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
43 unsigned long page_mask;
44 unsigned long page_size;
46 (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
49 int thread_number = 0;
54 unsigned long done_secs = 0;
56 static struct fio_mutex *startup_mutex;
57 static struct fio_mutex *writeout_mutex;
58 static volatile int fio_abort;
59 static int exit_value;
60 static struct itimerval itimer;
61 static pthread_t gtod_thread;
63 struct io_log *agg_io_log[2];
65 #define TERMINATE_ALL (-1)
66 #define JOB_START_TIMEOUT (5 * 1000)
68 void td_set_runstate(struct thread_data *td, int runstate)
70 if (td->runstate == runstate)
73 dprint(FD_PROCESS, "pid=%d: runstate %d -> %d\n", (int) td->pid,
74 td->runstate, runstate);
75 td->runstate = runstate;
78 static void terminate_threads(int group_id)
80 struct thread_data *td;
83 dprint(FD_PROCESS, "terminate group_id=%d\n", group_id);
86 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
87 dprint(FD_PROCESS, "setting terminate on %s/%d\n",
88 td->o.name, (int) td->pid);
90 td->o.start_delay = 0;
93 * if the thread is running, just let it exit
95 if (td->runstate < TD_RUNNING)
96 kill(td->pid, SIGQUIT);
98 struct ioengine_ops *ops = td->io_ops;
100 if (ops && (ops->flags & FIO_SIGQUIT))
101 kill(td->pid, SIGQUIT);
107 static void status_timer_arm(void)
109 itimer.it_value.tv_sec = 0;
110 itimer.it_value.tv_usec = DISK_UTIL_MSEC * 1000;
111 setitimer(ITIMER_REAL, &itimer, NULL);
114 static void sig_alrm(int fio_unused sig)
118 print_thread_status();
124 * Happens on thread runs with ctrl-c, ignore our own SIGQUIT
126 static void sig_quit(int sig)
130 static void sig_int(int sig)
133 printf("\nfio: terminating on signal %d\n", sig);
135 terminate_threads(TERMINATE_ALL);
139 static void sig_ill(int fio_unused sig)
144 log_err("fio: illegal instruction. your cpu does not support "
145 "the sse4.2 instruction for crc32c\n");
146 terminate_threads(TERMINATE_ALL);
150 static void set_sig_handlers(void)
152 struct sigaction act;
154 memset(&act, 0, sizeof(act));
155 act.sa_handler = sig_alrm;
156 act.sa_flags = SA_RESTART;
157 sigaction(SIGALRM, &act, NULL);
159 memset(&act, 0, sizeof(act));
160 act.sa_handler = sig_int;
161 act.sa_flags = SA_RESTART;
162 sigaction(SIGINT, &act, NULL);
164 memset(&act, 0, sizeof(act));
165 act.sa_handler = sig_ill;
166 act.sa_flags = SA_RESTART;
167 sigaction(SIGILL, &act, NULL);
169 memset(&act, 0, sizeof(act));
170 act.sa_handler = sig_quit;
171 act.sa_flags = SA_RESTART;
172 sigaction(SIGQUIT, &act, NULL);
176 * Check if we are above the minimum rate given.
178 static int __check_min_rate(struct thread_data *td, struct timeval *now,
181 unsigned long long bytes = 0;
182 unsigned long iops = 0;
185 unsigned int ratemin = 0;
186 unsigned int rate_iops = 0;
187 unsigned int rate_iops_min = 0;
189 if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
193 * allow a 2 second settle period in the beginning
195 if (mtime_since(&td->start, now) < 2000)
198 iops += td->io_blocks[ddir];
199 bytes += td->this_io_bytes[ddir];
200 ratemin += td->o.ratemin[ddir];
201 rate_iops += td->o.rate_iops[ddir];
202 rate_iops_min += td->o.rate_iops_min[ddir];
205 * if rate blocks is set, sample is running
207 if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
208 spent = mtime_since(&td->lastrate[ddir], now);
209 if (spent < td->o.ratecycle)
212 if (td->o.rate[ddir]) {
214 * check bandwidth specified rate
216 if (bytes < td->rate_bytes[ddir]) {
217 log_err("%s: min rate %u not met\n", td->o.name,
221 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
222 if (rate < ratemin ||
223 bytes < td->rate_bytes[ddir]) {
224 log_err("%s: min rate %u not met, got"
225 " %luKiB/sec\n", td->o.name,
232 * checks iops specified rate
234 if (iops < rate_iops) {
235 log_err("%s: min iops rate %u not met\n",
236 td->o.name, rate_iops);
239 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
240 if (rate < rate_iops_min ||
241 iops < td->rate_blocks[ddir]) {
242 log_err("%s: min iops rate %u not met,"
243 " got %lu\n", td->o.name,
244 rate_iops_min, rate);
250 td->rate_bytes[ddir] = bytes;
251 td->rate_blocks[ddir] = iops;
252 memcpy(&td->lastrate[ddir], now, sizeof(*now));
256 static int check_min_rate(struct thread_data *td, struct timeval *now,
257 unsigned long *bytes_done)
262 ret |= __check_min_rate(td, now, 0);
264 ret |= __check_min_rate(td, now, 1);
269 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
273 if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
280 * When job exits, we can cancel the in-flight IO if we are using async
281 * io. Attempt to do so.
283 static void cleanup_pending_aio(struct thread_data *td)
285 struct flist_head *entry, *n;
290 * get immediately available events, if any
292 r = io_u_queued_complete(td, 0, NULL);
297 * now cancel remaining active events
299 if (td->io_ops->cancel) {
300 flist_for_each_safe(entry, n, &td->io_u_busylist) {
301 io_u = flist_entry(entry, struct io_u, list);
304 * if the io_u isn't in flight, then that generally
305 * means someone leaked an io_u. complain but fix
306 * it up, so we don't stall here.
308 if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
309 log_err("fio: non-busy IO on busy list\n");
312 r = td->io_ops->cancel(td, io_u);
320 r = io_u_queued_complete(td, td->cur_depth, NULL);
324 * Helper to handle the final sync of a file. Works just like the normal
325 * io path, just does everything sync.
327 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
329 struct io_u *io_u = __get_io_u(td);
335 io_u->ddir = DDIR_SYNC;
338 if (td_io_prep(td, io_u)) {
344 ret = td_io_queue(td, io_u);
346 td_verror(td, io_u->error, "td_io_queue");
349 } else if (ret == FIO_Q_QUEUED) {
350 if (io_u_queued_complete(td, 1, NULL) < 0)
352 } else if (ret == FIO_Q_COMPLETED) {
354 td_verror(td, io_u->error, "td_io_queue");
358 if (io_u_sync_complete(td, io_u, NULL) < 0)
360 } else if (ret == FIO_Q_BUSY) {
361 if (td_io_commit(td))
369 static inline void update_tv_cache(struct thread_data *td)
371 if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
372 fio_gettime(&td->tv_cache, NULL);
376 * The main verify engine. Runs over the writes we previously submitted,
377 * reads the blocks back in, and checks the crc/md5 of the data.
379 static void do_verify(struct thread_data *td)
387 * sync io first and invalidate cache, to make sure we really
390 for_each_file(td, f, i) {
391 if (!fio_file_open(f))
393 if (fio_io_sync(td, f))
395 if (file_invalidate_cache(td, f))
402 td_set_runstate(td, TD_VERIFYING);
405 while (!td->terminate) {
408 io_u = __get_io_u(td);
414 if (runtime_exceeded(td, &td->tv_cache)) {
420 if (get_next_verify(td, io_u)) {
425 if (td_io_prep(td, io_u)) {
430 io_u->end_io = verify_io_u;
432 ret = td_io_queue(td, io_u);
434 case FIO_Q_COMPLETED:
437 else if (io_u->resid) {
438 int bytes = io_u->xfer_buflen - io_u->resid;
439 struct fio_file *f = io_u->file;
445 td_verror(td, EIO, "full resid");
450 io_u->xfer_buflen = io_u->resid;
451 io_u->xfer_buf += bytes;
452 io_u->offset += bytes;
454 td->ts.short_io_u[io_u->ddir]++;
456 if (io_u->offset == f->real_file_size)
459 requeue_io_u(td, &io_u);
462 ret = io_u_sync_complete(td, io_u, NULL);
470 requeue_io_u(td, &io_u);
471 ret2 = td_io_commit(td);
477 td_verror(td, -ret, "td_io_queue");
481 if (ret < 0 || td->error)
485 * if we can queue more, do so. but check if there are
486 * completed io_u's first.
488 full = queue_full(td) || ret == FIO_Q_BUSY;
489 if (full || !td->o.iodepth_batch_complete) {
490 min_events = td->o.iodepth_batch_complete;
491 if (full && !min_events)
496 * Reap required number of io units, if any,
497 * and do the verification on them through
498 * the callback handler
500 if (io_u_queued_complete(td, min_events, NULL) < 0) {
504 } while (full && (td->cur_depth > td->o.iodepth_low));
511 min_events = td->cur_depth;
514 ret = io_u_queued_complete(td, min_events, NULL);
516 cleanup_pending_aio(td);
518 td_set_runstate(td, TD_RUNNING);
522 * Main IO worker function. It retrieves io_u's to process and queues
523 * and reaps them, checking for rate and errors along the way.
525 static void do_io(struct thread_data *td)
530 if (in_ramp_time(td))
531 td_set_runstate(td, TD_RAMP);
533 td_set_runstate(td, TD_RUNNING);
535 while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) {
536 struct timeval comp_time;
537 unsigned long bytes_done[2] = { 0, 0 };
551 if (runtime_exceeded(td, &td->tv_cache)) {
558 * Add verification end_io handler, if asked to verify
559 * a previously written file.
561 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ) {
562 io_u->end_io = verify_io_u;
563 td_set_runstate(td, TD_VERIFYING);
564 } else if (in_ramp_time(td))
565 td_set_runstate(td, TD_RAMP);
567 td_set_runstate(td, TD_RUNNING);
569 ret = td_io_queue(td, io_u);
571 case FIO_Q_COMPLETED:
574 else if (io_u->resid) {
575 int bytes = io_u->xfer_buflen - io_u->resid;
576 struct fio_file *f = io_u->file;
582 td_verror(td, EIO, "full resid");
587 io_u->xfer_buflen = io_u->resid;
588 io_u->xfer_buf += bytes;
589 io_u->offset += bytes;
591 td->ts.short_io_u[io_u->ddir]++;
593 if (io_u->offset == f->real_file_size)
596 requeue_io_u(td, &io_u);
599 if (__should_check_rate(td, 0) ||
600 __should_check_rate(td, 1))
601 fio_gettime(&comp_time, NULL);
603 ret = io_u_sync_complete(td, io_u, bytes_done);
610 * if the engine doesn't have a commit hook,
611 * the io_u is really queued. if it does have such
612 * a hook, it has to call io_u_queued() itself.
614 if (td->io_ops->commit == NULL)
615 io_u_queued(td, io_u);
618 requeue_io_u(td, &io_u);
619 ret2 = td_io_commit(td);
629 if (ret < 0 || td->error)
633 * See if we need to complete some commands
635 full = queue_full(td) || ret == FIO_Q_BUSY;
636 if (full || !td->o.iodepth_batch_complete) {
637 min_evts = td->o.iodepth_batch_complete;
638 if (full && !min_evts)
641 if (__should_check_rate(td, 0) ||
642 __should_check_rate(td, 1))
643 fio_gettime(&comp_time, NULL);
646 ret = io_u_queued_complete(td, min_evts, bytes_done);
650 } while (full && (td->cur_depth > td->o.iodepth_low));
655 if (!(bytes_done[0] + bytes_done[1]))
658 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
659 if (check_min_rate(td, &comp_time, bytes_done)) {
660 if (exitall_on_terminate)
661 terminate_threads(td->groupid);
662 td_verror(td, EIO, "check_min_rate");
667 if (td->o.thinktime) {
668 unsigned long long b;
670 b = td->io_blocks[0] + td->io_blocks[1];
671 if (!(b % td->o.thinktime_blocks)) {
674 if (td->o.thinktime_spin)
675 usec_spin(td->o.thinktime_spin);
677 left = td->o.thinktime - td->o.thinktime_spin;
679 usec_sleep(td, left);
684 if (td->o.fill_device && td->error == ENOSPC) {
693 ret = io_u_queued_complete(td, i, NULL);
695 if (should_fsync(td) && td->o.end_fsync) {
696 td_set_runstate(td, TD_FSYNCING);
698 for_each_file(td, f, i) {
699 if (!fio_file_open(f))
705 cleanup_pending_aio(td);
708 * stop job if we failed doing any IO
710 if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
714 static void cleanup_io_u(struct thread_data *td)
716 struct flist_head *entry, *n;
719 flist_for_each_safe(entry, n, &td->io_u_freelist) {
720 io_u = flist_entry(entry, struct io_u, list);
722 flist_del(&io_u->list);
729 static int init_io_u(struct thread_data *td)
733 int cl_align, i, max_units;
736 max_units = td->o.iodepth;
737 max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
738 td->orig_buffer_size = (unsigned long long) max_bs
739 * (unsigned long long) max_units;
741 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
744 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
745 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
748 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
749 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
753 if (allocate_io_mem(td))
757 p = ALIGN(td->orig_buffer);
761 cl_align = os_cache_line_size();
763 for (i = 0; i < max_units; i++) {
769 if (posix_memalign(&ptr, cl_align, sizeof(*io_u))) {
770 log_err("fio: posix_memalign=%s\n", strerror(errno));
775 memset(io_u, 0, sizeof(*io_u));
776 INIT_FLIST_HEAD(&io_u->list);
778 if (!(td->io_ops->flags & FIO_NOIO)) {
779 io_u->buf = p + max_bs * i;
781 if (td_write(td) && !td->o.refill_buffers)
782 io_u_fill_buffer(td, io_u, max_bs);
786 io_u->flags = IO_U_F_FREE;
787 flist_add(&io_u->list, &td->io_u_freelist);
793 static int switch_ioscheduler(struct thread_data *td)
795 char tmp[256], tmp2[128];
799 if (td->io_ops->flags & FIO_DISKLESSIO)
802 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
804 f = fopen(tmp, "r+");
806 if (errno == ENOENT) {
807 log_err("fio: os or kernel doesn't support IO scheduler"
811 td_verror(td, errno, "fopen iosched");
818 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
819 if (ferror(f) || ret != 1) {
820 td_verror(td, errno, "fwrite");
828 * Read back and check that the selected scheduler is now the default.
830 ret = fread(tmp, 1, sizeof(tmp), f);
831 if (ferror(f) || ret < 0) {
832 td_verror(td, errno, "fread");
837 sprintf(tmp2, "[%s]", td->o.ioscheduler);
838 if (!strstr(tmp, tmp2)) {
839 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
840 td_verror(td, EINVAL, "iosched_switch");
849 static int keep_running(struct thread_data *td)
851 unsigned long long io_done;
855 if (td->o.time_based)
862 io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
864 if (io_done < td->o.size)
870 static void reset_io_counters(struct thread_data *td)
872 td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
873 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
875 td->rate_bytes[0] = td->rate_bytes[1] = 0;
876 td->rate_blocks[0] = td->rate_blocks[1] = 0;
878 td->last_was_sync = 0;
881 * reset file done count if we are to start over
883 if (td->o.time_based || td->o.loops)
884 td->nr_done_files = 0;
887 * Set the same seed to get repeatable runs
889 td_fill_rand_seeds(td);
892 void reset_all_stats(struct thread_data *td)
897 reset_io_counters(td);
899 for (i = 0; i < 2; i++) {
901 td->io_blocks[i] = 0;
902 td->io_issues[i] = 0;
903 td->ts.total_io_u[i] = 0;
906 fio_gettime(&tv, NULL);
907 memcpy(&td->epoch, &tv, sizeof(tv));
908 memcpy(&td->start, &tv, sizeof(tv));
911 static void clear_io_state(struct thread_data *td)
916 reset_io_counters(td);
919 for_each_file(td, f, i)
920 fio_file_clear_done(f);
923 static int exec_string(const char *string)
925 int ret, newlen = strlen(string) + 1 + 8;
928 str = malloc(newlen);
929 sprintf(str, "sh -c %s", string);
933 log_err("fio: exec of cmd <%s> failed\n", str);
940 * Entry point for the thread based jobs. The process based jobs end up
941 * here as well, after a little setup.
943 static void *thread_main(void *data)
945 unsigned long long runtime[2], elapsed;
946 struct thread_data *td = data;
949 if (!td->o.use_thread)
954 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
956 INIT_FLIST_HEAD(&td->io_u_freelist);
957 INIT_FLIST_HEAD(&td->io_u_busylist);
958 INIT_FLIST_HEAD(&td->io_u_requeues);
959 INIT_FLIST_HEAD(&td->io_log_list);
960 INIT_FLIST_HEAD(&td->io_hist_list);
961 td->io_hist_tree = RB_ROOT;
963 td_set_runstate(td, TD_INITIALIZED);
964 dprint(FD_MUTEX, "up startup_mutex\n");
965 fio_mutex_up(startup_mutex);
966 dprint(FD_MUTEX, "wait on td->mutex\n");
967 fio_mutex_down(td->mutex);
968 dprint(FD_MUTEX, "done waiting on td->mutex\n");
971 * the ->mutex mutex is now no longer used, close it to avoid
972 * eating a file descriptor
974 fio_mutex_remove(td->mutex);
977 * May alter parameters that init_io_u() will use, so we need to
986 if (td->o.cpumask_set && fio_setaffinity(td) == -1) {
987 td_verror(td, errno, "cpu_set_affinity");
992 * If we have a gettimeofday() thread, make sure we exclude that
993 * thread from this job
995 if (td->o.gtod_cpu) {
996 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
997 if (fio_setaffinity(td) == -1) {
998 td_verror(td, errno, "cpu_set_affinity");
1003 if (td->ioprio_set) {
1004 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1005 td_verror(td, errno, "ioprio_set");
1010 if (nice(td->o.nice) == -1) {
1011 td_verror(td, errno, "nice");
1015 if (td->o.ioscheduler && switch_ioscheduler(td))
1018 if (!td->o.create_serialize && setup_files(td))
1024 if (init_random_map(td))
1027 if (td->o.exec_prerun) {
1028 if (exec_string(td->o.exec_prerun))
1032 if (td->o.pre_read) {
1033 if (pre_read_files(td) < 0)
1037 fio_gettime(&td->epoch, NULL);
1038 getrusage(RUSAGE_SELF, &td->ts.ru_start);
1040 runtime[0] = runtime[1] = 0;
1042 while (keep_running(td)) {
1043 fio_gettime(&td->start, NULL);
1044 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
1045 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1047 if (td->o.ratemin[0] || td->o.ratemin[1])
1048 memcpy(&td->lastrate, &td->ts.stat_sample_time,
1049 sizeof(td->lastrate));
1054 prune_io_piece_log(td);
1060 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1061 elapsed = utime_since_now(&td->start);
1062 runtime[DDIR_READ] += elapsed;
1064 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1065 elapsed = utime_since_now(&td->start);
1066 runtime[DDIR_WRITE] += elapsed;
1069 if (td->error || td->terminate)
1072 if (!td->o.do_verify ||
1073 td->o.verify == VERIFY_NONE ||
1074 (td->io_ops->flags & FIO_UNIDIR))
1079 fio_gettime(&td->start, NULL);
1083 runtime[DDIR_READ] += utime_since_now(&td->start);
1085 if (td->error || td->terminate)
1089 update_rusage_stat(td);
1090 td->ts.runtime[0] = (runtime[0] + 999) / 1000;
1091 td->ts.runtime[1] = (runtime[1] + 999) / 1000;
1092 td->ts.total_run_time = mtime_since_now(&td->epoch);
1093 td->ts.io_bytes[0] = td->io_bytes[0];
1094 td->ts.io_bytes[1] = td->io_bytes[1];
1096 fio_mutex_down(writeout_mutex);
1097 if (td->ts.bw_log) {
1098 if (td->o.bw_log_file) {
1099 finish_log_named(td, td->ts.bw_log,
1100 td->o.bw_log_file, "bw");
1102 finish_log(td, td->ts.bw_log, "bw");
1104 if (td->ts.slat_log) {
1105 if (td->o.lat_log_file) {
1106 finish_log_named(td, td->ts.slat_log,
1107 td->o.lat_log_file, "slat");
1109 finish_log(td, td->ts.slat_log, "slat");
1111 if (td->ts.clat_log) {
1112 if (td->o.lat_log_file) {
1113 finish_log_named(td, td->ts.clat_log,
1114 td->o.lat_log_file, "clat");
1116 finish_log(td, td->ts.clat_log, "clat");
1118 fio_mutex_up(writeout_mutex);
1119 if (td->o.exec_postrun)
1120 exec_string(td->o.exec_postrun);
1122 if (exitall_on_terminate)
1123 terminate_threads(td->groupid);
1127 printf("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1129 close_and_free_files(td);
1133 if (td->o.cpumask_set) {
1134 int ret = fio_cpuset_exit(&td->o.cpumask);
1136 td_verror(td, ret, "fio_cpuset_exit");
1140 * do this very late, it will log file closing as well
1142 if (td->o.write_iolog_file)
1143 write_iolog_close(td);
1145 options_mem_free(td);
1146 td_set_runstate(td, TD_EXITED);
1147 return (void *) (unsigned long) td->error;
1151 * We cannot pass the td data into a forked process, so attach the td and
1152 * pass it to the thread worker.
1154 static int fork_main(int shmid, int offset)
1156 struct thread_data *td;
1159 data = shmat(shmid, NULL, 0);
1160 if (data == (void *) -1) {
1167 td = data + offset * sizeof(struct thread_data);
1168 ret = thread_main(td);
1170 return (int) (unsigned long) ret;
1174 * Run over the job map and reap the threads that have exited, if any.
1176 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1178 struct thread_data *td;
1179 int i, cputhreads, realthreads, pending, status, ret;
1182 * reap exited threads (TD_EXITED -> TD_REAPED)
1184 realthreads = pending = cputhreads = 0;
1185 for_each_td(td, i) {
1189 * ->io_ops is NULL for a thread that has closed its
1192 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1201 if (td->runstate == TD_REAPED)
1203 if (td->o.use_thread) {
1204 if (td->runstate == TD_EXITED) {
1205 td_set_runstate(td, TD_REAPED);
1212 if (td->runstate == TD_EXITED)
1216 * check if someone quit or got killed in an unusual way
1218 ret = waitpid(td->pid, &status, flags);
1220 if (errno == ECHILD) {
1221 log_err("fio: pid=%d disappeared %d\n",
1222 (int) td->pid, td->runstate);
1223 td_set_runstate(td, TD_REAPED);
1227 } else if (ret == td->pid) {
1228 if (WIFSIGNALED(status)) {
1229 int sig = WTERMSIG(status);
1232 log_err("fio: pid=%d, got signal=%d\n",
1233 (int) td->pid, sig);
1234 td_set_runstate(td, TD_REAPED);
1237 if (WIFEXITED(status)) {
1238 if (WEXITSTATUS(status) && !td->error)
1239 td->error = WEXITSTATUS(status);
1241 td_set_runstate(td, TD_REAPED);
1247 * thread is not dead, continue
1253 (*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
1254 (*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
1261 done_secs += mtime_since_now(&td->epoch) / 1000;
1264 if (*nr_running == cputhreads && !pending && realthreads)
1265 terminate_threads(TERMINATE_ALL);
1268 static void *gtod_thread_main(void *data)
1270 fio_mutex_up(startup_mutex);
1273 * As long as we have jobs around, update the clock. It would be nice
1274 * to have some way of NOT hammering that CPU with gettimeofday(),
1275 * but I'm not sure what to use outside of a simple CPU nop to relax
1276 * it - we don't want to lose precision.
1286 static int fio_start_gtod_thread(void)
1290 ret = pthread_create(>od_thread, NULL, gtod_thread_main, NULL);
1292 log_err("Can't create gtod thread: %s\n", strerror(ret));
1296 ret = pthread_detach(gtod_thread);
1298 log_err("Can't detatch gtod thread: %s\n", strerror(ret));
1302 dprint(FD_MUTEX, "wait on startup_mutex\n");
1303 fio_mutex_down(startup_mutex);
1304 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1309 * Main function for kicking off and reaping jobs, as needed.
1311 static void run_threads(void)
1313 struct thread_data *td;
1314 unsigned long spent;
1315 int i, todo, nr_running, m_rate, t_rate, nr_started;
1317 if (fio_pin_memory())
1320 if (fio_gtod_offload && fio_start_gtod_thread())
1323 if (!terse_output) {
1324 printf("Starting ");
1326 printf("%d thread%s", nr_thread,
1327 nr_thread > 1 ? "s" : "");
1331 printf("%d process%s", nr_process,
1332 nr_process > 1 ? "es" : "");
1340 todo = thread_number;
1343 m_rate = t_rate = 0;
1345 for_each_td(td, i) {
1346 print_status_init(td->thread_number - 1);
1348 if (!td->o.create_serialize) {
1354 * do file setup here so it happens sequentially,
1355 * we don't want X number of threads getting their
1356 * client data interspersed on disk
1358 if (setup_files(td)) {
1361 log_err("fio: pid=%d, err=%d/%s\n",
1362 (int) td->pid, td->error, td->verror);
1363 td_set_runstate(td, TD_REAPED);
1370 * for sharing to work, each job must always open
1371 * its own files. so close them, if we opened them
1374 for_each_file(td, f, i) {
1375 if (fio_file_open(f))
1376 td_io_close_file(td, f);
1378 assert(f->fd == -1);
1388 struct thread_data *map[MAX_JOBS];
1389 struct timeval this_start;
1390 int this_jobs = 0, left;
1393 * create threads (TD_NOT_CREATED -> TD_CREATED)
1395 for_each_td(td, i) {
1396 if (td->runstate != TD_NOT_CREATED)
1400 * never got a chance to start, killed by other
1401 * thread for some reason
1403 if (td->terminate) {
1408 if (td->o.start_delay) {
1409 spent = mtime_since_genesis();
1411 if (td->o.start_delay * 1000 > spent)
1415 if (td->o.stonewall && (nr_started || nr_running)) {
1416 dprint(FD_PROCESS, "%s: stonewall wait\n",
1422 * Set state to created. Thread will transition
1423 * to TD_INITIALIZED when it's done setting up.
1425 td_set_runstate(td, TD_CREATED);
1426 map[this_jobs++] = td;
1429 if (td->o.use_thread) {
1432 dprint(FD_PROCESS, "will pthread_create\n");
1433 ret = pthread_create(&td->thread, NULL,
1436 log_err("pthread_create: %s\n",
1441 ret = pthread_detach(td->thread);
1443 log_err("pthread_detach: %s",
1447 dprint(FD_PROCESS, "will fork\n");
1450 int ret = fork_main(shm_id, i);
1453 } else if (i == fio_debug_jobno)
1454 *fio_debug_jobp = pid;
1456 dprint(FD_MUTEX, "wait on startup_mutex\n");
1457 fio_mutex_down(startup_mutex);
1458 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1462 * Wait for the started threads to transition to
1465 fio_gettime(&this_start, NULL);
1467 while (left && !fio_abort) {
1468 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1473 for (i = 0; i < this_jobs; i++) {
1477 if (td->runstate == TD_INITIALIZED) {
1480 } else if (td->runstate >= TD_EXITED) {
1484 nr_running++; /* work-around... */
1490 log_err("fio: %d jobs failed to start\n", left);
1491 for (i = 0; i < this_jobs; i++) {
1495 kill(td->pid, SIGTERM);
1501 * start created threads (TD_INITIALIZED -> TD_RUNNING).
1503 for_each_td(td, i) {
1504 if (td->runstate != TD_INITIALIZED)
1507 if (in_ramp_time(td))
1508 td_set_runstate(td, TD_RAMP);
1510 td_set_runstate(td, TD_RUNNING);
1513 m_rate += td->o.ratemin[0] + td->o.ratemin[1];
1514 t_rate += td->o.rate[0] + td->o.rate[1];
1516 fio_mutex_up(td->mutex);
1519 reap_threads(&nr_running, &t_rate, &m_rate);
1525 while (nr_running) {
1526 reap_threads(&nr_running, &t_rate, &m_rate);
1534 int main(int argc, char *argv[])
1541 * We need locale for number printing, if it isn't set then just
1542 * go with the US format.
1544 if (!getenv("LC_NUMERIC"))
1545 setlocale(LC_NUMERIC, "en_US");
1547 if (parse_options(argc, argv))
1553 ps = sysconf(_SC_PAGESIZE);
1555 log_err("Failed to get page size\n");
1563 setup_log(&agg_io_log[DDIR_READ]);
1564 setup_log(&agg_io_log[DDIR_WRITE]);
1567 startup_mutex = fio_mutex_init(0);
1568 writeout_mutex = fio_mutex_init(1);
1579 __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1580 __finish_log(agg_io_log[DDIR_WRITE],
1581 "agg-write_bw.log");
1585 fio_mutex_remove(startup_mutex);
1586 fio_mutex_remove(writeout_mutex);