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
51 unsigned long page_mask;
52 unsigned long page_size;
54 #define PAGE_ALIGN(buf) \
55 (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
58 int thread_number = 0;
63 unsigned long done_secs = 0;
65 static struct fio_mutex *startup_mutex;
66 static struct fio_mutex *writeout_mutex;
67 static volatile int fio_abort;
68 static int exit_value;
69 static pthread_t gtod_thread;
70 static pthread_t disk_util_thread;
71 static struct flist_head *cgroup_list;
72 static char *cgroup_mnt;
74 unsigned long arch_flags = 0;
76 struct io_log *agg_io_log[2];
78 #define JOB_START_TIMEOUT (5 * 1000)
80 void td_set_runstate(struct thread_data *td, int runstate)
82 if (td->runstate == runstate)
85 dprint(FD_PROCESS, "pid=%d: runstate %d -> %d\n", (int) td->pid,
86 td->runstate, runstate);
87 td->runstate = runstate;
90 void fio_terminate_threads(int group_id)
92 struct thread_data *td;
95 dprint(FD_PROCESS, "terminate group_id=%d\n", group_id);
98 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
99 dprint(FD_PROCESS, "setting terminate on %s/%d\n",
100 td->o.name, (int) td->pid);
102 td->o.start_delay = 0;
105 * if the thread is running, just let it exit
109 else if (td->runstate < TD_RAMP)
110 kill(td->pid, SIGTERM);
112 struct ioengine_ops *ops = td->io_ops;
114 if (ops && (ops->flags & FIO_SIGTERM))
115 kill(td->pid, SIGTERM);
121 static void sig_int(int sig)
124 log_info("\nfio: terminating on signal %d\n", sig);
128 fio_terminate_threads(TERMINATE_ALL);
132 static void *disk_thread_main(void *data)
134 fio_mutex_up(startup_mutex);
137 usleep(DISK_UTIL_MSEC * 1000);
143 fio_server_send_status();
145 print_thread_status();
151 static int create_disk_util_thread(void)
155 ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
157 log_err("Can't create disk util thread: %s\n", strerror(ret));
161 ret = pthread_detach(disk_util_thread);
163 log_err("Can't detatch disk util thread: %s\n", strerror(ret));
167 dprint(FD_MUTEX, "wait on startup_mutex\n");
168 fio_mutex_down(startup_mutex);
169 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
173 static void set_sig_handlers(void)
175 struct sigaction act;
177 memset(&act, 0, sizeof(act));
178 act.sa_handler = sig_int;
179 act.sa_flags = SA_RESTART;
180 sigaction(SIGINT, &act, NULL);
182 memset(&act, 0, sizeof(act));
183 act.sa_handler = sig_int;
184 act.sa_flags = SA_RESTART;
185 sigaction(SIGTERM, &act, NULL);
188 memset(&act, 0, sizeof(act));
189 act.sa_handler = sig_int;
190 act.sa_flags = SA_RESTART;
191 sigaction(SIGPIPE, &act, NULL);
196 * Check if we are above the minimum rate given.
198 static int __check_min_rate(struct thread_data *td, struct timeval *now,
201 unsigned long long bytes = 0;
202 unsigned long iops = 0;
205 unsigned int ratemin = 0;
206 unsigned int rate_iops = 0;
207 unsigned int rate_iops_min = 0;
209 assert(ddir_rw(ddir));
211 if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
215 * allow a 2 second settle period in the beginning
217 if (mtime_since(&td->start, now) < 2000)
220 iops += td->io_blocks[ddir];
221 bytes += td->this_io_bytes[ddir];
222 ratemin += td->o.ratemin[ddir];
223 rate_iops += td->o.rate_iops[ddir];
224 rate_iops_min += td->o.rate_iops_min[ddir];
227 * if rate blocks is set, sample is running
229 if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
230 spent = mtime_since(&td->lastrate[ddir], now);
231 if (spent < td->o.ratecycle)
234 if (td->o.rate[ddir]) {
236 * check bandwidth specified rate
238 if (bytes < td->rate_bytes[ddir]) {
239 log_err("%s: min rate %u not met\n", td->o.name,
243 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
244 if (rate < ratemin ||
245 bytes < td->rate_bytes[ddir]) {
246 log_err("%s: min rate %u not met, got"
247 " %luKB/sec\n", td->o.name,
254 * checks iops specified rate
256 if (iops < rate_iops) {
257 log_err("%s: min iops rate %u not met\n",
258 td->o.name, rate_iops);
261 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
262 if (rate < rate_iops_min ||
263 iops < td->rate_blocks[ddir]) {
264 log_err("%s: min iops rate %u not met,"
265 " got %lu\n", td->o.name,
266 rate_iops_min, rate);
272 td->rate_bytes[ddir] = bytes;
273 td->rate_blocks[ddir] = iops;
274 memcpy(&td->lastrate[ddir], now, sizeof(*now));
278 static int check_min_rate(struct thread_data *td, struct timeval *now,
279 unsigned long *bytes_done)
284 ret |= __check_min_rate(td, now, 0);
286 ret |= __check_min_rate(td, now, 1);
291 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
295 if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
302 * When job exits, we can cancel the in-flight IO if we are using async
303 * io. Attempt to do so.
305 static void cleanup_pending_aio(struct thread_data *td)
307 struct flist_head *entry, *n;
312 * get immediately available events, if any
314 r = io_u_queued_complete(td, 0, NULL);
319 * now cancel remaining active events
321 if (td->io_ops->cancel) {
322 flist_for_each_safe(entry, n, &td->io_u_busylist) {
323 io_u = flist_entry(entry, struct io_u, list);
326 * if the io_u isn't in flight, then that generally
327 * means someone leaked an io_u. complain but fix
328 * it up, so we don't stall here.
330 if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
331 log_err("fio: non-busy IO on busy list\n");
334 r = td->io_ops->cancel(td, io_u);
342 r = io_u_queued_complete(td, td->cur_depth, NULL);
346 * Helper to handle the final sync of a file. Works just like the normal
347 * io path, just does everything sync.
349 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
351 struct io_u *io_u = __get_io_u(td);
357 io_u->ddir = DDIR_SYNC;
360 if (td_io_prep(td, io_u)) {
366 ret = td_io_queue(td, io_u);
368 td_verror(td, io_u->error, "td_io_queue");
371 } else if (ret == FIO_Q_QUEUED) {
372 if (io_u_queued_complete(td, 1, NULL) < 0)
374 } else if (ret == FIO_Q_COMPLETED) {
376 td_verror(td, io_u->error, "td_io_queue");
380 if (io_u_sync_complete(td, io_u, NULL) < 0)
382 } else if (ret == FIO_Q_BUSY) {
383 if (td_io_commit(td))
391 static inline void __update_tv_cache(struct thread_data *td)
393 fio_gettime(&td->tv_cache, NULL);
396 static inline void update_tv_cache(struct thread_data *td)
398 if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
399 __update_tv_cache(td);
402 static int break_on_this_error(struct thread_data *td, int *retptr)
406 if (ret < 0 || td->error) {
409 if (!td->o.continue_on_error)
417 if (td_non_fatal_error(err)) {
419 * Continue with the I/Os in case of
422 update_error_count(td, err);
426 } else if (td->o.fill_device && err == ENOSPC) {
428 * We expect to hit this error if
429 * fill_device option is set.
436 * Stop the I/O in case of a fatal
439 update_error_count(td, err);
448 * The main verify engine. Runs over the writes we previously submitted,
449 * reads the blocks back in, and checks the crc/md5 of the data.
451 static void do_verify(struct thread_data *td)
458 dprint(FD_VERIFY, "starting loop\n");
461 * sync io first and invalidate cache, to make sure we really
464 for_each_file(td, f, i) {
465 if (!fio_file_open(f))
467 if (fio_io_sync(td, f))
469 if (file_invalidate_cache(td, f))
476 td_set_runstate(td, TD_VERIFYING);
479 while (!td->terminate) {
484 if (runtime_exceeded(td, &td->tv_cache)) {
485 __update_tv_cache(td);
486 if (runtime_exceeded(td, &td->tv_cache)) {
492 io_u = __get_io_u(td);
496 if (get_next_verify(td, io_u)) {
501 if (td_io_prep(td, io_u)) {
506 if (td->o.verify_async)
507 io_u->end_io = verify_io_u_async;
509 io_u->end_io = verify_io_u;
511 ret = td_io_queue(td, io_u);
513 case FIO_Q_COMPLETED:
516 clear_io_u(td, io_u);
517 } else if (io_u->resid) {
518 int bytes = io_u->xfer_buflen - io_u->resid;
524 td_verror(td, EIO, "full resid");
529 io_u->xfer_buflen = io_u->resid;
530 io_u->xfer_buf += bytes;
531 io_u->offset += bytes;
533 if (ddir_rw(io_u->ddir))
534 td->ts.short_io_u[io_u->ddir]++;
537 if (io_u->offset == f->real_file_size)
540 requeue_io_u(td, &io_u);
543 ret = io_u_sync_complete(td, io_u, NULL);
551 requeue_io_u(td, &io_u);
552 ret2 = td_io_commit(td);
558 td_verror(td, -ret, "td_io_queue");
562 if (break_on_this_error(td, &ret))
566 * if we can queue more, do so. but check if there are
567 * completed io_u's first. Note that we can get BUSY even
568 * without IO queued, if the system is resource starved.
570 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
571 if (full || !td->o.iodepth_batch_complete) {
572 min_events = min(td->o.iodepth_batch_complete,
574 if (full && !min_events && td->o.iodepth_batch_complete != 0)
579 * Reap required number of io units, if any,
580 * and do the verification on them through
581 * the callback handler
583 if (io_u_queued_complete(td, min_events, NULL) < 0) {
587 } while (full && (td->cur_depth > td->o.iodepth_low));
594 min_events = td->cur_depth;
597 ret = io_u_queued_complete(td, min_events, NULL);
599 cleanup_pending_aio(td);
601 td_set_runstate(td, TD_RUNNING);
603 dprint(FD_VERIFY, "exiting loop\n");
607 * Main IO worker function. It retrieves io_u's to process and queues
608 * and reaps them, checking for rate and errors along the way.
610 static void do_io(struct thread_data *td)
615 if (in_ramp_time(td))
616 td_set_runstate(td, TD_RAMP);
618 td_set_runstate(td, TD_RUNNING);
620 while ( (td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
621 (!flist_empty(&td->trim_list)) ||
622 ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) ) {
623 struct timeval comp_time;
624 unsigned long bytes_done[2] = { 0, 0 };
634 if (runtime_exceeded(td, &td->tv_cache)) {
635 __update_tv_cache(td);
636 if (runtime_exceeded(td, &td->tv_cache)) {
647 * Add verification end_io handler, if asked to verify
648 * a previously written file.
650 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
652 if (td->o.verify_async)
653 io_u->end_io = verify_io_u_async;
655 io_u->end_io = verify_io_u;
656 td_set_runstate(td, TD_VERIFYING);
657 } else if (in_ramp_time(td))
658 td_set_runstate(td, TD_RAMP);
660 td_set_runstate(td, TD_RUNNING);
662 ret = td_io_queue(td, io_u);
664 case FIO_Q_COMPLETED:
667 clear_io_u(td, io_u);
668 } else if (io_u->resid) {
669 int bytes = io_u->xfer_buflen - io_u->resid;
670 struct fio_file *f = io_u->file;
676 td_verror(td, EIO, "full resid");
681 io_u->xfer_buflen = io_u->resid;
682 io_u->xfer_buf += bytes;
683 io_u->offset += bytes;
685 if (ddir_rw(io_u->ddir))
686 td->ts.short_io_u[io_u->ddir]++;
688 if (io_u->offset == f->real_file_size)
691 requeue_io_u(td, &io_u);
694 if (__should_check_rate(td, 0) ||
695 __should_check_rate(td, 1))
696 fio_gettime(&comp_time, NULL);
698 ret = io_u_sync_complete(td, io_u, bytes_done);
705 * if the engine doesn't have a commit hook,
706 * the io_u is really queued. if it does have such
707 * a hook, it has to call io_u_queued() itself.
709 if (td->io_ops->commit == NULL)
710 io_u_queued(td, io_u);
713 requeue_io_u(td, &io_u);
714 ret2 = td_io_commit(td);
724 if (break_on_this_error(td, &ret))
728 * See if we need to complete some commands. Note that we
729 * can get BUSY even without IO queued, if the system is
732 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
733 if (full || !td->o.iodepth_batch_complete) {
734 min_evts = min(td->o.iodepth_batch_complete,
736 if (full && !min_evts && td->o.iodepth_batch_complete != 0)
739 if (__should_check_rate(td, 0) ||
740 __should_check_rate(td, 1))
741 fio_gettime(&comp_time, NULL);
744 ret = io_u_queued_complete(td, min_evts, bytes_done);
748 } while (full && (td->cur_depth > td->o.iodepth_low));
753 if (!(bytes_done[0] + bytes_done[1]))
756 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
757 if (check_min_rate(td, &comp_time, bytes_done)) {
758 if (exitall_on_terminate)
759 fio_terminate_threads(td->groupid);
760 td_verror(td, EIO, "check_min_rate");
765 if (td->o.thinktime) {
766 unsigned long long b;
768 b = td->io_blocks[0] + td->io_blocks[1];
769 if (!(b % td->o.thinktime_blocks)) {
772 if (td->o.thinktime_spin)
773 usec_spin(td->o.thinktime_spin);
775 left = td->o.thinktime - td->o.thinktime_spin;
777 usec_sleep(td, left);
782 if (td->trim_entries)
783 printf("trim entries %ld\n", td->trim_entries);
785 if (td->o.fill_device && td->error == ENOSPC) {
794 ret = io_u_queued_complete(td, i, NULL);
795 if (td->o.fill_device && td->error == ENOSPC)
799 if (should_fsync(td) && td->o.end_fsync) {
800 td_set_runstate(td, TD_FSYNCING);
802 for_each_file(td, f, i) {
803 if (!fio_file_open(f))
809 cleanup_pending_aio(td);
812 * stop job if we failed doing any IO
814 if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
818 static void cleanup_io_u(struct thread_data *td)
820 struct flist_head *entry, *n;
823 flist_for_each_safe(entry, n, &td->io_u_freelist) {
824 io_u = flist_entry(entry, struct io_u, list);
826 flist_del(&io_u->list);
827 fio_memfree(io_u, sizeof(*io_u));
833 static int init_io_u(struct thread_data *td)
837 int cl_align, i, max_units;
840 max_units = td->o.iodepth;
841 max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
842 td->orig_buffer_size = (unsigned long long) max_bs
843 * (unsigned long long) max_units;
845 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
848 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
849 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
852 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
853 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
857 if (allocate_io_mem(td))
860 if (td->o.odirect || td->o.mem_align ||
861 (td->io_ops->flags & FIO_RAWIO))
862 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
866 cl_align = os_cache_line_size();
868 for (i = 0; i < max_units; i++) {
874 ptr = fio_memalign(cl_align, sizeof(*io_u));
876 log_err("fio: unable to allocate aligned memory\n");
881 memset(io_u, 0, sizeof(*io_u));
882 INIT_FLIST_HEAD(&io_u->list);
883 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
885 if (!(td->io_ops->flags & FIO_NOIO)) {
886 io_u->buf = p + max_bs * i;
887 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
890 io_u_fill_buffer(td, io_u, max_bs);
891 if (td_write(td) && td->o.verify_pattern_bytes) {
893 * Fill the buffer with the pattern if we are
894 * going to be doing writes.
896 fill_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
901 io_u->flags = IO_U_F_FREE;
902 flist_add(&io_u->list, &td->io_u_freelist);
908 static int switch_ioscheduler(struct thread_data *td)
910 char tmp[256], tmp2[128];
914 if (td->io_ops->flags & FIO_DISKLESSIO)
917 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
919 f = fopen(tmp, "r+");
921 if (errno == ENOENT) {
922 log_err("fio: os or kernel doesn't support IO scheduler"
926 td_verror(td, errno, "fopen iosched");
933 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
934 if (ferror(f) || ret != 1) {
935 td_verror(td, errno, "fwrite");
943 * Read back and check that the selected scheduler is now the default.
945 ret = fread(tmp, 1, sizeof(tmp), f);
946 if (ferror(f) || ret < 0) {
947 td_verror(td, errno, "fread");
952 sprintf(tmp2, "[%s]", td->o.ioscheduler);
953 if (!strstr(tmp, tmp2)) {
954 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
955 td_verror(td, EINVAL, "iosched_switch");
964 static int keep_running(struct thread_data *td)
966 unsigned long long io_done;
970 if (td->o.time_based)
977 io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
979 if (io_done < td->o.size)
985 static void reset_io_counters(struct thread_data *td)
987 td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0;
988 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
990 td->rate_bytes[0] = td->rate_bytes[1] = 0;
991 td->rate_blocks[0] = td->rate_blocks[1] = 0;
993 td->last_was_sync = 0;
996 * reset file done count if we are to start over
998 if (td->o.time_based || td->o.loops)
999 td->nr_done_files = 0;
1002 void reset_all_stats(struct thread_data *td)
1007 reset_io_counters(td);
1009 for (i = 0; i < 2; i++) {
1010 td->io_bytes[i] = 0;
1011 td->io_blocks[i] = 0;
1012 td->io_issues[i] = 0;
1013 td->ts.total_io_u[i] = 0;
1016 fio_gettime(&tv, NULL);
1017 td->ts.runtime[0] = 0;
1018 td->ts.runtime[1] = 0;
1019 memcpy(&td->epoch, &tv, sizeof(tv));
1020 memcpy(&td->start, &tv, sizeof(tv));
1023 static void clear_io_state(struct thread_data *td)
1028 reset_io_counters(td);
1031 for_each_file(td, f, i)
1032 fio_file_clear_done(f);
1035 * Set the same seed to get repeatable runs
1037 td_fill_rand_seeds(td);
1040 static int exec_string(const char *string)
1042 int ret, newlen = strlen(string) + 1 + 8;
1045 str = malloc(newlen);
1046 sprintf(str, "sh -c %s", string);
1050 log_err("fio: exec of cmd <%s> failed\n", str);
1057 * Entry point for the thread based jobs. The process based jobs end up
1058 * here as well, after a little setup.
1060 static void *thread_main(void *data)
1062 unsigned long long elapsed;
1063 struct thread_data *td = data;
1064 pthread_condattr_t attr;
1067 if (!td->o.use_thread) {
1073 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1075 INIT_FLIST_HEAD(&td->io_u_freelist);
1076 INIT_FLIST_HEAD(&td->io_u_busylist);
1077 INIT_FLIST_HEAD(&td->io_u_requeues);
1078 INIT_FLIST_HEAD(&td->io_log_list);
1079 INIT_FLIST_HEAD(&td->io_hist_list);
1080 INIT_FLIST_HEAD(&td->verify_list);
1081 INIT_FLIST_HEAD(&td->trim_list);
1082 pthread_mutex_init(&td->io_u_lock, NULL);
1083 td->io_hist_tree = RB_ROOT;
1085 pthread_condattr_init(&attr);
1086 pthread_cond_init(&td->verify_cond, &attr);
1087 pthread_cond_init(&td->free_cond, &attr);
1089 td_set_runstate(td, TD_INITIALIZED);
1090 dprint(FD_MUTEX, "up startup_mutex\n");
1091 fio_mutex_up(startup_mutex);
1092 dprint(FD_MUTEX, "wait on td->mutex\n");
1093 fio_mutex_down(td->mutex);
1094 dprint(FD_MUTEX, "done waiting on td->mutex\n");
1097 * the ->mutex mutex is now no longer used, close it to avoid
1098 * eating a file descriptor
1100 fio_mutex_remove(td->mutex);
1103 * A new gid requires privilege, so we need to do this before setting
1106 if (td->o.gid != -1U && setgid(td->o.gid)) {
1107 td_verror(td, errno, "setgid");
1110 if (td->o.uid != -1U && setuid(td->o.uid)) {
1111 td_verror(td, errno, "setuid");
1116 * If we have a gettimeofday() thread, make sure we exclude that
1117 * thread from this job
1120 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1123 * Set affinity first, in case it has an impact on the memory
1126 if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1127 td_verror(td, errno, "cpu_set_affinity");
1132 * May alter parameters that init_io_u() will use, so we need to
1141 if (td->o.verify_async && verify_async_init(td))
1144 if (td->ioprio_set) {
1145 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1146 td_verror(td, errno, "ioprio_set");
1151 if (td->o.cgroup_weight && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1154 if (nice(td->o.nice) == -1) {
1155 td_verror(td, errno, "nice");
1159 if (td->o.ioscheduler && switch_ioscheduler(td))
1162 if (!td->o.create_serialize && setup_files(td))
1168 if (init_random_map(td))
1171 if (td->o.exec_prerun) {
1172 if (exec_string(td->o.exec_prerun))
1176 if (td->o.pre_read) {
1177 if (pre_read_files(td) < 0)
1181 fio_gettime(&td->epoch, NULL);
1182 getrusage(RUSAGE_SELF, &td->ru_start);
1185 while (keep_running(td)) {
1186 fio_gettime(&td->start, NULL);
1187 memcpy(&td->stat_sample_time[0], &td->start, sizeof(td->start));
1188 memcpy(&td->stat_sample_time[1], &td->start, sizeof(td->start));
1189 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1191 if (td->o.ratemin[0] || td->o.ratemin[1])
1192 memcpy(&td->lastrate, &td->stat_sample_time,
1193 sizeof(td->lastrate));
1198 prune_io_piece_log(td);
1204 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1205 elapsed = utime_since_now(&td->start);
1206 td->ts.runtime[DDIR_READ] += elapsed;
1208 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1209 elapsed = utime_since_now(&td->start);
1210 td->ts.runtime[DDIR_WRITE] += elapsed;
1213 if (td->error || td->terminate)
1216 if (!td->o.do_verify ||
1217 td->o.verify == VERIFY_NONE ||
1218 (td->io_ops->flags & FIO_UNIDIR))
1223 fio_gettime(&td->start, NULL);
1227 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1229 if (td->error || td->terminate)
1233 update_rusage_stat(td);
1234 td->ts.runtime[0] = (td->ts.runtime[0] + 999) / 1000;
1235 td->ts.runtime[1] = (td->ts.runtime[1] + 999) / 1000;
1236 td->ts.total_run_time = mtime_since_now(&td->epoch);
1237 td->ts.io_bytes[0] = td->io_bytes[0];
1238 td->ts.io_bytes[1] = td->io_bytes[1];
1240 fio_mutex_down(writeout_mutex);
1242 if (td->o.bw_log_file) {
1243 finish_log_named(td, td->bw_log,
1244 td->o.bw_log_file, "bw");
1246 finish_log(td, td->bw_log, "bw");
1249 if (td->o.lat_log_file) {
1250 finish_log_named(td, td->lat_log,
1251 td->o.lat_log_file, "lat");
1253 finish_log(td, td->lat_log, "lat");
1256 if (td->o.lat_log_file) {
1257 finish_log_named(td, td->slat_log,
1258 td->o.lat_log_file, "slat");
1260 finish_log(td, td->slat_log, "slat");
1263 if (td->o.lat_log_file) {
1264 finish_log_named(td, td->clat_log,
1265 td->o.lat_log_file, "clat");
1267 finish_log(td, td->clat_log, "clat");
1269 fio_mutex_up(writeout_mutex);
1270 if (td->o.exec_postrun)
1271 exec_string(td->o.exec_postrun);
1273 if (exitall_on_terminate)
1274 fio_terminate_threads(td->groupid);
1278 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1281 if (td->o.verify_async)
1282 verify_async_exit(td);
1284 close_and_free_files(td);
1287 cgroup_shutdown(td, &cgroup_mnt);
1289 if (td->o.cpumask_set) {
1290 int ret = fio_cpuset_exit(&td->o.cpumask);
1292 td_verror(td, ret, "fio_cpuset_exit");
1296 * do this very late, it will log file closing as well
1298 if (td->o.write_iolog_file)
1299 write_iolog_close(td);
1301 options_mem_free(td);
1302 td_set_runstate(td, TD_EXITED);
1303 return (void *) (unsigned long) td->error;
1307 * We cannot pass the td data into a forked process, so attach the td and
1308 * pass it to the thread worker.
1310 static int fork_main(int shmid, int offset)
1312 struct thread_data *td;
1316 data = shmat(shmid, NULL, 0);
1317 if (data == (void *) -1) {
1325 * HP-UX inherits shm mappings?
1330 td = data + offset * sizeof(struct thread_data);
1331 ret = thread_main(td);
1333 return (int) (unsigned long) ret;
1337 * Run over the job map and reap the threads that have exited, if any.
1339 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1341 struct thread_data *td;
1342 int i, cputhreads, realthreads, pending, status, ret;
1345 * reap exited threads (TD_EXITED -> TD_REAPED)
1347 realthreads = pending = cputhreads = 0;
1348 for_each_td(td, i) {
1352 * ->io_ops is NULL for a thread that has closed its
1355 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1364 if (td->runstate == TD_REAPED)
1366 if (td->o.use_thread) {
1367 if (td->runstate == TD_EXITED) {
1368 td_set_runstate(td, TD_REAPED);
1375 if (td->runstate == TD_EXITED)
1379 * check if someone quit or got killed in an unusual way
1381 ret = waitpid(td->pid, &status, flags);
1383 if (errno == ECHILD) {
1384 log_err("fio: pid=%d disappeared %d\n",
1385 (int) td->pid, td->runstate);
1386 td_set_runstate(td, TD_REAPED);
1390 } else if (ret == td->pid) {
1391 if (WIFSIGNALED(status)) {
1392 int sig = WTERMSIG(status);
1395 log_err("fio: pid=%d, got signal=%d\n",
1396 (int) td->pid, sig);
1397 td_set_runstate(td, TD_REAPED);
1400 if (WIFEXITED(status)) {
1401 if (WEXITSTATUS(status) && !td->error)
1402 td->error = WEXITSTATUS(status);
1404 td_set_runstate(td, TD_REAPED);
1410 * thread is not dead, continue
1416 (*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
1417 (*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
1424 done_secs += mtime_since_now(&td->epoch) / 1000;
1427 if (*nr_running == cputhreads && !pending && realthreads)
1428 fio_terminate_threads(TERMINATE_ALL);
1431 static void *gtod_thread_main(void *data)
1433 fio_mutex_up(startup_mutex);
1436 * As long as we have jobs around, update the clock. It would be nice
1437 * to have some way of NOT hammering that CPU with gettimeofday(),
1438 * but I'm not sure what to use outside of a simple CPU nop to relax
1439 * it - we don't want to lose precision.
1449 static int fio_start_gtod_thread(void)
1451 pthread_attr_t attr;
1454 pthread_attr_init(&attr);
1455 pthread_attr_setstacksize(&attr, PTHREAD_STACK_MIN);
1456 ret = pthread_create(>od_thread, &attr, gtod_thread_main, NULL);
1457 pthread_attr_destroy(&attr);
1459 log_err("Can't create gtod thread: %s\n", strerror(ret));
1463 ret = pthread_detach(gtod_thread);
1465 log_err("Can't detatch gtod thread: %s\n", strerror(ret));
1469 dprint(FD_MUTEX, "wait on startup_mutex\n");
1470 fio_mutex_down(startup_mutex);
1471 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1476 * Main function for kicking off and reaping jobs, as needed.
1478 static void run_threads(void)
1480 struct thread_data *td;
1481 unsigned long spent;
1482 int i, todo, nr_running, m_rate, t_rate, nr_started;
1484 if (fio_pin_memory())
1487 if (fio_gtod_offload && fio_start_gtod_thread())
1492 if (!terse_output) {
1493 log_info("Starting ");
1495 log_info("%d thread%s", nr_thread,
1496 nr_thread > 1 ? "s" : "");
1500 log_info("%d process%s", nr_process,
1501 nr_process > 1 ? "es" : "");
1507 todo = thread_number;
1510 m_rate = t_rate = 0;
1512 for_each_td(td, i) {
1513 print_status_init(td->thread_number - 1);
1515 if (!td->o.create_serialize)
1519 * do file setup here so it happens sequentially,
1520 * we don't want X number of threads getting their
1521 * client data interspersed on disk
1523 if (setup_files(td)) {
1526 log_err("fio: pid=%d, err=%d/%s\n",
1527 (int) td->pid, td->error, td->verror);
1528 td_set_runstate(td, TD_REAPED);
1535 * for sharing to work, each job must always open
1536 * its own files. so close them, if we opened them
1539 for_each_file(td, f, j) {
1540 if (fio_file_open(f))
1541 td_io_close_file(td, f);
1549 struct thread_data *map[REAL_MAX_JOBS];
1550 struct timeval this_start;
1551 int this_jobs = 0, left;
1554 * create threads (TD_NOT_CREATED -> TD_CREATED)
1556 for_each_td(td, i) {
1557 if (td->runstate != TD_NOT_CREATED)
1561 * never got a chance to start, killed by other
1562 * thread for some reason
1564 if (td->terminate) {
1569 if (td->o.start_delay) {
1570 spent = mtime_since_genesis();
1572 if (td->o.start_delay * 1000 > spent)
1576 if (td->o.stonewall && (nr_started || nr_running)) {
1577 dprint(FD_PROCESS, "%s: stonewall wait\n",
1585 * Set state to created. Thread will transition
1586 * to TD_INITIALIZED when it's done setting up.
1588 td_set_runstate(td, TD_CREATED);
1589 map[this_jobs++] = td;
1592 if (td->o.use_thread) {
1595 dprint(FD_PROCESS, "will pthread_create\n");
1596 ret = pthread_create(&td->thread, NULL,
1599 log_err("pthread_create: %s\n",
1604 ret = pthread_detach(td->thread);
1606 log_err("pthread_detach: %s",
1610 dprint(FD_PROCESS, "will fork\n");
1613 int ret = fork_main(shm_id, i);
1616 } else if (i == fio_debug_jobno)
1617 *fio_debug_jobp = pid;
1619 dprint(FD_MUTEX, "wait on startup_mutex\n");
1620 if (fio_mutex_down_timeout(startup_mutex, 10)) {
1621 log_err("fio: job startup hung? exiting.\n");
1622 fio_terminate_threads(TERMINATE_ALL);
1627 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1631 * Wait for the started threads to transition to
1634 fio_gettime(&this_start, NULL);
1636 while (left && !fio_abort) {
1637 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1642 for (i = 0; i < this_jobs; i++) {
1646 if (td->runstate == TD_INITIALIZED) {
1649 } else if (td->runstate >= TD_EXITED) {
1653 nr_running++; /* work-around... */
1659 log_err("fio: %d jobs failed to start\n", left);
1660 for (i = 0; i < this_jobs; i++) {
1664 kill(td->pid, SIGTERM);
1670 * start created threads (TD_INITIALIZED -> TD_RUNNING).
1672 for_each_td(td, i) {
1673 if (td->runstate != TD_INITIALIZED)
1676 if (in_ramp_time(td))
1677 td_set_runstate(td, TD_RAMP);
1679 td_set_runstate(td, TD_RUNNING);
1682 m_rate += td->o.ratemin[0] + td->o.ratemin[1];
1683 t_rate += td->o.rate[0] + td->o.rate[1];
1685 fio_mutex_up(td->mutex);
1688 reap_threads(&nr_running, &t_rate, &m_rate);
1694 while (nr_running) {
1695 reap_threads(&nr_running, &t_rate, &m_rate);
1698 fio_server_idle_loop();
1710 return fio_handle_clients();
1711 if (exec_profile && load_profile(exec_profile))
1718 setup_log(&agg_io_log[DDIR_READ]);
1719 setup_log(&agg_io_log[DDIR_WRITE]);
1722 startup_mutex = fio_mutex_init(0);
1723 if (startup_mutex == NULL)
1725 writeout_mutex = fio_mutex_init(1);
1726 if (writeout_mutex == NULL)
1730 create_disk_util_thread();
1732 cgroup_list = smalloc(sizeof(*cgroup_list));
1733 INIT_FLIST_HEAD(cgroup_list);
1740 __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1741 __finish_log(agg_io_log[DDIR_WRITE],
1742 "agg-write_bw.log");
1746 cgroup_kill(cgroup_list);
1750 fio_mutex_remove(startup_mutex);
1751 fio_mutex_remove(writeout_mutex);
1755 void reset_fio_state(void)
1764 int main(int argc, char *argv[], char *envp[])
1773 * We need locale for number printing, if it isn't set then just
1774 * go with the US format.
1776 if (!getenv("LC_NUMERIC"))
1777 setlocale(LC_NUMERIC, "en_US");
1779 ps = sysconf(_SC_PAGESIZE);
1781 log_err("Failed to get page size\n");
1788 fio_keywords_init();
1790 if (parse_options(argc, argv))