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
49 unsigned long page_mask;
50 unsigned long page_size;
52 #define PAGE_ALIGN(buf) \
53 (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
56 int thread_number = 0;
61 unsigned long done_secs = 0;
63 static struct fio_mutex *startup_mutex;
64 static struct fio_mutex *writeout_mutex;
65 static volatile int fio_abort;
66 static int exit_value;
67 static timer_t ival_timer;
68 static pthread_t gtod_thread;
69 static struct flist_head *cgroup_list;
70 static char *cgroup_mnt;
72 struct io_log *agg_io_log[2];
74 #define TERMINATE_ALL (-1)
75 #define JOB_START_TIMEOUT (5 * 1000)
77 void td_set_runstate(struct thread_data *td, int runstate)
79 if (td->runstate == runstate)
82 dprint(FD_PROCESS, "pid=%d: runstate %d -> %d\n", (int) td->pid,
83 td->runstate, runstate);
84 td->runstate = runstate;
87 static void terminate_threads(int group_id)
89 struct thread_data *td;
92 dprint(FD_PROCESS, "terminate group_id=%d\n", group_id);
95 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
96 dprint(FD_PROCESS, "setting terminate on %s/%d\n",
97 td->o.name, (int) td->pid);
99 td->o.start_delay = 0;
102 * if the thread is running, just let it exit
104 if (td->runstate < TD_RUNNING)
105 kill(td->pid, SIGTERM);
107 struct ioengine_ops *ops = td->io_ops;
109 if (ops && (ops->flags & FIO_SIGTERM))
110 kill(td->pid, SIGTERM);
116 static void status_timer_arm(void)
118 struct itimerspec value;
120 value.it_value.tv_sec = 0;
121 value.it_value.tv_nsec = DISK_UTIL_MSEC * 1000000;
122 value.it_interval.tv_sec = 0;
123 value.it_interval.tv_nsec = DISK_UTIL_MSEC * 1000000;
125 timer_settime(ival_timer, 0, &value, NULL);
128 static void ival_fn(union sigval sig)
132 print_thread_status();
137 * Happens on thread runs with ctrl-c, ignore our own SIGQUIT
139 static void sig_quit(int sig)
143 static void sig_int(int sig)
146 log_info("\nfio: terminating on signal %d\n", sig);
148 terminate_threads(TERMINATE_ALL);
152 static void posix_timer_teardown(void)
154 timer_delete(ival_timer);
157 static void posix_timer_setup(void)
161 memset(&evt, 0, sizeof(evt));
162 evt.sigev_notify = SIGEV_THREAD;
163 evt.sigev_notify_function = ival_fn;
165 if (timer_create(FIO_TIMER_CLOCK, &evt, &ival_timer) < 0)
166 perror("timer_create");
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_quit;
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 if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
384 fio_gettime(&td->tv_cache, NULL);
387 static int break_on_this_error(struct thread_data *td, int *retptr)
391 if (ret < 0 || td->error) {
394 if (!td->o.continue_on_error)
402 if (td_non_fatal_error(err)) {
404 * Continue with the I/Os in case of
407 update_error_count(td, err);
411 } else if (td->o.fill_device && err == ENOSPC) {
413 * We expect to hit this error if
414 * fill_device option is set.
421 * Stop the I/O in case of a fatal
424 update_error_count(td, err);
433 * The main verify engine. Runs over the writes we previously submitted,
434 * reads the blocks back in, and checks the crc/md5 of the data.
436 static void do_verify(struct thread_data *td)
443 dprint(FD_VERIFY, "starting loop\n");
446 * sync io first and invalidate cache, to make sure we really
449 for_each_file(td, f, i) {
450 if (!fio_file_open(f))
452 if (fio_io_sync(td, f))
454 if (file_invalidate_cache(td, f))
461 td_set_runstate(td, TD_VERIFYING);
464 while (!td->terminate) {
469 if (runtime_exceeded(td, &td->tv_cache)) {
474 io_u = __get_io_u(td);
478 if (get_next_verify(td, io_u)) {
483 if (td_io_prep(td, io_u)) {
488 if (td->o.verify_async)
489 io_u->end_io = verify_io_u_async;
491 io_u->end_io = verify_io_u;
493 ret = td_io_queue(td, io_u);
495 case FIO_Q_COMPLETED:
498 clear_io_u(td, io_u);
499 } else if (io_u->resid) {
500 int bytes = io_u->xfer_buflen - io_u->resid;
501 struct fio_file *f = io_u->file;
507 td_verror(td, EIO, "full resid");
512 io_u->xfer_buflen = io_u->resid;
513 io_u->xfer_buf += bytes;
514 io_u->offset += bytes;
516 if (ddir_rw(io_u->ddir))
517 td->ts.short_io_u[io_u->ddir]++;
519 if (io_u->offset == f->real_file_size)
522 requeue_io_u(td, &io_u);
525 ret = io_u_sync_complete(td, io_u, NULL);
533 requeue_io_u(td, &io_u);
534 ret2 = td_io_commit(td);
540 td_verror(td, -ret, "td_io_queue");
544 if (break_on_this_error(td, &ret))
548 * if we can queue more, do so. but check if there are
549 * completed io_u's first.
551 full = queue_full(td) || ret == FIO_Q_BUSY;
552 if (full || !td->o.iodepth_batch_complete) {
553 min_events = min(td->o.iodepth_batch_complete,
555 if (full && !min_events)
560 * Reap required number of io units, if any,
561 * and do the verification on them through
562 * the callback handler
564 if (io_u_queued_complete(td, min_events, NULL) < 0) {
568 } while (full && (td->cur_depth > td->o.iodepth_low));
575 min_events = td->cur_depth;
578 ret = io_u_queued_complete(td, min_events, NULL);
580 cleanup_pending_aio(td);
582 td_set_runstate(td, TD_RUNNING);
584 dprint(FD_VERIFY, "exiting loop\n");
588 * Main IO worker function. It retrieves io_u's to process and queues
589 * and reaps them, checking for rate and errors along the way.
591 static void do_io(struct thread_data *td)
596 if (in_ramp_time(td))
597 td_set_runstate(td, TD_RAMP);
599 td_set_runstate(td, TD_RUNNING);
601 while ( (td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
602 (!flist_empty(&td->trim_list)) ||
603 ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) ) {
604 struct timeval comp_time;
605 unsigned long bytes_done[2] = { 0, 0 };
615 if (runtime_exceeded(td, &td->tv_cache)) {
625 * Add verification end_io handler, if asked to verify
626 * a previously written file.
628 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
630 if (td->o.verify_async)
631 io_u->end_io = verify_io_u_async;
633 io_u->end_io = verify_io_u;
634 td_set_runstate(td, TD_VERIFYING);
635 } else if (in_ramp_time(td))
636 td_set_runstate(td, TD_RAMP);
638 td_set_runstate(td, TD_RUNNING);
640 ret = td_io_queue(td, io_u);
642 case FIO_Q_COMPLETED:
645 clear_io_u(td, io_u);
646 } else if (io_u->resid) {
647 int bytes = io_u->xfer_buflen - io_u->resid;
648 struct fio_file *f = io_u->file;
654 td_verror(td, EIO, "full resid");
659 io_u->xfer_buflen = io_u->resid;
660 io_u->xfer_buf += bytes;
661 io_u->offset += bytes;
663 if (ddir_rw(io_u->ddir))
664 td->ts.short_io_u[io_u->ddir]++;
666 if (io_u->offset == f->real_file_size)
669 requeue_io_u(td, &io_u);
672 if (__should_check_rate(td, 0) ||
673 __should_check_rate(td, 1))
674 fio_gettime(&comp_time, NULL);
676 ret = io_u_sync_complete(td, io_u, bytes_done);
683 * if the engine doesn't have a commit hook,
684 * the io_u is really queued. if it does have such
685 * a hook, it has to call io_u_queued() itself.
687 if (td->io_ops->commit == NULL)
688 io_u_queued(td, io_u);
691 requeue_io_u(td, &io_u);
692 ret2 = td_io_commit(td);
702 if (break_on_this_error(td, &ret))
706 * See if we need to complete some commands
708 full = queue_full(td) || ret == FIO_Q_BUSY;
709 if (full || !td->o.iodepth_batch_complete) {
710 min_evts = min(td->o.iodepth_batch_complete,
712 if (full && !min_evts)
715 if (__should_check_rate(td, 0) ||
716 __should_check_rate(td, 1))
717 fio_gettime(&comp_time, NULL);
720 ret = io_u_queued_complete(td, min_evts, bytes_done);
724 } while (full && (td->cur_depth > td->o.iodepth_low));
729 if (!(bytes_done[0] + bytes_done[1]))
732 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
733 if (check_min_rate(td, &comp_time, bytes_done)) {
734 if (exitall_on_terminate)
735 terminate_threads(td->groupid);
736 td_verror(td, EIO, "check_min_rate");
741 if (td->o.thinktime) {
742 unsigned long long b;
744 b = td->io_blocks[0] + td->io_blocks[1];
745 if (!(b % td->o.thinktime_blocks)) {
748 if (td->o.thinktime_spin)
749 usec_spin(td->o.thinktime_spin);
751 left = td->o.thinktime - td->o.thinktime_spin;
753 usec_sleep(td, left);
758 if (td->trim_entries)
759 printf("trim entries %ld\n", td->trim_entries);
761 if (td->o.fill_device && td->error == ENOSPC) {
770 ret = io_u_queued_complete(td, i, NULL);
772 if (should_fsync(td) && td->o.end_fsync) {
773 td_set_runstate(td, TD_FSYNCING);
775 for_each_file(td, f, i) {
776 if (!fio_file_open(f))
782 cleanup_pending_aio(td);
785 * stop job if we failed doing any IO
787 if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
791 static void cleanup_io_u(struct thread_data *td)
793 struct flist_head *entry, *n;
796 flist_for_each_safe(entry, n, &td->io_u_freelist) {
797 io_u = flist_entry(entry, struct io_u, list);
799 flist_del(&io_u->list);
806 static int init_io_u(struct thread_data *td)
810 int cl_align, i, max_units;
813 max_units = td->o.iodepth;
814 max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
815 td->orig_buffer_size = (unsigned long long) max_bs
816 * (unsigned long long) max_units;
818 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
821 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
822 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
825 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
826 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
830 if (allocate_io_mem(td))
833 if (td->o.odirect || td->o.mem_align ||
834 (td->io_ops->flags & FIO_RAWIO))
835 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
839 cl_align = os_cache_line_size();
841 for (i = 0; i < max_units; i++) {
847 if (posix_memalign(&ptr, cl_align, sizeof(*io_u))) {
848 log_err("fio: posix_memalign=%s\n", strerror(errno));
853 memset(io_u, 0, sizeof(*io_u));
854 INIT_FLIST_HEAD(&io_u->list);
855 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
857 if (!(td->io_ops->flags & FIO_NOIO)) {
858 io_u->buf = p + max_bs * i;
859 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
861 if (td_write(td) && !td->o.refill_buffers)
862 io_u_fill_buffer(td, io_u, max_bs);
863 else if (td_write(td) && td->o.verify_pattern_bytes) {
865 * Fill the buffer with the pattern if we are
866 * going to be doing writes.
868 fill_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
873 io_u->flags = IO_U_F_FREE;
874 flist_add(&io_u->list, &td->io_u_freelist);
880 static int switch_ioscheduler(struct thread_data *td)
882 char tmp[256], tmp2[128];
886 if (td->io_ops->flags & FIO_DISKLESSIO)
889 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
891 f = fopen(tmp, "r+");
893 if (errno == ENOENT) {
894 log_err("fio: os or kernel doesn't support IO scheduler"
898 td_verror(td, errno, "fopen iosched");
905 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
906 if (ferror(f) || ret != 1) {
907 td_verror(td, errno, "fwrite");
915 * Read back and check that the selected scheduler is now the default.
917 ret = fread(tmp, 1, sizeof(tmp), f);
918 if (ferror(f) || ret < 0) {
919 td_verror(td, errno, "fread");
924 sprintf(tmp2, "[%s]", td->o.ioscheduler);
925 if (!strstr(tmp, tmp2)) {
926 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
927 td_verror(td, EINVAL, "iosched_switch");
936 static int keep_running(struct thread_data *td)
938 unsigned long long io_done;
942 if (td->o.time_based)
949 io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
951 if (io_done < td->o.size)
957 static void reset_io_counters(struct thread_data *td)
959 td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
960 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
962 td->rate_bytes[0] = td->rate_bytes[1] = 0;
963 td->rate_blocks[0] = td->rate_blocks[1] = 0;
965 td->last_was_sync = 0;
968 * reset file done count if we are to start over
970 if (td->o.time_based || td->o.loops)
971 td->nr_done_files = 0;
974 * Set the same seed to get repeatable runs
976 td_fill_rand_seeds(td);
979 void reset_all_stats(struct thread_data *td)
984 reset_io_counters(td);
986 for (i = 0; i < 2; i++) {
988 td->io_blocks[i] = 0;
989 td->io_issues[i] = 0;
990 td->ts.total_io_u[i] = 0;
993 fio_gettime(&tv, NULL);
994 td->ts.runtime[0] = 0;
995 td->ts.runtime[1] = 0;
996 memcpy(&td->epoch, &tv, sizeof(tv));
997 memcpy(&td->start, &tv, sizeof(tv));
1000 static void clear_io_state(struct thread_data *td)
1005 reset_io_counters(td);
1008 for_each_file(td, f, i)
1009 fio_file_clear_done(f);
1012 static int exec_string(const char *string)
1014 int ret, newlen = strlen(string) + 1 + 8;
1017 str = malloc(newlen);
1018 sprintf(str, "sh -c %s", string);
1022 log_err("fio: exec of cmd <%s> failed\n", str);
1029 * Entry point for the thread based jobs. The process based jobs end up
1030 * here as well, after a little setup.
1032 static void *thread_main(void *data)
1034 unsigned long long elapsed;
1035 struct thread_data *td = data;
1036 pthread_condattr_t attr;
1039 if (!td->o.use_thread)
1044 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1046 INIT_FLIST_HEAD(&td->io_u_freelist);
1047 INIT_FLIST_HEAD(&td->io_u_busylist);
1048 INIT_FLIST_HEAD(&td->io_u_requeues);
1049 INIT_FLIST_HEAD(&td->io_log_list);
1050 INIT_FLIST_HEAD(&td->io_hist_list);
1051 INIT_FLIST_HEAD(&td->verify_list);
1052 INIT_FLIST_HEAD(&td->trim_list);
1053 pthread_mutex_init(&td->io_u_lock, NULL);
1054 td->io_hist_tree = RB_ROOT;
1056 pthread_condattr_init(&attr);
1057 pthread_cond_init(&td->verify_cond, &attr);
1058 pthread_cond_init(&td->free_cond, &attr);
1060 td_set_runstate(td, TD_INITIALIZED);
1061 dprint(FD_MUTEX, "up startup_mutex\n");
1062 fio_mutex_up(startup_mutex);
1063 dprint(FD_MUTEX, "wait on td->mutex\n");
1064 fio_mutex_down(td->mutex);
1065 dprint(FD_MUTEX, "done waiting on td->mutex\n");
1068 * the ->mutex mutex is now no longer used, close it to avoid
1069 * eating a file descriptor
1071 fio_mutex_remove(td->mutex);
1074 * A new gid requires privilege, so we need to do this before setting
1077 if (td->o.gid != -1U && setgid(td->o.gid)) {
1078 td_verror(td, errno, "setgid");
1081 if (td->o.uid != -1U && setuid(td->o.uid)) {
1082 td_verror(td, errno, "setuid");
1087 * May alter parameters that init_io_u() will use, so we need to
1096 if (td->o.verify_async && verify_async_init(td))
1099 if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1100 td_verror(td, errno, "cpu_set_affinity");
1105 * If we have a gettimeofday() thread, make sure we exclude that
1106 * thread from this job
1108 if (td->o.gtod_cpu) {
1109 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1110 if (fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1111 td_verror(td, errno, "cpu_set_affinity");
1116 if (td->ioprio_set) {
1117 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1118 td_verror(td, errno, "ioprio_set");
1123 if (td->o.cgroup_weight && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1126 if (nice(td->o.nice) == -1) {
1127 td_verror(td, errno, "nice");
1131 if (td->o.ioscheduler && switch_ioscheduler(td))
1134 if (!td->o.create_serialize && setup_files(td))
1140 if (init_random_map(td))
1143 if (td->o.exec_prerun) {
1144 if (exec_string(td->o.exec_prerun))
1148 if (td->o.pre_read) {
1149 if (pre_read_files(td) < 0)
1153 fio_gettime(&td->epoch, NULL);
1154 getrusage(RUSAGE_SELF, &td->ts.ru_start);
1157 while (keep_running(td)) {
1158 fio_gettime(&td->start, NULL);
1159 memcpy(&td->ts.stat_sample_time[0], &td->start,
1161 memcpy(&td->ts.stat_sample_time[1], &td->start,
1163 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1165 if (td->o.ratemin[0] || td->o.ratemin[1])
1166 memcpy(&td->lastrate, &td->ts.stat_sample_time,
1167 sizeof(td->lastrate));
1172 prune_io_piece_log(td);
1178 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1179 elapsed = utime_since_now(&td->start);
1180 td->ts.runtime[DDIR_READ] += elapsed;
1182 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1183 elapsed = utime_since_now(&td->start);
1184 td->ts.runtime[DDIR_WRITE] += elapsed;
1187 if (td->error || td->terminate)
1190 if (!td->o.do_verify ||
1191 td->o.verify == VERIFY_NONE ||
1192 (td->io_ops->flags & FIO_UNIDIR))
1197 fio_gettime(&td->start, NULL);
1201 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1203 if (td->error || td->terminate)
1207 update_rusage_stat(td);
1208 td->ts.runtime[0] = (td->ts.runtime[0] + 999) / 1000;
1209 td->ts.runtime[1] = (td->ts.runtime[1] + 999) / 1000;
1210 td->ts.total_run_time = mtime_since_now(&td->epoch);
1211 td->ts.io_bytes[0] = td->io_bytes[0];
1212 td->ts.io_bytes[1] = td->io_bytes[1];
1214 fio_mutex_down(writeout_mutex);
1215 if (td->ts.bw_log) {
1216 if (td->o.bw_log_file) {
1217 finish_log_named(td, td->ts.bw_log,
1218 td->o.bw_log_file, "bw");
1220 finish_log(td, td->ts.bw_log, "bw");
1222 if (td->ts.lat_log) {
1223 if (td->o.lat_log_file) {
1224 finish_log_named(td, td->ts.lat_log,
1225 td->o.lat_log_file, "lat");
1227 finish_log(td, td->ts.lat_log, "lat");
1229 if (td->ts.slat_log) {
1230 if (td->o.lat_log_file) {
1231 finish_log_named(td, td->ts.slat_log,
1232 td->o.lat_log_file, "slat");
1234 finish_log(td, td->ts.slat_log, "slat");
1236 if (td->ts.clat_log) {
1237 if (td->o.lat_log_file) {
1238 finish_log_named(td, td->ts.clat_log,
1239 td->o.lat_log_file, "clat");
1241 finish_log(td, td->ts.clat_log, "clat");
1243 fio_mutex_up(writeout_mutex);
1244 if (td->o.exec_postrun)
1245 exec_string(td->o.exec_postrun);
1247 if (exitall_on_terminate)
1248 terminate_threads(td->groupid);
1252 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1255 if (td->o.verify_async)
1256 verify_async_exit(td);
1258 close_and_free_files(td);
1261 cgroup_shutdown(td, &cgroup_mnt);
1263 if (td->o.cpumask_set) {
1264 int ret = fio_cpuset_exit(&td->o.cpumask);
1266 td_verror(td, ret, "fio_cpuset_exit");
1270 * do this very late, it will log file closing as well
1272 if (td->o.write_iolog_file)
1273 write_iolog_close(td);
1275 options_mem_free(td);
1276 td_set_runstate(td, TD_EXITED);
1277 return (void *) (unsigned long) td->error;
1281 * We cannot pass the td data into a forked process, so attach the td and
1282 * pass it to the thread worker.
1284 static int fork_main(int shmid, int offset)
1286 struct thread_data *td;
1289 data = shmat(shmid, NULL, 0);
1290 if (data == (void *) -1) {
1297 td = data + offset * sizeof(struct thread_data);
1298 ret = thread_main(td);
1300 return (int) (unsigned long) ret;
1304 * Run over the job map and reap the threads that have exited, if any.
1306 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1308 struct thread_data *td;
1309 int i, cputhreads, realthreads, pending, status, ret;
1312 * reap exited threads (TD_EXITED -> TD_REAPED)
1314 realthreads = pending = cputhreads = 0;
1315 for_each_td(td, i) {
1319 * ->io_ops is NULL for a thread that has closed its
1322 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1331 if (td->runstate == TD_REAPED)
1333 if (td->o.use_thread) {
1334 if (td->runstate == TD_EXITED) {
1335 td_set_runstate(td, TD_REAPED);
1342 if (td->runstate == TD_EXITED)
1346 * check if someone quit or got killed in an unusual way
1348 ret = waitpid(td->pid, &status, flags);
1350 if (errno == ECHILD) {
1351 log_err("fio: pid=%d disappeared %d\n",
1352 (int) td->pid, td->runstate);
1353 td_set_runstate(td, TD_REAPED);
1357 } else if (ret == td->pid) {
1358 if (WIFSIGNALED(status)) {
1359 int sig = WTERMSIG(status);
1362 log_err("fio: pid=%d, got signal=%d\n",
1363 (int) td->pid, sig);
1364 td_set_runstate(td, TD_REAPED);
1367 if (WIFEXITED(status)) {
1368 if (WEXITSTATUS(status) && !td->error)
1369 td->error = WEXITSTATUS(status);
1371 td_set_runstate(td, TD_REAPED);
1377 * thread is not dead, continue
1383 (*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
1384 (*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
1391 done_secs += mtime_since_now(&td->epoch) / 1000;
1394 if (*nr_running == cputhreads && !pending && realthreads)
1395 terminate_threads(TERMINATE_ALL);
1398 static void *gtod_thread_main(void *data)
1400 fio_mutex_up(startup_mutex);
1403 * As long as we have jobs around, update the clock. It would be nice
1404 * to have some way of NOT hammering that CPU with gettimeofday(),
1405 * but I'm not sure what to use outside of a simple CPU nop to relax
1406 * it - we don't want to lose precision.
1416 static int fio_start_gtod_thread(void)
1418 pthread_attr_t attr;
1421 pthread_attr_init(&attr);
1422 pthread_attr_setstacksize(&attr, PTHREAD_STACK_MIN);
1423 ret = pthread_create(>od_thread, &attr, gtod_thread_main, NULL);
1424 pthread_attr_destroy(&attr);
1426 log_err("Can't create gtod thread: %s\n", strerror(ret));
1430 ret = pthread_detach(gtod_thread);
1432 log_err("Can't detatch gtod thread: %s\n", strerror(ret));
1436 dprint(FD_MUTEX, "wait on startup_mutex\n");
1437 fio_mutex_down(startup_mutex);
1438 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1443 * Main function for kicking off and reaping jobs, as needed.
1445 static void run_threads(void)
1447 struct thread_data *td;
1448 unsigned long spent;
1449 int i, todo, nr_running, m_rate, t_rate, nr_started;
1451 if (fio_pin_memory())
1454 if (fio_gtod_offload && fio_start_gtod_thread())
1457 if (!terse_output) {
1458 log_info("Starting ");
1460 log_info("%d thread%s", nr_thread,
1461 nr_thread > 1 ? "s" : "");
1465 log_info("%d process%s", nr_process,
1466 nr_process > 1 ? "es" : "");
1474 todo = thread_number;
1477 m_rate = t_rate = 0;
1479 for_each_td(td, i) {
1480 print_status_init(td->thread_number - 1);
1482 if (!td->o.create_serialize) {
1488 * do file setup here so it happens sequentially,
1489 * we don't want X number of threads getting their
1490 * client data interspersed on disk
1492 if (setup_files(td)) {
1495 log_err("fio: pid=%d, err=%d/%s\n",
1496 (int) td->pid, td->error, td->verror);
1497 td_set_runstate(td, TD_REAPED);
1504 * for sharing to work, each job must always open
1505 * its own files. so close them, if we opened them
1508 for_each_file(td, f, i) {
1509 if (fio_file_open(f))
1510 td_io_close_file(td, f);
1520 struct thread_data *map[MAX_JOBS];
1521 struct timeval this_start;
1522 int this_jobs = 0, left;
1525 * create threads (TD_NOT_CREATED -> TD_CREATED)
1527 for_each_td(td, i) {
1528 if (td->runstate != TD_NOT_CREATED)
1532 * never got a chance to start, killed by other
1533 * thread for some reason
1535 if (td->terminate) {
1540 if (td->o.start_delay) {
1541 spent = mtime_since_genesis();
1543 if (td->o.start_delay * 1000 > spent)
1547 if (td->o.stonewall && (nr_started || nr_running)) {
1548 dprint(FD_PROCESS, "%s: stonewall wait\n",
1554 * Set state to created. Thread will transition
1555 * to TD_INITIALIZED when it's done setting up.
1557 td_set_runstate(td, TD_CREATED);
1558 map[this_jobs++] = td;
1561 if (td->o.use_thread) {
1564 dprint(FD_PROCESS, "will pthread_create\n");
1565 ret = pthread_create(&td->thread, NULL,
1568 log_err("pthread_create: %s\n",
1573 ret = pthread_detach(td->thread);
1575 log_err("pthread_detach: %s",
1579 dprint(FD_PROCESS, "will fork\n");
1582 int ret = fork_main(shm_id, i);
1585 } else if (i == fio_debug_jobno)
1586 *fio_debug_jobp = pid;
1588 dprint(FD_MUTEX, "wait on startup_mutex\n");
1589 if (fio_mutex_down_timeout(startup_mutex, 10)) {
1590 log_err("fio: job startup hung? exiting.\n");
1591 terminate_threads(TERMINATE_ALL);
1596 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1600 * Wait for the started threads to transition to
1603 fio_gettime(&this_start, NULL);
1605 while (left && !fio_abort) {
1606 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1611 for (i = 0; i < this_jobs; i++) {
1615 if (td->runstate == TD_INITIALIZED) {
1618 } else if (td->runstate >= TD_EXITED) {
1622 nr_running++; /* work-around... */
1628 log_err("fio: %d jobs failed to start\n", left);
1629 for (i = 0; i < this_jobs; i++) {
1633 kill(td->pid, SIGTERM);
1639 * start created threads (TD_INITIALIZED -> TD_RUNNING).
1641 for_each_td(td, i) {
1642 if (td->runstate != TD_INITIALIZED)
1645 if (in_ramp_time(td))
1646 td_set_runstate(td, TD_RAMP);
1648 td_set_runstate(td, TD_RUNNING);
1651 m_rate += td->o.ratemin[0] + td->o.ratemin[1];
1652 t_rate += td->o.rate[0] + td->o.rate[1];
1654 fio_mutex_up(td->mutex);
1657 reap_threads(&nr_running, &t_rate, &m_rate);
1663 while (nr_running) {
1664 reap_threads(&nr_running, &t_rate, &m_rate);
1672 int main(int argc, char *argv[])
1677 init_rand(&__fio_rand_state);
1680 * We need locale for number printing, if it isn't set then just
1681 * go with the US format.
1683 if (!getenv("LC_NUMERIC"))
1684 setlocale(LC_NUMERIC, "en_US");
1686 ps = sysconf(_SC_PAGESIZE);
1688 log_err("Failed to get page size\n");
1695 fio_keywords_init();
1697 if (parse_options(argc, argv))
1700 if (exec_profile && load_profile(exec_profile))
1707 setup_log(&agg_io_log[DDIR_READ]);
1708 setup_log(&agg_io_log[DDIR_WRITE]);
1711 startup_mutex = fio_mutex_init(0);
1712 if (startup_mutex == NULL)
1714 writeout_mutex = fio_mutex_init(1);
1715 if (writeout_mutex == NULL)
1720 posix_timer_setup();
1723 cgroup_list = smalloc(sizeof(*cgroup_list));
1724 INIT_FLIST_HEAD(cgroup_list);
1731 __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1732 __finish_log(agg_io_log[DDIR_WRITE],
1733 "agg-write_bw.log");
1737 cgroup_kill(cgroup_list);
1741 posix_timer_teardown();
1742 fio_mutex_remove(startup_mutex);
1743 fio_mutex_remove(writeout_mutex);