data)
1012 unsigned long long runtime[2], elapsed;
1013 struct thread_data *td = data;
1014 pthread_condattr_t attr;
1017 if (!td->o.use_thread)
1022 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1024 INIT_FLIST_HEAD(&td->io_u_freelist);
1025 INIT_FLIST_HEAD(&td->io_u_busylist);
1026 INIT_FLIST_HEAD(&td->io_u_requeues);
1027 INIT_FLIST_HEAD(&td->io_log_list);
1028 INIT_FLIST_HEAD(&td->io_hist_list);
1029 INIT_FLIST_HEAD(&td->verify_list);
1030 pthread_mutex_init(&td->io_u_lock, NULL);
1031 td->io_hist_tree = RB_ROOT;
1033 pthread_condattr_init(&attr);
1034 pthread_cond_init(&td->verify_cond, &attr);
1035 pthread_cond_init(&td->free_cond, &attr);
1037 td_set_runstate(td, TD_INITIALIZED);
1038 dprint(FD_MUTEX, "up startup_mutex\n");
1039 fio_mutex_up(startup_mutex);
1040 dprint(FD_MUTEX, "wait on td->mutex\n");
1041 fio_mutex_down(td->mutex);
1042 dprint(FD_MUTEX, "done waiting on td->mutex\n");
1045 * the ->mutex mutex is now no longer used, close it to avoid
1046 * eating a file descriptor
1048 fio_mutex_remove(td->mutex);
1050 if (td->o.uid != -1U && setuid(td->o.uid)) {
1051 td_verror(td, errno, "setuid");
1054 if (td->o.gid != -1U && setgid(td->o.gid)) {
1055 td_verror(td, errno, "setgid");
1060 * May alter parameters that init_io_u() will use, so we need to
1069 if (td->o.verify_async && verify_async_init(td))
1072 if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1073 td_verror(td, errno, "cpu_set_affinity");
1078 * If we have a gettimeofday() thread, make sure we exclude that
1079 * thread from this job
1081 if (td->o.gtod_cpu) {
1082 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1083 if (fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1084 td_verror(td, errno, "cpu_set_affinity");
1089 if (td->ioprio_set) {
1090 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1091 td_verror(td, errno, "ioprio_set");
1096 if (td->o.cgroup_weight && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1099 if (nice(td->o.nice) == -1) {
1100 td_verror(td, errno, "nice");
1104 if (td->o.ioscheduler && switch_ioscheduler(td))
1107 if (!td->o.create_serialize && setup_files(td))
1113 if (init_random_map(td))
1116 if (td->o.exec_prerun) {
1117 if (exec_string(td->o.exec_prerun))
1121 if (td->o.pre_read) {
1122 if (pre_read_files(td) < 0)
1126 fio_gettime(&td->epoch, NULL);
1127 getrusage(RUSAGE_SELF, &td->ts.ru_start);
1129 runtime[0] = runtime[1] = 0;
1131 while (keep_running(td)) {
1132 fio_gettime(&td->start, NULL);
1133 memcpy(&td->ts.stat_sample_time[0], &td->start,
1135 memcpy(&td->ts.stat_sample_time[1], &td->start,
1137 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1139 if (td->o.ratemin[0] || td->o.ratemin[1])
1140 memcpy(&td->lastrate, &td->ts.stat_sample_time,
1141 sizeof(td->lastrate));
1146 prune_io_piece_log(td);
1152 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1153 elapsed = utime_since_now(&td->start);
1154 runtime[DDIR_READ] += elapsed;
1156 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1157 elapsed = utime_since_now(&td->start);
1158 runtime[DDIR_WRITE] += elapsed;
1161 if (td->error || td->terminate)
1164 if (!td->o.do_verify ||
1165 td->o.verify == VERIFY_NONE ||
1166 (td->io_ops->flags & FIO_UNIDIR))
1171 fio_gettime(&td->start, NULL);
1175 runtime[DDIR_READ] += utime_since_now(&td->start);
1177 if (td->error || td->terminate)
1181 update_rusage_stat(td);
1182 td->ts.runtime[0] = (runtime[0] + 999) / 1000;
1183 td->ts.runtime[1] = (runtime[1] + 999) / 1000;
1184 td->ts.total_run_time = mtime_since_now(&td->epoch);
1185 td->ts.io_bytes[0] = td->io_bytes[0];
1186 td->ts.io_bytes[1] = td->io_bytes[1];
1188 fio_mutex_down(writeout_mutex);
1189 if (td->ts.bw_log) {
1190 if (td->o.bw_log_file) {
1191 finish_log_named(td, td->ts.bw_log,
1192 td->o.bw_log_file, "bw");
1194 finish_log(td, td->ts.bw_log, "bw");
1196 if (td->ts.slat_log) {
1197 if (td->o.lat_log_file) {
1198 finish_log_named(td, td->ts.slat_log,
1199 td->o.lat_log_file, "slat");
1201 finish_log(td, td->ts.slat_log, "slat");
1203 if (td->ts.clat_log) {
1204 if (td->o.lat_log_file) {
1205 finish_log_named(td, td->ts.clat_log,
1206 td->o.lat_log_file, "clat");
1208 finish_log(td, td->ts.clat_log, "clat");
1210 fio_mutex_up(writeout_mutex);
1211 if (td->o.exec_postrun)
1212 exec_string(td->o.exec_postrun);
1214 if (exitall_on_terminate)
1215 terminate_threads(td->groupid);
1219 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1222 if (td->o.verify_async)
1223 verify_async_exit(td);
1225 close_and_free_files(td);
1228 cgroup_shutdown(td, &cgroup_mnt);
1230 if (td->o.cpumask_set) {
1231 int ret = fio_cpuset_exit(&td->o.cpumask);
1233 td_verror(td, ret, "fio_cpuset_exit");
1237 * do this very late, it will log file closing as well
1239 if (td->o.write_iolog_file)
1240 write_iolog_close(td);
1242 options_mem_free(td);
1243 td_set_runstate(td, TD_EXITED);
1244 return (void *) (unsigned long) td->error;
1248 * We cannot pass the td data into a forked process, so attach the td and
1249 * pass it to the thread worker.
1251 static int fork_main(int shmid, int offset)
1253 struct thread_data *td;
1256 data = shmat(shmid, NULL, 0);
1257 if (data == (void *) -1) {
1264 td = data + offset * sizeof(struct thread_data);
1265 ret = thread_main(td);
1267 return (int) (unsigned long) ret;
1271 * Run over the job map and reap the threads that have exited, if any.
1273 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1275 struct thread_data *td;
1276 int i, cputhreads, realthreads, pending, status, ret;
1279 * reap exited threads (TD_EXITED -> TD_REAPED)
1281 realthreads = pending = cputhreads = 0;
1282 for_each_td(td, i) {
1286 * ->io_ops is NULL for a thread that has closed its
1289 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1298 if (td->runstate == TD_REAPED)
1300 if (td->o.use_thread) {
1301 if (td->runstate == TD_EXITED) {
1302 td_set_runstate(td, TD_REAPED);
1309 if (td->runstate == TD_EXITED)
1313 * check if someone quit or got killed in an unusual way
1315 ret = waitpid(td->pid, &status, flags);
1317 if (errno == ECHILD) {
1318 log_err("fio: pid=%d disappeared %d\n",
1319 (int) td->pid, td->runstate);
1320 td_set_runstate(td, TD_REAPED);
1324 } else if (ret == td->pid) {
1325 if (WIFSIGNALED(status)) {
1326 int sig = WTERMSIG(status);
1329 log_err("fio: pid=%d, got signal=%d\n",
1330 (int) td->pid, sig);
1331 td_set_runstate(td, TD_REAPED);
1334 if (WIFEXITED(status)) {
1335 if (WEXITSTATUS(status) && !td->error)
1336 td->error = WEXITSTATUS(status);
1338 td_set_runstate(td, TD_REAPED);
1344 * thread is not dead, continue
1350 (*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
1351 (*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
1358 done_secs += mtime_since_now(&td->epoch) / 1000;
1361 if (*nr_running == cputhreads && !pending && realthreads)
1362 terminate_threads(TERMINATE_ALL);
1365 static void *gtod_thread_main(void *data)
1367 fio_mutex_up(startup_mutex);
1370 * As long as we have jobs around, update the clock. It would be nice
1371 * to have some way of NOT hammering that CPU with gettimeofday(),
1372 * but I'm not sure what to use outside of a simple CPU nop to relax
1373 * it - we don't want to lose precision.
1383 static int fio_start_gtod_thread(void)
1387 ret = pthread_create(>od_thread, NULL, gtod_thread_main, NULL);
1389 log_err("Can't create gtod thread: %s\n", strerror(ret));
1393 ret = pthread_detach(gtod_thread);
1395 log_err("Can't detatch gtod thread: %s\n", strerror(ret));
1399 dprint(FD_MUTEX, "wait on startup_mutex\n");
1400 fio_mutex_down(startup_mutex);
1401 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1406 * Main function for kicking off and reaping jobs, as needed.
1408 static void run_threads(void)
1410 struct thread_data *td;
1411 unsigned long spent;
1412 int i, todo, nr_running, m_rate, t_rate, nr_started;
1414 if (fio_pin_memory())
1417 if (fio_gtod_offload && fio_start_gtod_thread())
1420 if (!terse_output) {
1421 log_info("Starting ");
1423 log_info("%d thread%s", nr_thread,
1424 nr_thread > 1 ? "s" : "");
1428 log_info("%d process%s", nr_process,
1429 nr_process > 1 ? "es" : "");
1437 todo = thread_number;
1440 m_rate = t_rate = 0;
1442 for_each_td(td, i) {
1443 print_status_init(td->thread_number - 1);
1445 if (!td->o.create_serialize) {
1451 * do file setup here so it happens sequentially,
1452 * we don't want X number of threads getting their
1453 * client data interspersed on disk
1455 if (setup_files(td)) {
1458 log_err("fio: pid=%d, err=%d/%s\n",
1459 (int) td->pid, td->error, td->verror);
1460 td_set_runstate(td, TD_REAPED);
1467 * for sharing to work, each job must always open
1468 * its own files. so close them, if we opened them
1471 for_each_file(td, f, i) {
1472 if (fio_file_open(f))
1473 td_io_close_file(td, f);
1483 struct thread_data *map[MAX_JOBS];
1484 struct timeval this_start;
1485 int this_jobs = 0, left;
1488 * create threads (TD_NOT_CREATED -> TD_CREATED)
1490 for_each_td(td, i) {
1491 if (td->runstate != TD_NOT_CREATED)
1495 * never got a chance to start, killed by other
1496 * thread for some reason
1498 if (td->terminate) {
1503 if (td->o.start_delay) {
1504 spent = mtime_since_genesis();
1506 if (td->o.start_delay * 1000 > spent)
1510 if (td->o.stonewall && (nr_started || nr_running)) {
1511 dprint(FD_PROCESS, "%s: stonewall wait\n",
1517 * Set state to created. Thread will transition
1518 * to TD_INITIALIZED when it's done setting up.
1520 td_set_runstate(td, TD_CREATED);
1521 map[this_jobs++] = td;
1524 if (td->o.use_thread) {
1527 dprint(FD_PROCESS, "will pthread_create\n");
1528 ret = pthread_create(&td->thread, NULL,
1531 log_err("pthread_create: %s\n",
1536 ret = pthread_detach(td->thread);
1538 log_err("pthread_detach: %s",
1542 dprint(FD_PROCESS, "will fork\n");
1545 int ret = fork_main(shm_id, i);
1548 } else if (i == fio_debug_jobno)
1549 *fio_debug_jobp = pid;
1551 dprint(FD_MUTEX, "wait on startup_mutex\n");
1552 if (fio_mutex_down_timeout(startup_mutex, 10)) {
1553 log_err("fio: job startup hung? exiting.\n");
1554 terminate_threads(TERMINATE_ALL);
1559 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1563 * Wait for the started threads to transition to
1566 fio_gettime(&this_start, NULL);
1568 while (left && !fio_abort) {
1569 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1574 for (i = 0; i < this_jobs; i++) {
1578 if (td->runstate == TD_INITIALIZED) {
1581 } else if (td->runstate >= TD_EXITED) {
1585 nr_running++; /* work-around... */
1591 log_err("fio: %d jobs failed to start\n", left);
1592 for (i = 0; i < this_jobs; i++) {
1596 kill(td->pid, SIGTERM);
1602 * start created threads (TD_INITIALIZED -> TD_RUNNING).
1604 for_each_td(td, i) {
1605 if (td->runstate != TD_INITIALIZED)
1608 if (in_ramp_time(td))
1609 td_set_runstate(td, TD_RAMP);
1611 td_set_runstate(td, TD_RUNNING);
1614 m_rate += td->o.ratemin[0] + td->o.ratemin[1];
1615 t_rate += td->o.rate[0] + td->o.rate[1];
1617 fio_mutex_up(td->mutex);
1620 reap_threads(&nr_running, &t_rate, &m_rate);
1626 while (nr_running) {
1627 reap_threads(&nr_running, &t_rate, &m_rate);
1635 int main(int argc, char *argv[])
1640 init_rand(&__fio_rand_state);
1643 * We need locale for number printing, if it isn't set then just
1644 * go with the US format.
1646 if (!getenv("LC_NUMERIC"))
1647 setlocale(LC_NUMERIC, "en_US");
1649 ps = sysconf(_SC_PAGESIZE);
1651 log_err("Failed to get page size\n");
1658 fio_keywords_init();
1660 if (parse_options(argc, argv))
1663 if (exec_profile && load_profile(exec_profile))
1670 setup_log(&agg_io_log[DDIR_READ]);
1671 setup_log(&agg_io_log[DDIR_WRITE]);
1674 startup_mutex = fio_mutex_init(0);
1675 writeout_mutex = fio_mutex_init(1);
1681 cgroup_list = smalloc(sizeof(*cgroup_list));
1682 INIT_FLIST_HEAD(cgroup_list);
1689 __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1690 __finish_log(agg_io_log[DDIR_WRITE],
1691 "agg-write_bw.log");
1695 cgroup_kill(cgroup_list);
1699 fio_mutex_remove(startup_mutex);
1700 fio_mutex_remove(writeout_mutex);