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 TERMINATE_ALL (-1)
79 #define JOB_START_TIMEOUT (5 * 1000)
81 void td_set_runstate(struct thread_data *td, int runstate)
83 if (td->runstate == runstate)
86 dprint(FD_PROCESS, "pid=%d: runstate %d -> %d\n", (int) td->pid,
87 td->runstate, runstate);
88 td->runstate = runstate;
91 static void terminate_threads(int group_id)
93 struct thread_data *td;
96 dprint(FD_PROCESS, "terminate group_id=%d\n", group_id);
99 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
100 dprint(FD_PROCESS, "setting terminate on %s/%d\n",
101 td->o.name, (int) td->pid);
103 td->o.start_delay = 0;
106 * if the thread is running, just let it exit
110 else if (td->runstate < TD_RAMP)
111 kill(td->pid, SIGTERM);
113 struct ioengine_ops *ops = td->io_ops;
115 if (ops && (ops->flags & FIO_SIGTERM))
116 kill(td->pid, SIGTERM);
122 static void sig_int(int sig)
125 log_info("\nfio: terminating on signal %d\n", sig);
129 terminate_threads(TERMINATE_ALL);
133 static void *disk_thread_main(void *data)
135 fio_mutex_up(startup_mutex);
138 usleep(DISK_UTIL_MSEC * 1000);
142 print_thread_status();
148 static int create_disk_util_thread(void)
152 ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
154 log_err("Can't create disk util thread: %s\n", strerror(ret));
158 ret = pthread_detach(disk_util_thread);
160 log_err("Can't detatch disk util thread: %s\n", strerror(ret));
164 dprint(FD_MUTEX, "wait on startup_mutex\n");
165 fio_mutex_down(startup_mutex);
166 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
170 static void set_sig_handlers(void)
172 struct sigaction act;
174 memset(&act, 0, sizeof(act));
175 act.sa_handler = sig_int;
176 act.sa_flags = SA_RESTART;
177 sigaction(SIGINT, &act, NULL);
179 memset(&act, 0, sizeof(act));
180 act.sa_handler = sig_int;
181 act.sa_flags = SA_RESTART;
182 sigaction(SIGTERM, &act, NULL);
186 * Check if we are above the minimum rate given.
188 static int __check_min_rate(struct thread_data *td, struct timeval *now,
191 unsigned long long bytes = 0;
192 unsigned long iops = 0;
195 unsigned int ratemin = 0;
196 unsigned int rate_iops = 0;
197 unsigned int rate_iops_min = 0;
199 assert(ddir_rw(ddir));
201 if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
205 * allow a 2 second settle period in the beginning
207 if (mtime_since(&td->start, now) < 2000)
210 iops += td->io_blocks[ddir];
211 bytes += td->this_io_bytes[ddir];
212 ratemin += td->o.ratemin[ddir];
213 rate_iops += td->o.rate_iops[ddir];
214 rate_iops_min += td->o.rate_iops_min[ddir];
217 * if rate blocks is set, sample is running
219 if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
220 spent = mtime_since(&td->lastrate[ddir], now);
221 if (spent < td->o.ratecycle)
224 if (td->o.rate[ddir]) {
226 * check bandwidth specified rate
228 if (bytes < td->rate_bytes[ddir]) {
229 log_err("%s: min rate %u not met\n", td->o.name,
233 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
234 if (rate < ratemin ||
235 bytes < td->rate_bytes[ddir]) {
236 log_err("%s: min rate %u not met, got"
237 " %luKB/sec\n", td->o.name,
244 * checks iops specified rate
246 if (iops < rate_iops) {
247 log_err("%s: min iops rate %u not met\n",
248 td->o.name, rate_iops);
251 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
252 if (rate < rate_iops_min ||
253 iops < td->rate_blocks[ddir]) {
254 log_err("%s: min iops rate %u not met,"
255 " got %lu\n", td->o.name,
256 rate_iops_min, rate);
262 td->rate_bytes[ddir] = bytes;
263 td->rate_blocks[ddir] = iops;
264 memcpy(&td->lastrate[ddir], now, sizeof(*now));
268 static int check_min_rate(struct thread_data *td, struct timeval *now,
269 unsigned long *bytes_done)
274 ret |= __check_min_rate(td, now, 0);
276 ret |= __check_min_rate(td, now, 1);
281 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
285 if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
292 * When job exits, we can cancel the in-flight IO if we are using async
293 * io. Attempt to do so.
295 static void cleanup_pending_aio(struct thread_data *td)
297 struct flist_head *entry, *n;
302 * get immediately available events, if any
304 r = io_u_queued_complete(td, 0, NULL);
309 * now cancel remaining active events
311 if (td->io_ops->cancel) {
312 flist_for_each_safe(entry, n, &td->io_u_busylist) {
313 io_u = flist_entry(entry, struct io_u, list);
316 * if the io_u isn't in flight, then that generally
317 * means someone leaked an io_u. complain but fix
318 * it up, so we don't stall here.
320 if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
321 log_err("fio: non-busy IO on busy list\n");
324 r = td->io_ops->cancel(td, io_u);
332 r = io_u_queued_complete(td, td->cur_depth, NULL);
336 * Helper to handle the final sync of a file. Works just like the normal
337 * io path, just does everything sync.
339 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
341 struct io_u *io_u = __get_io_u(td);
347 io_u->ddir = DDIR_SYNC;
350 if (td_io_prep(td, io_u)) {
356 ret = td_io_queue(td, io_u);
358 td_verror(td, io_u->error, "td_io_queue");
361 } else if (ret == FIO_Q_QUEUED) {
362 if (io_u_queued_complete(td, 1, NULL) < 0)
364 } else if (ret == FIO_Q_COMPLETED) {
366 td_verror(td, io_u->error, "td_io_queue");
370 if (io_u_sync_complete(td, io_u, NULL) < 0)
372 } else if (ret == FIO_Q_BUSY) {
373 if (td_io_commit(td))
381 static inline void __update_tv_cache(struct thread_data *td)
383 fio_gettime(&td->tv_cache, NULL);
386 static inline void update_tv_cache(struct thread_data *td)
388 if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
389 __update_tv_cache(td);
392 static int break_on_this_error(struct thread_data *td, int *retptr)
396 if (ret < 0 || td->error) {
399 if (!td->o.continue_on_error)
407 if (td_non_fatal_error(err)) {
409 * Continue with the I/Os in case of
412 update_error_count(td, err);
416 } else if (td->o.fill_device && err == ENOSPC) {
418 * We expect to hit this error if
419 * fill_device option is set.
426 * Stop the I/O in case of a fatal
429 update_error_count(td, err);
438 * The main verify engine. Runs over the writes we previously submitted,
439 * reads the blocks back in, and checks the crc/md5 of the data.
441 static void do_verify(struct thread_data *td)
448 dprint(FD_VERIFY, "starting loop\n");
451 * sync io first and invalidate cache, to make sure we really
454 for_each_file(td, f, i) {
455 if (!fio_file_open(f))
457 if (fio_io_sync(td, f))
459 if (file_invalidate_cache(td, f))
466 td_set_runstate(td, TD_VERIFYING);
469 while (!td->terminate) {
474 if (runtime_exceeded(td, &td->tv_cache)) {
475 __update_tv_cache(td);
476 if (runtime_exceeded(td, &td->tv_cache)) {
482 io_u = __get_io_u(td);
486 if (get_next_verify(td, io_u)) {
491 if (td_io_prep(td, io_u)) {
496 if (td->o.verify_async)
497 io_u->end_io = verify_io_u_async;
499 io_u->end_io = verify_io_u;
501 ret = td_io_queue(td, io_u);
503 case FIO_Q_COMPLETED:
506 clear_io_u(td, io_u);
507 } else if (io_u->resid) {
508 int bytes = io_u->xfer_buflen - io_u->resid;
514 td_verror(td, EIO, "full resid");
519 io_u->xfer_buflen = io_u->resid;
520 io_u->xfer_buf += bytes;
521 io_u->offset += bytes;
523 if (ddir_rw(io_u->ddir))
524 td->ts.short_io_u[io_u->ddir]++;
527 if (io_u->offset == f->real_file_size)
530 requeue_io_u(td, &io_u);
533 ret = io_u_sync_complete(td, io_u, NULL);
541 requeue_io_u(td, &io_u);
542 ret2 = td_io_commit(td);
548 td_verror(td, -ret, "td_io_queue");
552 if (break_on_this_error(td, &ret))
556 * if we can queue more, do so. but check if there are
557 * completed io_u's first. Note that we can get BUSY even
558 * without IO queued, if the system is resource starved.
560 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
561 if (full || !td->o.iodepth_batch_complete) {
562 min_events = min(td->o.iodepth_batch_complete,
564 if (full && !min_events && td->o.iodepth_batch_complete != 0)
569 * Reap required number of io units, if any,
570 * and do the verification on them through
571 * the callback handler
573 if (io_u_queued_complete(td, min_events, NULL) < 0) {
577 } while (full && (td->cur_depth > td->o.iodepth_low));
584 min_events = td->cur_depth;
587 ret = io_u_queued_complete(td, min_events, NULL);
589 cleanup_pending_aio(td);
591 td_set_runstate(td, TD_RUNNING);
593 dprint(FD_VERIFY, "exiting loop\n");
597 * Main IO worker function. It retrieves io_u's to process and queues
598 * and reaps them, checking for rate and errors along the way.
600 static void do_io(struct thread_data *td)
605 if (in_ramp_time(td))
606 td_set_runstate(td, TD_RAMP);
608 td_set_runstate(td, TD_RUNNING);
610 while ( (td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
611 (!flist_empty(&td->trim_list)) ||
612 ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) ) {
613 struct timeval comp_time;
614 unsigned long bytes_done[2] = { 0, 0 };
624 if (runtime_exceeded(td, &td->tv_cache)) {
625 __update_tv_cache(td);
626 if (runtime_exceeded(td, &td->tv_cache)) {
637 * Add verification end_io handler, if asked to verify
638 * a previously written file.
640 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
642 if (td->o.verify_async)
643 io_u->end_io = verify_io_u_async;
645 io_u->end_io = verify_io_u;
646 td_set_runstate(td, TD_VERIFYING);
647 } else if (in_ramp_time(td))
648 td_set_runstate(td, TD_RAMP);
650 td_set_runstate(td, TD_RUNNING);
652 ret = td_io_queue(td, io_u);
654 case FIO_Q_COMPLETED:
657 clear_io_u(td, io_u);
658 } else if (io_u->resid) {
659 int bytes = io_u->xfer_buflen - io_u->resid;
660 struct fio_file *f = io_u->file;
666 td_verror(td, EIO, "full resid");
671 io_u->xfer_buflen = io_u->resid;
672 io_u->xfer_buf += bytes;
673 io_u->offset += bytes;
675 if (ddir_rw(io_u->ddir))
676 td->ts.short_io_u[io_u->ddir]++;
678 if (io_u->offset == f->real_file_size)
681 requeue_io_u(td, &io_u);
684 if (__should_check_rate(td, 0) ||
685 __should_check_rate(td, 1))
686 fio_gettime(&comp_time, NULL);
688 ret = io_u_sync_complete(td, io_u, bytes_done);
695 * if the engine doesn't have a commit hook,
696 * the io_u is really queued. if it does have such
697 * a hook, it has to call io_u_queued() itself.
699 if (td->io_ops->commit == NULL)
700 io_u_queued(td, io_u);
703 requeue_io_u(td, &io_u);
704 ret2 = td_io_commit(td);
714 if (break_on_this_error(td, &ret))
718 * See if we need to complete some commands. Note that we
719 * can get BUSY even without IO queued, if the system is
722 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
723 if (full || !td->o.iodepth_batch_complete) {
724 min_evts = min(td->o.iodepth_batch_complete,
726 if (full && !min_evts && td->o.iodepth_batch_complete != 0)
729 if (__should_check_rate(td, 0) ||
730 __should_check_rate(td, 1))
731 fio_gettime(&comp_time, NULL);
734 ret = io_u_queued_complete(td, min_evts, bytes_done);
738 } while (full && (td->cur_depth > td->o.iodepth_low));
743 if (!(bytes_done[0] + bytes_done[1]))
746 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
747 if (check_min_rate(td, &comp_time, bytes_done)) {
748 if (exitall_on_terminate)
749 terminate_threads(td->groupid);
750 td_verror(td, EIO, "check_min_rate");
755 if (td->o.thinktime) {
756 unsigned long long b;
758 b = td->io_blocks[0] + td->io_blocks[1];
759 if (!(b % td->o.thinktime_blocks)) {
762 if (td->o.thinktime_spin)
763 usec_spin(td->o.thinktime_spin);
765 left = td->o.thinktime - td->o.thinktime_spin;
767 usec_sleep(td, left);
772 if (td->trim_entries)
773 printf("trim entries %ld\n", td->trim_entries);
775 if (td->o.fill_device && td->error == ENOSPC) {
784 ret = io_u_queued_complete(td, i, NULL);
785 if (td->o.fill_device && td->error == ENOSPC)
789 if (should_fsync(td) && td->o.end_fsync) {
790 td_set_runstate(td, TD_FSYNCING);
792 for_each_file(td, f, i) {
793 if (!fio_file_open(f))
799 cleanup_pending_aio(td);
802 * stop job if we failed doing any IO
804 if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
808 static void cleanup_io_u(struct thread_data *td)
810 struct flist_head *entry, *n;
813 flist_for_each_safe(entry, n, &td->io_u_freelist) {
814 io_u = flist_entry(entry, struct io_u, list);
816 flist_del(&io_u->list);
817 fio_memfree(io_u, sizeof(*io_u));
823 static int init_io_u(struct thread_data *td)
827 int cl_align, i, max_units;
830 max_units = td->o.iodepth;
831 max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
832 td->orig_buffer_size = (unsigned long long) max_bs
833 * (unsigned long long) max_units;
835 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
838 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
839 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
842 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
843 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
847 if (allocate_io_mem(td))
850 if (td->o.odirect || td->o.mem_align ||
851 (td->io_ops->flags & FIO_RAWIO))
852 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
856 cl_align = os_cache_line_size();
858 for (i = 0; i < max_units; i++) {
864 ptr = fio_memalign(cl_align, sizeof(*io_u));
866 log_err("fio: unable to allocate aligned memory\n");
871 memset(io_u, 0, sizeof(*io_u));
872 INIT_FLIST_HEAD(&io_u->list);
873 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
875 if (!(td->io_ops->flags & FIO_NOIO)) {
876 io_u->buf = p + max_bs * i;
877 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
880 io_u_fill_buffer(td, io_u, max_bs);
881 if (td_write(td) && td->o.verify_pattern_bytes) {
883 * Fill the buffer with the pattern if we are
884 * going to be doing writes.
886 fill_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
891 io_u->flags = IO_U_F_FREE;
892 flist_add(&io_u->list, &td->io_u_freelist);
898 static int switch_ioscheduler(struct thread_data *td)
900 char tmp[256], tmp2[128];
904 if (td->io_ops->flags & FIO_DISKLESSIO)
907 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
909 f = fopen(tmp, "r+");
911 if (errno == ENOENT) {
912 log_err("fio: os or kernel doesn't support IO scheduler"
916 td_verror(td, errno, "fopen iosched");
923 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
924 if (ferror(f) || ret != 1) {
925 td_verror(td, errno, "fwrite");
933 * Read back and check that the selected scheduler is now the default.
935 ret = fread(tmp, 1, sizeof(tmp), f);
936 if (ferror(f) || ret < 0) {
937 td_verror(td, errno, "fread");
942 sprintf(tmp2, "[%s]", td->o.ioscheduler);
943 if (!strstr(tmp, tmp2)) {
944 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
945 td_verror(td, EINVAL, "iosched_switch");
954 static int keep_running(struct thread_data *td)
956 unsigned long long io_done;
960 if (td->o.time_based)
967 io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
969 if (io_done < td->o.size)
975 static void reset_io_counters(struct thread_data *td)
977 td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0;
978 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
980 td->rate_bytes[0] = td->rate_bytes[1] = 0;
981 td->rate_blocks[0] = td->rate_blocks[1] = 0;
983 td->last_was_sync = 0;
986 * reset file done count if we are to start over
988 if (td->o.time_based || td->o.loops)
989 td->nr_done_files = 0;
992 void reset_all_stats(struct thread_data *td)
997 reset_io_counters(td);
999 for (i = 0; i < 2; i++) {
1000 td->io_bytes[i] = 0;
1001 td->io_blocks[i] = 0;
1002 td->io_issues[i] = 0;
1003 td->ts.total_io_u[i] = 0;
1006 fio_gettime(&tv, NULL);
1007 td->ts.runtime[0] = 0;
1008 td->ts.runtime[1] = 0;
1009 memcpy(&td->epoch, &tv, sizeof(tv));
1010 memcpy(&td->start, &tv, sizeof(tv));
1013 static void clear_io_state(struct thread_data *td)
1018 reset_io_counters(td);
1021 for_each_file(td, f, i)
1022 fio_file_clear_done(f);
1025 * Set the same seed to get repeatable runs
1027 td_fill_rand_seeds(td);
1030 static int exec_string(const char *string)
1032 int ret, newlen = strlen(string) + 1 + 8;
1035 str = malloc(newlen);
1036 sprintf(str, "sh -c %s", string);
1040 log_err("fio: exec of cmd <%s> failed\n", str);
1047 * Entry point for the thread based jobs. The process based jobs end up
1048 * here as well, after a little setup.
1050 static void *thread_main(void *data)
1052 unsigned long long elapsed;
1053 struct thread_data *td = data;
1054 pthread_condattr_t attr;
1057 if (!td->o.use_thread) {
1063 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1065 INIT_FLIST_HEAD(&td->io_u_freelist);
1066 INIT_FLIST_HEAD(&td->io_u_busylist);
1067 INIT_FLIST_HEAD(&td->io_u_requeues);
1068 INIT_FLIST_HEAD(&td->io_log_list);
1069 INIT_FLIST_HEAD(&td->io_hist_list);
1070 INIT_FLIST_HEAD(&td->verify_list);
1071 INIT_FLIST_HEAD(&td->trim_list);
1072 pthread_mutex_init(&td->io_u_lock, NULL);
1073 td->io_hist_tree = RB_ROOT;
1075 pthread_condattr_init(&attr);
1076 pthread_cond_init(&td->verify_cond, &attr);
1077 pthread_cond_init(&td->free_cond, &attr);
1079 td_set_runstate(td, TD_INITIALIZED);
1080 dprint(FD_MUTEX, "up startup_mutex\n");
1081 fio_mutex_up(startup_mutex);
1082 dprint(FD_MUTEX, "wait on td->mutex\n");
1083 fio_mutex_down(td->mutex);
1084 dprint(FD_MUTEX, "done waiting on td->mutex\n");
1087 * the ->mutex mutex is now no longer used, close it to avoid
1088 * eating a file descriptor
1090 fio_mutex_remove(td->mutex);
1093 * A new gid requires privilege, so we need to do this before setting
1096 if (td->o.gid != -1U && setgid(td->o.gid)) {
1097 td_verror(td, errno, "setgid");
1100 if (td->o.uid != -1U && setuid(td->o.uid)) {
1101 td_verror(td, errno, "setuid");
1106 * If we have a gettimeofday() thread, make sure we exclude that
1107 * thread from this job
1110 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1113 * Set affinity first, in case it has an impact on the memory
1116 if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1117 td_verror(td, errno, "cpu_set_affinity");
1122 * May alter parameters that init_io_u() will use, so we need to
1131 if (td->o.verify_async && verify_async_init(td))
1134 if (td->ioprio_set) {
1135 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1136 td_verror(td, errno, "ioprio_set");
1141 if (td->o.cgroup_weight && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1144 if (nice(td->o.nice) == -1) {
1145 td_verror(td, errno, "nice");
1149 if (td->o.ioscheduler && switch_ioscheduler(td))
1152 if (!td->o.create_serialize && setup_files(td))
1158 if (init_random_map(td))
1161 if (td->o.exec_prerun) {
1162 if (exec_string(td->o.exec_prerun))
1166 if (td->o.pre_read) {
1167 if (pre_read_files(td) < 0)
1171 fio_gettime(&td->epoch, NULL);
1172 getrusage(RUSAGE_SELF, &td->ru_start);
1175 while (keep_running(td)) {
1176 fio_gettime(&td->start, NULL);
1177 memcpy(&td->stat_sample_time[0], &td->start, sizeof(td->start));
1178 memcpy(&td->stat_sample_time[1], &td->start, sizeof(td->start));
1179 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1181 if (td->o.ratemin[0] || td->o.ratemin[1])
1182 memcpy(&td->lastrate, &td->stat_sample_time,
1183 sizeof(td->lastrate));
1188 prune_io_piece_log(td);
1194 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1195 elapsed = utime_since_now(&td->start);
1196 td->ts.runtime[DDIR_READ] += elapsed;
1198 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1199 elapsed = utime_since_now(&td->start);
1200 td->ts.runtime[DDIR_WRITE] += elapsed;
1203 if (td->error || td->terminate)
1206 if (!td->o.do_verify ||
1207 td->o.verify == VERIFY_NONE ||
1208 (td->io_ops->flags & FIO_UNIDIR))
1213 fio_gettime(&td->start, NULL);
1217 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1219 if (td->error || td->terminate)
1223 update_rusage_stat(td);
1224 td->ts.runtime[0] = (td->ts.runtime[0] + 999) / 1000;
1225 td->ts.runtime[1] = (td->ts.runtime[1] + 999) / 1000;
1226 td->ts.total_run_time = mtime_since_now(&td->epoch);
1227 td->ts.io_bytes[0] = td->io_bytes[0];
1228 td->ts.io_bytes[1] = td->io_bytes[1];
1230 fio_mutex_down(writeout_mutex);
1232 if (td->o.bw_log_file) {
1233 finish_log_named(td, td->bw_log,
1234 td->o.bw_log_file, "bw");
1236 finish_log(td, td->bw_log, "bw");
1239 if (td->o.lat_log_file) {
1240 finish_log_named(td, td->lat_log,
1241 td->o.lat_log_file, "lat");
1243 finish_log(td, td->lat_log, "lat");
1246 if (td->o.lat_log_file) {
1247 finish_log_named(td, td->slat_log,
1248 td->o.lat_log_file, "slat");
1250 finish_log(td, td->slat_log, "slat");
1253 if (td->o.lat_log_file) {
1254 finish_log_named(td, td->clat_log,
1255 td->o.lat_log_file, "clat");
1257 finish_log(td, td->clat_log, "clat");
1259 fio_mutex_up(writeout_mutex);
1260 if (td->o.exec_postrun)
1261 exec_string(td->o.exec_postrun);
1263 if (exitall_on_terminate)
1264 terminate_threads(td->groupid);
1268 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1271 if (td->o.verify_async)
1272 verify_async_exit(td);
1274 close_and_free_files(td);
1277 cgroup_shutdown(td, &cgroup_mnt);
1279 if (td->o.cpumask_set) {
1280 int ret = fio_cpuset_exit(&td->o.cpumask);
1282 td_verror(td, ret, "fio_cpuset_exit");
1286 * do this very late, it will log file closing as well
1288 if (td->o.write_iolog_file)
1289 write_iolog_close(td);
1291 options_mem_free(td);
1292 td_set_runstate(td, TD_EXITED);
1293 return (void *) (unsigned long) td->error;
1297 * We cannot pass the td data into a forked process, so attach the td and
1298 * pass it to the thread worker.
1300 static int fork_main(int shmid, int offset)
1302 struct thread_data *td;
1306 data = shmat(shmid, NULL, 0);
1307 if (data == (void *) -1) {
1315 * HP-UX inherits shm mappings?
1320 td = data + offset * sizeof(struct thread_data);
1321 ret = thread_main(td);
1323 return (int) (unsigned long) ret;
1327 * Run over the job map and reap the threads that have exited, if any.
1329 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1331 struct thread_data *td;
1332 int i, cputhreads, realthreads, pending, status, ret;
1335 * reap exited threads (TD_EXITED -> TD_REAPED)
1337 realthreads = pending = cputhreads = 0;
1338 for_each_td(td, i) {
1342 * ->io_ops is NULL for a thread that has closed its
1345 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1354 if (td->runstate == TD_REAPED)
1356 if (td->o.use_thread) {
1357 if (td->runstate == TD_EXITED) {
1358 td_set_runstate(td, TD_REAPED);
1365 if (td->runstate == TD_EXITED)
1369 * check if someone quit or got killed in an unusual way
1371 ret = waitpid(td->pid, &status, flags);
1373 if (errno == ECHILD) {
1374 log_err("fio: pid=%d disappeared %d\n",
1375 (int) td->pid, td->runstate);
1376 td_set_runstate(td, TD_REAPED);
1380 } else if (ret == td->pid) {
1381 if (WIFSIGNALED(status)) {
1382 int sig = WTERMSIG(status);
1385 log_err("fio: pid=%d, got signal=%d\n",
1386 (int) td->pid, sig);
1387 td_set_runstate(td, TD_REAPED);
1390 if (WIFEXITED(status)) {
1391 if (WEXITSTATUS(status) && !td->error)
1392 td->error = WEXITSTATUS(status);
1394 td_set_runstate(td, TD_REAPED);
1400 * thread is not dead, continue
1406 (*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
1407 (*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
1414 done_secs += mtime_since_now(&td->epoch) / 1000;
1417 if (*nr_running == cputhreads && !pending && realthreads)
1418 terminate_threads(TERMINATE_ALL);
1421 static void *gtod_thread_main(void *data)
1423 fio_mutex_up(startup_mutex);
1426 * As long as we have jobs around, update the clock. It would be nice
1427 * to have some way of NOT hammering that CPU with gettimeofday(),
1428 * but I'm not sure what to use outside of a simple CPU nop to relax
1429 * it - we don't want to lose precision.
1439 static int fio_start_gtod_thread(void)
1441 pthread_attr_t attr;
1444 pthread_attr_init(&attr);
1445 pthread_attr_setstacksize(&attr, PTHREAD_STACK_MIN);
1446 ret = pthread_create(>od_thread, &attr, gtod_thread_main, NULL);
1447 pthread_attr_destroy(&attr);
1449 log_err("Can't create gtod thread: %s\n", strerror(ret));
1453 ret = pthread_detach(gtod_thread);
1455 log_err("Can't detatch gtod thread: %s\n", strerror(ret));
1459 dprint(FD_MUTEX, "wait on startup_mutex\n");
1460 fio_mutex_down(startup_mutex);
1461 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1466 * Main function for kicking off and reaping jobs, as needed.
1468 static void run_threads(void)
1470 struct thread_data *td;
1471 unsigned long spent;
1472 int i, todo, nr_running, m_rate, t_rate, nr_started;
1474 if (fio_pin_memory())
1477 if (fio_gtod_offload && fio_start_gtod_thread())
1480 if (!terse_output) {
1481 log_info("Starting ");
1483 log_info("%d thread%s", nr_thread,
1484 nr_thread > 1 ? "s" : "");
1488 log_info("%d process%s", nr_process,
1489 nr_process > 1 ? "es" : "");
1497 todo = thread_number;
1500 m_rate = t_rate = 0;
1502 for_each_td(td, i) {
1503 print_status_init(td->thread_number - 1);
1505 if (!td->o.create_serialize)
1509 * do file setup here so it happens sequentially,
1510 * we don't want X number of threads getting their
1511 * client data interspersed on disk
1513 if (setup_files(td)) {
1516 log_err("fio: pid=%d, err=%d/%s\n",
1517 (int) td->pid, td->error, td->verror);
1518 td_set_runstate(td, TD_REAPED);
1525 * for sharing to work, each job must always open
1526 * its own files. so close them, if we opened them
1529 for_each_file(td, f, j) {
1530 if (fio_file_open(f))
1531 td_io_close_file(td, f);
1539 struct thread_data *map[REAL_MAX_JOBS];
1540 struct timeval this_start;
1541 int this_jobs = 0, left;
1544 * create threads (TD_NOT_CREATED -> TD_CREATED)
1546 for_each_td(td, i) {
1547 if (td->runstate != TD_NOT_CREATED)
1551 * never got a chance to start, killed by other
1552 * thread for some reason
1554 if (td->terminate) {
1559 if (td->o.start_delay) {
1560 spent = mtime_since_genesis();
1562 if (td->o.start_delay * 1000 > spent)
1566 if (td->o.stonewall && (nr_started || nr_running)) {
1567 dprint(FD_PROCESS, "%s: stonewall wait\n",
1575 * Set state to created. Thread will transition
1576 * to TD_INITIALIZED when it's done setting up.
1578 td_set_runstate(td, TD_CREATED);
1579 map[this_jobs++] = td;
1582 if (td->o.use_thread) {
1585 dprint(FD_PROCESS, "will pthread_create\n");
1586 ret = pthread_create(&td->thread, NULL,
1589 log_err("pthread_create: %s\n",
1594 ret = pthread_detach(td->thread);
1596 log_err("pthread_detach: %s",
1600 dprint(FD_PROCESS, "will fork\n");
1603 int ret = fork_main(shm_id, i);
1606 } else if (i == fio_debug_jobno)
1607 *fio_debug_jobp = pid;
1609 dprint(FD_MUTEX, "wait on startup_mutex\n");
1610 if (fio_mutex_down_timeout(startup_mutex, 10)) {
1611 log_err("fio: job startup hung? exiting.\n");
1612 terminate_threads(TERMINATE_ALL);
1617 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1621 * Wait for the started threads to transition to
1624 fio_gettime(&this_start, NULL);
1626 while (left && !fio_abort) {
1627 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1632 for (i = 0; i < this_jobs; i++) {
1636 if (td->runstate == TD_INITIALIZED) {
1639 } else if (td->runstate >= TD_EXITED) {
1643 nr_running++; /* work-around... */
1649 log_err("fio: %d jobs failed to start\n", left);
1650 for (i = 0; i < this_jobs; i++) {
1654 kill(td->pid, SIGTERM);
1660 * start created threads (TD_INITIALIZED -> TD_RUNNING).
1662 for_each_td(td, i) {
1663 if (td->runstate != TD_INITIALIZED)
1666 if (in_ramp_time(td))
1667 td_set_runstate(td, TD_RAMP);
1669 td_set_runstate(td, TD_RUNNING);
1672 m_rate += td->o.ratemin[0] + td->o.ratemin[1];
1673 t_rate += td->o.rate[0] + td->o.rate[1];
1675 fio_mutex_up(td->mutex);
1678 reap_threads(&nr_running, &t_rate, &m_rate);
1684 while (nr_running) {
1685 reap_threads(&nr_running, &t_rate, &m_rate);
1696 return fio_handle_clients();
1697 if (exec_profile && load_profile(exec_profile))
1704 setup_log(&agg_io_log[DDIR_READ]);
1705 setup_log(&agg_io_log[DDIR_WRITE]);
1708 startup_mutex = fio_mutex_init(0);
1709 if (startup_mutex == NULL)
1711 writeout_mutex = fio_mutex_init(1);
1712 if (writeout_mutex == NULL)
1716 create_disk_util_thread();
1718 cgroup_list = smalloc(sizeof(*cgroup_list));
1719 INIT_FLIST_HEAD(cgroup_list);
1726 __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1727 __finish_log(agg_io_log[DDIR_WRITE],
1728 "agg-write_bw.log");
1732 cgroup_kill(cgroup_list);
1736 fio_mutex_remove(startup_mutex);
1737 fio_mutex_remove(writeout_mutex);
1741 void reset_fio_state(void)
1750 int main(int argc, char *argv[], char *envp[])
1759 * We need locale for number printing, if it isn't set then just
1760 * go with the US format.
1762 if (!getenv("LC_NUMERIC"))
1763 setlocale(LC_NUMERIC, "en_US");
1765 ps = sysconf(_SC_PAGESIZE);
1767 log_err("Failed to get page size\n");
1774 fio_keywords_init();
1776 if (parse_options(argc, argv))