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 unsigned int thread_number = 0;
59 unsigned int nr_process = 0;
60 unsigned int nr_thread = 0;
63 unsigned long done_secs = 0;
66 * Just expose an empty list, if the OS does not support disk util stats
68 #ifndef FIO_HAVE_DISK_UTIL
69 FLIST_HEAD(disk_list);
72 static struct fio_mutex *startup_mutex;
73 static struct fio_mutex *writeout_mutex;
74 static volatile int fio_abort;
75 static int exit_value;
76 static pthread_t disk_util_thread;
77 static struct flist_head *cgroup_list;
78 static char *cgroup_mnt;
80 unsigned long arch_flags = 0;
82 struct io_log *agg_io_log[2];
84 #define JOB_START_TIMEOUT (5 * 1000)
86 static const char *fio_os_strings[os_nr] = {
98 static const char *fio_arch_strings[arch_nr] = {
114 const char *fio_get_os_string(int nr)
117 return fio_os_strings[nr];
122 const char *fio_get_arch_string(int nr)
125 return fio_arch_strings[nr];
130 void td_set_runstate(struct thread_data *td, int runstate)
132 if (td->runstate == runstate)
135 dprint(FD_PROCESS, "pid=%d: runstate %d -> %d\n", (int) td->pid,
136 td->runstate, runstate);
137 td->runstate = runstate;
140 void fio_terminate_threads(int group_id)
142 struct thread_data *td;
145 dprint(FD_PROCESS, "terminate group_id=%d\n", group_id);
148 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
149 dprint(FD_PROCESS, "setting terminate on %s/%d\n",
150 td->o.name, (int) td->pid);
152 td->o.start_delay = 0;
155 * if the thread is running, just let it exit
159 else if (td->runstate < TD_RAMP)
160 kill(td->pid, SIGTERM);
162 struct ioengine_ops *ops = td->io_ops;
164 if (ops && (ops->flags & FIO_SIGTERM))
165 kill(td->pid, SIGTERM);
171 static void sig_int(int sig)
175 fio_server_got_signal(sig);
177 log_info("\nfio: terminating on signal %d\n", sig);
182 fio_terminate_threads(TERMINATE_ALL);
186 static void *disk_thread_main(void *data)
188 fio_mutex_up(startup_mutex);
191 usleep(DISK_UTIL_MSEC * 1000);
197 print_thread_status();
203 static int create_disk_util_thread(void)
207 ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
209 log_err("Can't create disk util thread: %s\n", strerror(ret));
213 ret = pthread_detach(disk_util_thread);
215 log_err("Can't detatch disk util thread: %s\n", strerror(ret));
219 dprint(FD_MUTEX, "wait on startup_mutex\n");
220 fio_mutex_down(startup_mutex);
221 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
225 static void set_sig_handlers(void)
227 struct sigaction act;
229 memset(&act, 0, sizeof(act));
230 act.sa_handler = sig_int;
231 act.sa_flags = SA_RESTART;
232 sigaction(SIGINT, &act, NULL);
234 memset(&act, 0, sizeof(act));
235 act.sa_handler = sig_int;
236 act.sa_flags = SA_RESTART;
237 sigaction(SIGTERM, &act, NULL);
240 memset(&act, 0, sizeof(act));
241 act.sa_handler = sig_int;
242 act.sa_flags = SA_RESTART;
243 sigaction(SIGPIPE, &act, NULL);
248 * Check if we are above the minimum rate given.
250 static int __check_min_rate(struct thread_data *td, struct timeval *now,
253 unsigned long long bytes = 0;
254 unsigned long iops = 0;
257 unsigned int ratemin = 0;
258 unsigned int rate_iops = 0;
259 unsigned int rate_iops_min = 0;
261 assert(ddir_rw(ddir));
263 if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
267 * allow a 2 second settle period in the beginning
269 if (mtime_since(&td->start, now) < 2000)
272 iops += td->this_io_blocks[ddir];
273 bytes += td->this_io_bytes[ddir];
274 ratemin += td->o.ratemin[ddir];
275 rate_iops += td->o.rate_iops[ddir];
276 rate_iops_min += td->o.rate_iops_min[ddir];
279 * if rate blocks is set, sample is running
281 if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
282 spent = mtime_since(&td->lastrate[ddir], now);
283 if (spent < td->o.ratecycle)
286 if (td->o.rate[ddir]) {
288 * check bandwidth specified rate
290 if (bytes < td->rate_bytes[ddir]) {
291 log_err("%s: min rate %u not met\n", td->o.name,
295 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
296 if (rate < ratemin ||
297 bytes < td->rate_bytes[ddir]) {
298 log_err("%s: min rate %u not met, got"
299 " %luKB/sec\n", td->o.name,
306 * checks iops specified rate
308 if (iops < rate_iops) {
309 log_err("%s: min iops rate %u not met\n",
310 td->o.name, rate_iops);
313 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
314 if (rate < rate_iops_min ||
315 iops < td->rate_blocks[ddir]) {
316 log_err("%s: min iops rate %u not met,"
317 " got %lu\n", td->o.name,
318 rate_iops_min, rate);
324 td->rate_bytes[ddir] = bytes;
325 td->rate_blocks[ddir] = iops;
326 memcpy(&td->lastrate[ddir], now, sizeof(*now));
330 static int check_min_rate(struct thread_data *td, struct timeval *now,
331 unsigned long *bytes_done)
336 ret |= __check_min_rate(td, now, 0);
338 ret |= __check_min_rate(td, now, 1);
343 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
345 if (in_ramp_time(td))
349 if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
356 * When job exits, we can cancel the in-flight IO if we are using async
357 * io. Attempt to do so.
359 static void cleanup_pending_aio(struct thread_data *td)
361 struct flist_head *entry, *n;
366 * get immediately available events, if any
368 r = io_u_queued_complete(td, 0, NULL);
373 * now cancel remaining active events
375 if (td->io_ops->cancel) {
376 flist_for_each_safe(entry, n, &td->io_u_busylist) {
377 io_u = flist_entry(entry, struct io_u, list);
380 * if the io_u isn't in flight, then that generally
381 * means someone leaked an io_u. complain but fix
382 * it up, so we don't stall here.
384 if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
385 log_err("fio: non-busy IO on busy list\n");
388 r = td->io_ops->cancel(td, io_u);
396 r = io_u_queued_complete(td, td->cur_depth, NULL);
400 * Helper to handle the final sync of a file. Works just like the normal
401 * io path, just does everything sync.
403 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
405 struct io_u *io_u = __get_io_u(td);
411 io_u->ddir = DDIR_SYNC;
414 if (td_io_prep(td, io_u)) {
420 ret = td_io_queue(td, io_u);
422 td_verror(td, io_u->error, "td_io_queue");
425 } else if (ret == FIO_Q_QUEUED) {
426 if (io_u_queued_complete(td, 1, NULL) < 0)
428 } else if (ret == FIO_Q_COMPLETED) {
430 td_verror(td, io_u->error, "td_io_queue");
434 if (io_u_sync_complete(td, io_u, NULL) < 0)
436 } else if (ret == FIO_Q_BUSY) {
437 if (td_io_commit(td))
445 static inline void __update_tv_cache(struct thread_data *td)
447 fio_gettime(&td->tv_cache, NULL);
450 static inline void update_tv_cache(struct thread_data *td)
452 if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
453 __update_tv_cache(td);
456 static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
461 if (ret < 0 || td->error) {
469 if (!(td->o.continue_on_error & td_error_type(ddir, err)))
472 if (td_non_fatal_error(err)) {
474 * Continue with the I/Os in case of
477 update_error_count(td, err);
481 } else if (td->o.fill_device && err == ENOSPC) {
483 * We expect to hit this error if
484 * fill_device option is set.
491 * Stop the I/O in case of a fatal
494 update_error_count(td, err);
503 * The main verify engine. Runs over the writes we previously submitted,
504 * reads the blocks back in, and checks the crc/md5 of the data.
506 static void do_verify(struct thread_data *td)
513 dprint(FD_VERIFY, "starting loop\n");
516 * sync io first and invalidate cache, to make sure we really
519 for_each_file(td, f, i) {
520 if (!fio_file_open(f))
522 if (fio_io_sync(td, f))
524 if (file_invalidate_cache(td, f))
531 td_set_runstate(td, TD_VERIFYING);
534 while (!td->terminate) {
539 if (runtime_exceeded(td, &td->tv_cache)) {
540 __update_tv_cache(td);
541 if (runtime_exceeded(td, &td->tv_cache)) {
547 io_u = __get_io_u(td);
551 if (get_next_verify(td, io_u)) {
556 if (td_io_prep(td, io_u)) {
561 if (td->o.verify_async)
562 io_u->end_io = verify_io_u_async;
564 io_u->end_io = verify_io_u;
566 ret = td_io_queue(td, io_u);
568 case FIO_Q_COMPLETED:
571 clear_io_u(td, io_u);
572 } else if (io_u->resid) {
573 int bytes = io_u->xfer_buflen - io_u->resid;
579 td_verror(td, EIO, "full resid");
584 io_u->xfer_buflen = io_u->resid;
585 io_u->xfer_buf += bytes;
586 io_u->offset += bytes;
588 if (ddir_rw(io_u->ddir))
589 td->ts.short_io_u[io_u->ddir]++;
592 if (io_u->offset == f->real_file_size)
595 requeue_io_u(td, &io_u);
598 ret = io_u_sync_complete(td, io_u, NULL);
606 requeue_io_u(td, &io_u);
607 ret2 = td_io_commit(td);
613 td_verror(td, -ret, "td_io_queue");
617 if (break_on_this_error(td, io_u->ddir, &ret))
621 * if we can queue more, do so. but check if there are
622 * completed io_u's first. Note that we can get BUSY even
623 * without IO queued, if the system is resource starved.
625 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
626 if (full || !td->o.iodepth_batch_complete) {
627 min_events = min(td->o.iodepth_batch_complete,
629 if (full && !min_events && td->o.iodepth_batch_complete != 0)
634 * Reap required number of io units, if any,
635 * and do the verification on them through
636 * the callback handler
638 if (io_u_queued_complete(td, min_events, NULL) < 0) {
642 } while (full && (td->cur_depth > td->o.iodepth_low));
649 min_events = td->cur_depth;
652 ret = io_u_queued_complete(td, min_events, NULL);
654 cleanup_pending_aio(td);
656 td_set_runstate(td, TD_RUNNING);
658 dprint(FD_VERIFY, "exiting loop\n");
662 * Main IO worker function. It retrieves io_u's to process and queues
663 * and reaps them, checking for rate and errors along the way.
665 static void do_io(struct thread_data *td)
670 if (in_ramp_time(td))
671 td_set_runstate(td, TD_RAMP);
673 td_set_runstate(td, TD_RUNNING);
675 while ( (td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
676 (!flist_empty(&td->trim_list)) ||
677 ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) ) {
678 struct timeval comp_time;
679 unsigned long bytes_done[2] = { 0, 0 };
690 if (runtime_exceeded(td, &td->tv_cache)) {
691 __update_tv_cache(td);
692 if (runtime_exceeded(td, &td->tv_cache)) {
705 * Add verification end_io handler, if asked to verify
706 * a previously written file.
708 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
710 if (td->o.verify_async)
711 io_u->end_io = verify_io_u_async;
713 io_u->end_io = verify_io_u;
714 td_set_runstate(td, TD_VERIFYING);
715 } else if (in_ramp_time(td))
716 td_set_runstate(td, TD_RAMP);
718 td_set_runstate(td, TD_RUNNING);
720 ret = td_io_queue(td, io_u);
722 case FIO_Q_COMPLETED:
725 clear_io_u(td, io_u);
726 } else if (io_u->resid) {
727 int bytes = io_u->xfer_buflen - io_u->resid;
728 struct fio_file *f = io_u->file;
734 td_verror(td, EIO, "full resid");
739 io_u->xfer_buflen = io_u->resid;
740 io_u->xfer_buf += bytes;
741 io_u->offset += bytes;
743 if (ddir_rw(io_u->ddir))
744 td->ts.short_io_u[io_u->ddir]++;
746 if (io_u->offset == f->real_file_size)
749 requeue_io_u(td, &io_u);
752 if (__should_check_rate(td, 0) ||
753 __should_check_rate(td, 1))
754 fio_gettime(&comp_time, NULL);
756 ret = io_u_sync_complete(td, io_u, bytes_done);
763 * if the engine doesn't have a commit hook,
764 * the io_u is really queued. if it does have such
765 * a hook, it has to call io_u_queued() itself.
767 if (td->io_ops->commit == NULL)
768 io_u_queued(td, io_u);
771 requeue_io_u(td, &io_u);
772 ret2 = td_io_commit(td);
782 if (break_on_this_error(td, ddir, &ret))
786 * See if we need to complete some commands. Note that we
787 * can get BUSY even without IO queued, if the system is
790 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
791 if (full || !td->o.iodepth_batch_complete) {
792 min_evts = min(td->o.iodepth_batch_complete,
794 if (full && !min_evts && td->o.iodepth_batch_complete != 0)
797 if (__should_check_rate(td, 0) ||
798 __should_check_rate(td, 1))
799 fio_gettime(&comp_time, NULL);
802 ret = io_u_queued_complete(td, min_evts, bytes_done);
806 } while (full && (td->cur_depth > td->o.iodepth_low));
811 if (!(bytes_done[0] + bytes_done[1]))
814 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
815 if (check_min_rate(td, &comp_time, bytes_done)) {
816 if (exitall_on_terminate)
817 fio_terminate_threads(td->groupid);
818 td_verror(td, EIO, "check_min_rate");
823 if (td->o.thinktime) {
824 unsigned long long b;
826 b = td->io_blocks[0] + td->io_blocks[1];
827 if (!(b % td->o.thinktime_blocks)) {
830 if (td->o.thinktime_spin)
831 usec_spin(td->o.thinktime_spin);
833 left = td->o.thinktime - td->o.thinktime_spin;
835 usec_sleep(td, left);
840 if (td->trim_entries)
841 log_err("fio: %d trim entries leaked?\n", td->trim_entries);
843 if (td->o.fill_device && td->error == ENOSPC) {
852 ret = io_u_queued_complete(td, i, NULL);
853 if (td->o.fill_device && td->error == ENOSPC)
857 if (should_fsync(td) && td->o.end_fsync) {
858 td_set_runstate(td, TD_FSYNCING);
860 for_each_file(td, f, i) {
861 if (!fio_file_open(f))
867 cleanup_pending_aio(td);
870 * stop job if we failed doing any IO
872 if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
876 static void cleanup_io_u(struct thread_data *td)
878 struct flist_head *entry, *n;
881 flist_for_each_safe(entry, n, &td->io_u_freelist) {
882 io_u = flist_entry(entry, struct io_u, list);
884 flist_del(&io_u->list);
885 fio_memfree(io_u, sizeof(*io_u));
891 static int init_io_u(struct thread_data *td)
895 int cl_align, i, max_units;
898 max_units = td->o.iodepth;
899 max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
900 td->orig_buffer_size = (unsigned long long) max_bs
901 * (unsigned long long) max_units;
903 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
906 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
907 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
910 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
911 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
915 if (allocate_io_mem(td))
918 if (td->o.odirect || td->o.mem_align ||
919 (td->io_ops->flags & FIO_RAWIO))
920 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
924 cl_align = os_cache_line_size();
926 for (i = 0; i < max_units; i++) {
932 ptr = fio_memalign(cl_align, sizeof(*io_u));
934 log_err("fio: unable to allocate aligned memory\n");
939 memset(io_u, 0, sizeof(*io_u));
940 INIT_FLIST_HEAD(&io_u->list);
941 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
943 if (!(td->io_ops->flags & FIO_NOIO)) {
945 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
948 io_u_fill_buffer(td, io_u, max_bs);
949 if (td_write(td) && td->o.verify_pattern_bytes) {
951 * Fill the buffer with the pattern if we are
952 * going to be doing writes.
954 fill_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
959 io_u->flags = IO_U_F_FREE;
960 flist_add(&io_u->list, &td->io_u_freelist);
967 static int switch_ioscheduler(struct thread_data *td)
969 char tmp[256], tmp2[128];
973 if (td->io_ops->flags & FIO_DISKLESSIO)
976 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
978 f = fopen(tmp, "r+");
980 if (errno == ENOENT) {
981 log_err("fio: os or kernel doesn't support IO scheduler"
985 td_verror(td, errno, "fopen iosched");
992 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
993 if (ferror(f) || ret != 1) {
994 td_verror(td, errno, "fwrite");
1002 * Read back and check that the selected scheduler is now the default.
1004 ret = fread(tmp, 1, sizeof(tmp), f);
1005 if (ferror(f) || ret < 0) {
1006 td_verror(td, errno, "fread");
1011 sprintf(tmp2, "[%s]", td->o.ioscheduler);
1012 if (!strstr(tmp, tmp2)) {
1013 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1014 td_verror(td, EINVAL, "iosched_switch");
1023 static int keep_running(struct thread_data *td)
1025 unsigned long long io_done;
1029 if (td->o.time_based)
1036 io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
1037 + td->io_skip_bytes;
1038 if (io_done < td->o.size)
1044 static void reset_io_counters(struct thread_data *td)
1046 td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0;
1047 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
1048 td->stat_io_blocks[0] = td->stat_io_blocks[1] = 0;
1049 td->this_io_blocks[0] = td->this_io_blocks[1] = 0;
1051 td->rate_bytes[0] = td->rate_bytes[1] = 0;
1052 td->rate_blocks[0] = td->rate_blocks[1] = 0;
1054 td->last_was_sync = 0;
1057 * reset file done count if we are to start over
1059 if (td->o.time_based || td->o.loops)
1060 td->nr_done_files = 0;
1063 void reset_all_stats(struct thread_data *td)
1068 reset_io_counters(td);
1070 for (i = 0; i < 2; i++) {
1071 td->io_bytes[i] = 0;
1072 td->io_blocks[i] = 0;
1073 td->io_issues[i] = 0;
1074 td->ts.total_io_u[i] = 0;
1077 fio_gettime(&tv, NULL);
1078 td->ts.runtime[0] = 0;
1079 td->ts.runtime[1] = 0;
1080 memcpy(&td->epoch, &tv, sizeof(tv));
1081 memcpy(&td->start, &tv, sizeof(tv));
1084 static void clear_io_state(struct thread_data *td)
1089 reset_io_counters(td);
1092 for_each_file(td, f, i)
1093 fio_file_clear_done(f);
1096 * Set the same seed to get repeatable runs
1098 td_fill_rand_seeds(td);
1101 static int exec_string(const char *string)
1103 int ret, newlen = strlen(string) + 1 + 8;
1106 str = malloc(newlen);
1107 sprintf(str, "sh -c %s", string);
1111 log_err("fio: exec of cmd <%s> failed\n", str);
1118 * Entry point for the thread based jobs. The process based jobs end up
1119 * here as well, after a little setup.
1121 static void *thread_main(void *data)
1123 unsigned long long elapsed;
1124 struct thread_data *td = data;
1125 pthread_condattr_t attr;
1128 if (!td->o.use_thread) {
1134 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1136 INIT_FLIST_HEAD(&td->io_u_freelist);
1137 INIT_FLIST_HEAD(&td->io_u_busylist);
1138 INIT_FLIST_HEAD(&td->io_u_requeues);
1139 INIT_FLIST_HEAD(&td->io_log_list);
1140 INIT_FLIST_HEAD(&td->io_hist_list);
1141 INIT_FLIST_HEAD(&td->verify_list);
1142 INIT_FLIST_HEAD(&td->trim_list);
1143 pthread_mutex_init(&td->io_u_lock, NULL);
1144 td->io_hist_tree = RB_ROOT;
1146 pthread_condattr_init(&attr);
1147 pthread_cond_init(&td->verify_cond, &attr);
1148 pthread_cond_init(&td->free_cond, &attr);
1150 td_set_runstate(td, TD_INITIALIZED);
1151 dprint(FD_MUTEX, "up startup_mutex\n");
1152 fio_mutex_up(startup_mutex);
1153 dprint(FD_MUTEX, "wait on td->mutex\n");
1154 fio_mutex_down(td->mutex);
1155 dprint(FD_MUTEX, "done waiting on td->mutex\n");
1158 * the ->mutex mutex is now no longer used, close it to avoid
1159 * eating a file descriptor
1161 fio_mutex_remove(td->mutex);
1164 * A new gid requires privilege, so we need to do this before setting
1167 if (td->o.gid != -1U && setgid(td->o.gid)) {
1168 td_verror(td, errno, "setgid");
1171 if (td->o.uid != -1U && setuid(td->o.uid)) {
1172 td_verror(td, errno, "setuid");
1177 * If we have a gettimeofday() thread, make sure we exclude that
1178 * thread from this job
1181 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1184 * Set affinity first, in case it has an impact on the memory
1187 if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1188 td_verror(td, errno, "cpu_set_affinity");
1193 * May alter parameters that init_io_u() will use, so we need to
1202 if (td->o.verify_async && verify_async_init(td))
1205 if (td->ioprio_set) {
1206 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1207 td_verror(td, errno, "ioprio_set");
1212 if (td->o.cgroup_weight && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1215 if (nice(td->o.nice) == -1) {
1216 td_verror(td, errno, "nice");
1220 if (td->o.ioscheduler && switch_ioscheduler(td))
1223 if (!td->o.create_serialize && setup_files(td))
1229 if (init_random_map(td))
1232 if (td->o.exec_prerun) {
1233 if (exec_string(td->o.exec_prerun))
1237 if (td->o.pre_read) {
1238 if (pre_read_files(td) < 0)
1242 fio_gettime(&td->epoch, NULL);
1243 getrusage(RUSAGE_SELF, &td->ru_start);
1246 while (keep_running(td)) {
1247 fio_gettime(&td->start, NULL);
1248 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1249 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1250 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1252 if (td->o.ratemin[0] || td->o.ratemin[1]) {
1253 memcpy(&td->lastrate[0], &td->bw_sample_time,
1254 sizeof(td->bw_sample_time));
1255 memcpy(&td->lastrate[1], &td->bw_sample_time,
1256 sizeof(td->bw_sample_time));
1262 prune_io_piece_log(td);
1268 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1269 elapsed = utime_since_now(&td->start);
1270 td->ts.runtime[DDIR_READ] += elapsed;
1272 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1273 elapsed = utime_since_now(&td->start);
1274 td->ts.runtime[DDIR_WRITE] += elapsed;
1277 if (td->error || td->terminate)
1280 if (!td->o.do_verify ||
1281 td->o.verify == VERIFY_NONE ||
1282 (td->io_ops->flags & FIO_UNIDIR))
1287 fio_gettime(&td->start, NULL);
1291 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1293 if (td->error || td->terminate)
1297 update_rusage_stat(td);
1298 td->ts.runtime[0] = (td->ts.runtime[0] + 999) / 1000;
1299 td->ts.runtime[1] = (td->ts.runtime[1] + 999) / 1000;
1300 td->ts.total_run_time = mtime_since_now(&td->epoch);
1301 td->ts.io_bytes[0] = td->io_bytes[0];
1302 td->ts.io_bytes[1] = td->io_bytes[1];
1304 fio_mutex_down(writeout_mutex);
1306 if (td->o.bw_log_file) {
1307 finish_log_named(td, td->bw_log,
1308 td->o.bw_log_file, "bw");
1310 finish_log(td, td->bw_log, "bw");
1313 if (td->o.lat_log_file) {
1314 finish_log_named(td, td->lat_log,
1315 td->o.lat_log_file, "lat");
1317 finish_log(td, td->lat_log, "lat");
1320 if (td->o.lat_log_file) {
1321 finish_log_named(td, td->slat_log,
1322 td->o.lat_log_file, "slat");
1324 finish_log(td, td->slat_log, "slat");
1327 if (td->o.lat_log_file) {
1328 finish_log_named(td, td->clat_log,
1329 td->o.lat_log_file, "clat");
1331 finish_log(td, td->clat_log, "clat");
1334 if (td->o.iops_log_file) {
1335 finish_log_named(td, td->iops_log,
1336 td->o.iops_log_file, "iops");
1338 finish_log(td, td->iops_log, "iops");
1341 fio_mutex_up(writeout_mutex);
1342 if (td->o.exec_postrun)
1343 exec_string(td->o.exec_postrun);
1345 if (exitall_on_terminate)
1346 fio_terminate_threads(td->groupid);
1350 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1353 if (td->o.verify_async)
1354 verify_async_exit(td);
1356 close_and_free_files(td);
1359 cgroup_shutdown(td, &cgroup_mnt);
1361 if (td->o.cpumask_set) {
1362 int ret = fio_cpuset_exit(&td->o.cpumask);
1364 td_verror(td, ret, "fio_cpuset_exit");
1368 * do this very late, it will log file closing as well
1370 if (td->o.write_iolog_file)
1371 write_iolog_close(td);
1373 td_set_runstate(td, TD_EXITED);
1374 return (void *) (unsigned long) td->error;
1378 * We cannot pass the td data into a forked process, so attach the td and
1379 * pass it to the thread worker.
1381 static int fork_main(int shmid, int offset)
1383 struct thread_data *td;
1387 data = shmat(shmid, NULL, 0);
1388 if (data == (void *) -1) {
1396 * HP-UX inherits shm mappings?
1401 td = data + offset * sizeof(struct thread_data);
1402 ret = thread_main(td);
1404 return (int) (unsigned long) ret;
1408 * Run over the job map and reap the threads that have exited, if any.
1410 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1411 unsigned int *m_rate)
1413 struct thread_data *td;
1414 unsigned int cputhreads, realthreads, pending;
1418 * reap exited threads (TD_EXITED -> TD_REAPED)
1420 realthreads = pending = cputhreads = 0;
1421 for_each_td(td, i) {
1425 * ->io_ops is NULL for a thread that has closed its
1428 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1437 if (td->runstate == TD_REAPED)
1439 if (td->o.use_thread) {
1440 if (td->runstate == TD_EXITED) {
1441 td_set_runstate(td, TD_REAPED);
1448 if (td->runstate == TD_EXITED)
1452 * check if someone quit or got killed in an unusual way
1454 ret = waitpid(td->pid, &status, flags);
1456 if (errno == ECHILD) {
1457 log_err("fio: pid=%d disappeared %d\n",
1458 (int) td->pid, td->runstate);
1459 td_set_runstate(td, TD_REAPED);
1463 } else if (ret == td->pid) {
1464 if (WIFSIGNALED(status)) {
1465 int sig = WTERMSIG(status);
1468 log_err("fio: pid=%d, got signal=%d\n",
1469 (int) td->pid, sig);
1470 td_set_runstate(td, TD_REAPED);
1473 if (WIFEXITED(status)) {
1474 if (WEXITSTATUS(status) && !td->error)
1475 td->error = WEXITSTATUS(status);
1477 td_set_runstate(td, TD_REAPED);
1483 * thread is not dead, continue
1489 (*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
1490 (*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
1497 done_secs += mtime_since_now(&td->epoch) / 1000;
1500 if (*nr_running == cputhreads && !pending && realthreads)
1501 fio_terminate_threads(TERMINATE_ALL);
1505 * Main function for kicking off and reaping jobs, as needed.
1507 static void run_threads(void)
1509 struct thread_data *td;
1510 unsigned long spent;
1511 unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
1513 if (fio_pin_memory())
1516 if (fio_gtod_offload && fio_start_gtod_thread())
1521 if (!terse_output) {
1522 log_info("Starting ");
1524 log_info("%d thread%s", nr_thread,
1525 nr_thread > 1 ? "s" : "");
1529 log_info("%d process%s", nr_process,
1530 nr_process > 1 ? "es" : "");
1536 todo = thread_number;
1539 m_rate = t_rate = 0;
1541 for_each_td(td, i) {
1542 print_status_init(td->thread_number - 1);
1544 if (!td->o.create_serialize)
1548 * do file setup here so it happens sequentially,
1549 * we don't want X number of threads getting their
1550 * client data interspersed on disk
1552 if (setup_files(td)) {
1555 log_err("fio: pid=%d, err=%d/%s\n",
1556 (int) td->pid, td->error, td->verror);
1557 td_set_runstate(td, TD_REAPED);
1564 * for sharing to work, each job must always open
1565 * its own files. so close them, if we opened them
1568 for_each_file(td, f, j) {
1569 if (fio_file_open(f))
1570 td_io_close_file(td, f);
1578 struct thread_data *map[REAL_MAX_JOBS];
1579 struct timeval this_start;
1580 int this_jobs = 0, left;
1583 * create threads (TD_NOT_CREATED -> TD_CREATED)
1585 for_each_td(td, i) {
1586 if (td->runstate != TD_NOT_CREATED)
1590 * never got a chance to start, killed by other
1591 * thread for some reason
1593 if (td->terminate) {
1598 if (td->o.start_delay) {
1599 spent = mtime_since_genesis();
1601 if (td->o.start_delay * 1000 > spent)
1605 if (td->o.stonewall && (nr_started || nr_running)) {
1606 dprint(FD_PROCESS, "%s: stonewall wait\n",
1614 * Set state to created. Thread will transition
1615 * to TD_INITIALIZED when it's done setting up.
1617 td_set_runstate(td, TD_CREATED);
1618 map[this_jobs++] = td;
1621 if (td->o.use_thread) {
1624 dprint(FD_PROCESS, "will pthread_create\n");
1625 ret = pthread_create(&td->thread, NULL,
1628 log_err("pthread_create: %s\n",
1633 ret = pthread_detach(td->thread);
1635 log_err("pthread_detach: %s",
1639 dprint(FD_PROCESS, "will fork\n");
1642 int ret = fork_main(shm_id, i);
1645 } else if (i == fio_debug_jobno)
1646 *fio_debug_jobp = pid;
1648 dprint(FD_MUTEX, "wait on startup_mutex\n");
1649 if (fio_mutex_down_timeout(startup_mutex, 10)) {
1650 log_err("fio: job startup hung? exiting.\n");
1651 fio_terminate_threads(TERMINATE_ALL);
1656 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1660 * Wait for the started threads to transition to
1663 fio_gettime(&this_start, NULL);
1665 while (left && !fio_abort) {
1666 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1671 for (i = 0; i < this_jobs; i++) {
1675 if (td->runstate == TD_INITIALIZED) {
1678 } else if (td->runstate >= TD_EXITED) {
1682 nr_running++; /* work-around... */
1688 log_err("fio: %d jobs failed to start\n", left);
1689 for (i = 0; i < this_jobs; i++) {
1693 kill(td->pid, SIGTERM);
1699 * start created threads (TD_INITIALIZED -> TD_RUNNING).
1701 for_each_td(td, i) {
1702 if (td->runstate != TD_INITIALIZED)
1705 if (in_ramp_time(td))
1706 td_set_runstate(td, TD_RAMP);
1708 td_set_runstate(td, TD_RUNNING);
1711 m_rate += td->o.ratemin[0] + td->o.ratemin[1];
1712 t_rate += td->o.rate[0] + td->o.rate[1];
1714 fio_mutex_up(td->mutex);
1717 reap_threads(&nr_running, &t_rate, &m_rate);
1721 fio_server_idle_loop();
1727 while (nr_running) {
1728 reap_threads(&nr_running, &t_rate, &m_rate);
1731 fio_server_idle_loop();
1742 struct thread_data *td;
1746 return fio_handle_clients();
1748 if (load_profile(exec_profile))
1751 exec_profile = NULL;
1757 setup_log(&agg_io_log[DDIR_READ], 0);
1758 setup_log(&agg_io_log[DDIR_WRITE], 0);
1761 startup_mutex = fio_mutex_init(0);
1762 if (startup_mutex == NULL)
1764 writeout_mutex = fio_mutex_init(1);
1765 if (writeout_mutex == NULL)
1769 create_disk_util_thread();
1771 cgroup_list = smalloc(sizeof(*cgroup_list));
1772 INIT_FLIST_HEAD(cgroup_list);
1779 __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1780 __finish_log(agg_io_log[DDIR_WRITE],
1781 "agg-write_bw.log");
1786 fio_options_free(td);
1788 cgroup_kill(cgroup_list);
1792 fio_mutex_remove(startup_mutex);
1793 fio_mutex_remove(writeout_mutex);
1797 void reset_fio_state(void)
1806 static int endian_check(void)
1817 else if (u.c[0] == 0x12)
1820 #if defined(FIO_LITTLE_ENDIAN)
1823 #elif defined(FIO_BIG_ENDIAN)
1836 int main(int argc, char *argv[], char *envp[])
1840 if (endian_check()) {
1841 log_err("fio: endianness settings appear wrong.\n");
1842 log_err("fio: please report this to fio@vger.kernel.org\n");
1851 * We need locale for number printing, if it isn't set then just
1852 * go with the US format.
1854 if (!getenv("LC_NUMERIC"))
1855 setlocale(LC_NUMERIC, "en_US");
1857 ps = sysconf(_SC_PAGESIZE);
1859 log_err("Failed to get page size\n");
1866 fio_keywords_init();
1868 if (parse_options(argc, argv))