2 * fio - the flexible io tester
4 * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
5 * Copyright (C) 2006-2012 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
52 static pthread_t disk_util_thread;
53 static struct fio_mutex *disk_thread_mutex;
54 static struct fio_mutex *startup_mutex;
55 static struct fio_mutex *writeout_mutex;
56 static struct flist_head *cgroup_list;
57 static char *cgroup_mnt;
58 static int exit_value;
59 static volatile int fio_abort;
61 struct io_log *agg_io_log[DDIR_RWDIR_CNT];
64 unsigned int thread_number = 0;
65 unsigned int nr_process = 0;
66 unsigned int nr_thread = 0;
69 unsigned long done_secs = 0;
71 #define PAGE_ALIGN(buf) \
72 (char *) (((uintptr_t) (buf) + page_mask) & ~page_mask)
74 #define JOB_START_TIMEOUT (5 * 1000)
76 static void sig_int(int sig)
80 fio_server_got_signal(sig);
82 log_info("\nfio: terminating on signal %d\n", sig);
87 fio_terminate_threads(TERMINATE_ALL);
91 static void sig_show_status(int sig)
93 show_running_run_stats();
96 static void set_sig_handlers(void)
100 memset(&act, 0, sizeof(act));
101 act.sa_handler = sig_int;
102 act.sa_flags = SA_RESTART;
103 sigaction(SIGINT, &act, NULL);
105 memset(&act, 0, sizeof(act));
106 act.sa_handler = sig_int;
107 act.sa_flags = SA_RESTART;
108 sigaction(SIGTERM, &act, NULL);
110 memset(&act, 0, sizeof(act));
111 act.sa_handler = sig_show_status;
112 act.sa_flags = SA_RESTART;
113 sigaction(SIGUSR1, &act, NULL);
116 memset(&act, 0, sizeof(act));
117 act.sa_handler = sig_int;
118 act.sa_flags = SA_RESTART;
119 sigaction(SIGPIPE, &act, NULL);
124 * Check if we are above the minimum rate given.
126 static int __check_min_rate(struct thread_data *td, struct timeval *now,
129 unsigned long long bytes = 0;
130 unsigned long iops = 0;
133 unsigned int ratemin = 0;
134 unsigned int rate_iops = 0;
135 unsigned int rate_iops_min = 0;
137 assert(ddir_rw(ddir));
139 if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
143 * allow a 2 second settle period in the beginning
145 if (mtime_since(&td->start, now) < 2000)
148 iops += td->this_io_blocks[ddir];
149 bytes += td->this_io_bytes[ddir];
150 ratemin += td->o.ratemin[ddir];
151 rate_iops += td->o.rate_iops[ddir];
152 rate_iops_min += td->o.rate_iops_min[ddir];
155 * if rate blocks is set, sample is running
157 if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
158 spent = mtime_since(&td->lastrate[ddir], now);
159 if (spent < td->o.ratecycle)
162 if (td->o.rate[ddir]) {
164 * check bandwidth specified rate
166 if (bytes < td->rate_bytes[ddir]) {
167 log_err("%s: min rate %u not met\n", td->o.name,
171 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
172 if (rate < ratemin ||
173 bytes < td->rate_bytes[ddir]) {
174 log_err("%s: min rate %u not met, got"
175 " %luKB/sec\n", td->o.name,
182 * checks iops specified rate
184 if (iops < rate_iops) {
185 log_err("%s: min iops rate %u not met\n",
186 td->o.name, rate_iops);
189 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
190 if (rate < rate_iops_min ||
191 iops < td->rate_blocks[ddir]) {
192 log_err("%s: min iops rate %u not met,"
193 " got %lu\n", td->o.name,
194 rate_iops_min, rate);
200 td->rate_bytes[ddir] = bytes;
201 td->rate_blocks[ddir] = iops;
202 memcpy(&td->lastrate[ddir], now, sizeof(*now));
206 static int check_min_rate(struct thread_data *td, struct timeval *now,
207 unsigned long *bytes_done)
211 if (bytes_done[DDIR_READ])
212 ret |= __check_min_rate(td, now, DDIR_READ);
213 if (bytes_done[DDIR_WRITE])
214 ret |= __check_min_rate(td, now, DDIR_WRITE);
215 if (bytes_done[DDIR_TRIM])
216 ret |= __check_min_rate(td, now, DDIR_TRIM);
222 * When job exits, we can cancel the in-flight IO if we are using async
223 * io. Attempt to do so.
225 static void cleanup_pending_aio(struct thread_data *td)
227 struct flist_head *entry, *n;
232 * get immediately available events, if any
234 r = io_u_queued_complete(td, 0, NULL);
239 * now cancel remaining active events
241 if (td->io_ops->cancel) {
242 flist_for_each_safe(entry, n, &td->io_u_busylist) {
243 io_u = flist_entry(entry, struct io_u, list);
246 * if the io_u isn't in flight, then that generally
247 * means someone leaked an io_u. complain but fix
248 * it up, so we don't stall here.
250 if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
251 log_err("fio: non-busy IO on busy list\n");
254 r = td->io_ops->cancel(td, io_u);
262 r = io_u_queued_complete(td, td->cur_depth, NULL);
266 * Helper to handle the final sync of a file. Works just like the normal
267 * io path, just does everything sync.
269 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
271 struct io_u *io_u = __get_io_u(td);
277 io_u->ddir = DDIR_SYNC;
280 if (td_io_prep(td, io_u)) {
286 ret = td_io_queue(td, io_u);
288 td_verror(td, io_u->error, "td_io_queue");
291 } else if (ret == FIO_Q_QUEUED) {
292 if (io_u_queued_complete(td, 1, NULL) < 0)
294 } else if (ret == FIO_Q_COMPLETED) {
296 td_verror(td, io_u->error, "td_io_queue");
300 if (io_u_sync_complete(td, io_u, NULL) < 0)
302 } else if (ret == FIO_Q_BUSY) {
303 if (td_io_commit(td))
311 static inline void __update_tv_cache(struct thread_data *td)
313 fio_gettime(&td->tv_cache, NULL);
316 static inline void update_tv_cache(struct thread_data *td)
318 if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
319 __update_tv_cache(td);
322 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
324 if (in_ramp_time(td))
328 if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
334 static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
339 if (ret < 0 || td->error) {
341 enum error_type_bit eb;
346 eb = td_error_type(ddir, err);
347 if (!(td->o.continue_on_error & (1 << eb)))
350 if (td_non_fatal_error(td, eb, err)) {
352 * Continue with the I/Os in case of
355 update_error_count(td, err);
359 } else if (td->o.fill_device && err == ENOSPC) {
361 * We expect to hit this error if
362 * fill_device option is set.
369 * Stop the I/O in case of a fatal
372 update_error_count(td, err);
381 * The main verify engine. Runs over the writes we previously submitted,
382 * reads the blocks back in, and checks the crc/md5 of the data.
384 static void do_verify(struct thread_data *td)
391 dprint(FD_VERIFY, "starting loop\n");
394 * sync io first and invalidate cache, to make sure we really
397 for_each_file(td, f, i) {
398 if (!fio_file_open(f))
400 if (fio_io_sync(td, f))
402 if (file_invalidate_cache(td, f))
409 td_set_runstate(td, TD_VERIFYING);
412 while (!td->terminate) {
417 if (runtime_exceeded(td, &td->tv_cache)) {
418 __update_tv_cache(td);
419 if (runtime_exceeded(td, &td->tv_cache)) {
425 if (flow_threshold_exceeded(td))
428 io_u = __get_io_u(td);
432 if (get_next_verify(td, io_u)) {
437 if (td_io_prep(td, io_u)) {
442 if (td->o.verify_async)
443 io_u->end_io = verify_io_u_async;
445 io_u->end_io = verify_io_u;
447 ret = td_io_queue(td, io_u);
449 case FIO_Q_COMPLETED:
452 clear_io_u(td, io_u);
453 } else if (io_u->resid) {
454 int bytes = io_u->xfer_buflen - io_u->resid;
460 td_verror(td, EIO, "full resid");
465 io_u->xfer_buflen = io_u->resid;
466 io_u->xfer_buf += bytes;
467 io_u->offset += bytes;
469 if (ddir_rw(io_u->ddir))
470 td->ts.short_io_u[io_u->ddir]++;
473 if (io_u->offset == f->real_file_size)
476 requeue_io_u(td, &io_u);
479 ret = io_u_sync_complete(td, io_u, NULL);
487 requeue_io_u(td, &io_u);
488 ret2 = td_io_commit(td);
494 td_verror(td, -ret, "td_io_queue");
498 if (break_on_this_error(td, io_u->ddir, &ret))
502 * if we can queue more, do so. but check if there are
503 * completed io_u's first. Note that we can get BUSY even
504 * without IO queued, if the system is resource starved.
506 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
507 if (full || !td->o.iodepth_batch_complete) {
508 min_events = min(td->o.iodepth_batch_complete,
511 * if the queue is full, we MUST reap at least 1 event
513 if (full && !min_events)
518 * Reap required number of io units, if any,
519 * and do the verification on them through
520 * the callback handler
522 if (io_u_queued_complete(td, min_events, NULL) < 0) {
526 } while (full && (td->cur_depth > td->o.iodepth_low));
533 min_events = td->cur_depth;
536 ret = io_u_queued_complete(td, min_events, NULL);
538 cleanup_pending_aio(td);
540 td_set_runstate(td, TD_RUNNING);
542 dprint(FD_VERIFY, "exiting loop\n");
545 static int io_bytes_exceeded(struct thread_data *td)
547 unsigned long long bytes;
550 bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE];
551 else if (td_write(td))
552 bytes = td->this_io_bytes[DDIR_WRITE];
553 else if (td_read(td))
554 bytes = td->this_io_bytes[DDIR_READ];
556 bytes = td->this_io_bytes[DDIR_TRIM];
558 return bytes >= td->o.size;
562 * Main IO worker function. It retrieves io_u's to process and queues
563 * and reaps them, checking for rate and errors along the way.
565 static void do_io(struct thread_data *td)
570 if (in_ramp_time(td))
571 td_set_runstate(td, TD_RAMP);
573 td_set_runstate(td, TD_RUNNING);
575 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
576 (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td) ||
578 struct timeval comp_time;
579 unsigned long bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
590 if (runtime_exceeded(td, &td->tv_cache)) {
591 __update_tv_cache(td);
592 if (runtime_exceeded(td, &td->tv_cache)) {
598 if (flow_threshold_exceeded(td))
608 * Add verification end_io handler if:
609 * - Asked to verify (!td_rw(td))
610 * - Or the io_u is from our verify list (mixed write/ver)
612 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
613 ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
614 if (td->o.verify_async)
615 io_u->end_io = verify_io_u_async;
617 io_u->end_io = verify_io_u;
618 td_set_runstate(td, TD_VERIFYING);
619 } else if (in_ramp_time(td))
620 td_set_runstate(td, TD_RAMP);
622 td_set_runstate(td, TD_RUNNING);
624 ret = td_io_queue(td, io_u);
626 case FIO_Q_COMPLETED:
629 clear_io_u(td, io_u);
630 } else if (io_u->resid) {
631 int bytes = io_u->xfer_buflen - io_u->resid;
632 struct fio_file *f = io_u->file;
638 td_verror(td, EIO, "full resid");
643 io_u->xfer_buflen = io_u->resid;
644 io_u->xfer_buf += bytes;
645 io_u->offset += bytes;
647 if (ddir_rw(io_u->ddir))
648 td->ts.short_io_u[io_u->ddir]++;
650 if (io_u->offset == f->real_file_size)
653 requeue_io_u(td, &io_u);
656 if (__should_check_rate(td, DDIR_READ) ||
657 __should_check_rate(td, DDIR_WRITE) ||
658 __should_check_rate(td, DDIR_TRIM))
659 fio_gettime(&comp_time, NULL);
661 ret = io_u_sync_complete(td, io_u, bytes_done);
668 * if the engine doesn't have a commit hook,
669 * the io_u is really queued. if it does have such
670 * a hook, it has to call io_u_queued() itself.
672 if (td->io_ops->commit == NULL)
673 io_u_queued(td, io_u);
676 requeue_io_u(td, &io_u);
677 ret2 = td_io_commit(td);
687 if (break_on_this_error(td, ddir, &ret))
691 * See if we need to complete some commands. Note that we
692 * can get BUSY even without IO queued, if the system is
695 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
696 if (full || !td->o.iodepth_batch_complete) {
697 min_evts = min(td->o.iodepth_batch_complete,
700 * if the queue is full, we MUST reap at least 1 event
702 if (full && !min_evts)
705 if (__should_check_rate(td, DDIR_READ) ||
706 __should_check_rate(td, DDIR_WRITE) ||
707 __should_check_rate(td, DDIR_TRIM))
708 fio_gettime(&comp_time, NULL);
711 ret = io_u_queued_complete(td, min_evts, bytes_done);
715 } while (full && (td->cur_depth > td->o.iodepth_low));
720 if (!ddir_rw_sum(bytes_done))
723 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
724 if (check_min_rate(td, &comp_time, bytes_done)) {
725 if (exitall_on_terminate)
726 fio_terminate_threads(td->groupid);
727 td_verror(td, EIO, "check_min_rate");
732 if (td->o.thinktime) {
733 unsigned long long b;
735 b = ddir_rw_sum(td->io_blocks);
736 if (!(b % td->o.thinktime_blocks)) {
739 if (td->o.thinktime_spin)
740 usec_spin(td->o.thinktime_spin);
742 left = td->o.thinktime - td->o.thinktime_spin;
744 usec_sleep(td, left);
749 if (td->trim_entries)
750 log_err("fio: %d trim entries leaked?\n", td->trim_entries);
752 if (td->o.fill_device && td->error == ENOSPC) {
761 ret = io_u_queued_complete(td, i, NULL);
762 if (td->o.fill_device && td->error == ENOSPC)
766 if (should_fsync(td) && td->o.end_fsync) {
767 td_set_runstate(td, TD_FSYNCING);
769 for_each_file(td, f, i) {
770 if (!fio_file_open(f))
776 cleanup_pending_aio(td);
779 * stop job if we failed doing any IO
781 if (!ddir_rw_sum(td->this_io_bytes))
785 static void cleanup_io_u(struct thread_data *td)
787 struct flist_head *entry, *n;
790 flist_for_each_safe(entry, n, &td->io_u_freelist) {
791 io_u = flist_entry(entry, struct io_u, list);
793 flist_del(&io_u->list);
794 fio_memfree(io_u, sizeof(*io_u));
800 static int init_io_u(struct thread_data *td)
803 unsigned int max_bs, min_write;
804 int cl_align, i, max_units;
808 max_units = td->o.iodepth;
809 max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
810 max_bs = max(td->o.max_bs[DDIR_TRIM], max_bs);
811 min_write = td->o.min_bs[DDIR_WRITE];
812 td->orig_buffer_size = (unsigned long long) max_bs
813 * (unsigned long long) max_units;
815 if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
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 (data_xfer && 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 ptr = fio_memalign(cl_align, sizeof(*io_u));
849 log_err("fio: unable to allocate aligned memory\n");
854 memset(io_u, 0, sizeof(*io_u));
855 INIT_FLIST_HEAD(&io_u->list);
856 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
860 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
863 io_u_fill_buffer(td, io_u, min_write, max_bs);
864 if (td_write(td) && td->o.verify_pattern_bytes) {
866 * Fill the buffer with the pattern if we are
867 * going to be doing writes.
869 fill_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
874 io_u->flags = IO_U_F_FREE;
875 flist_add(&io_u->list, &td->io_u_freelist);
882 static int switch_ioscheduler(struct thread_data *td)
884 char tmp[256], tmp2[128];
888 if (td->io_ops->flags & FIO_DISKLESSIO)
891 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
893 f = fopen(tmp, "r+");
895 if (errno == ENOENT) {
896 log_err("fio: os or kernel doesn't support IO scheduler"
900 td_verror(td, errno, "fopen iosched");
907 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
908 if (ferror(f) || ret != 1) {
909 td_verror(td, errno, "fwrite");
917 * Read back and check that the selected scheduler is now the default.
919 ret = fread(tmp, 1, sizeof(tmp), f);
920 if (ferror(f) || ret < 0) {
921 td_verror(td, errno, "fread");
926 sprintf(tmp2, "[%s]", td->o.ioscheduler);
927 if (!strstr(tmp, tmp2)) {
928 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
929 td_verror(td, EINVAL, "iosched_switch");
938 static int keep_running(struct thread_data *td)
942 if (td->o.time_based)
949 if (ddir_rw_sum(td->io_bytes) < td->o.size)
955 static int exec_string(const char *string)
957 int ret, newlen = strlen(string) + 1 + 8;
960 str = malloc(newlen);
961 sprintf(str, "sh -c %s", string);
965 log_err("fio: exec of cmd <%s> failed\n", str);
972 * Entry point for the thread based jobs. The process based jobs end up
973 * here as well, after a little setup.
975 static void *thread_main(void *data)
977 unsigned long long elapsed;
978 struct thread_data *td = data;
979 pthread_condattr_t attr;
982 if (!td->o.use_thread) {
988 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
990 INIT_FLIST_HEAD(&td->io_u_freelist);
991 INIT_FLIST_HEAD(&td->io_u_busylist);
992 INIT_FLIST_HEAD(&td->io_u_requeues);
993 INIT_FLIST_HEAD(&td->io_log_list);
994 INIT_FLIST_HEAD(&td->io_hist_list);
995 INIT_FLIST_HEAD(&td->verify_list);
996 INIT_FLIST_HEAD(&td->trim_list);
997 pthread_mutex_init(&td->io_u_lock, NULL);
998 td->io_hist_tree = RB_ROOT;
1000 pthread_condattr_init(&attr);
1001 pthread_cond_init(&td->verify_cond, &attr);
1002 pthread_cond_init(&td->free_cond, &attr);
1004 td_set_runstate(td, TD_INITIALIZED);
1005 dprint(FD_MUTEX, "up startup_mutex\n");
1006 fio_mutex_up(startup_mutex);
1007 dprint(FD_MUTEX, "wait on td->mutex\n");
1008 fio_mutex_down(td->mutex);
1009 dprint(FD_MUTEX, "done waiting on td->mutex\n");
1012 * the ->mutex mutex is now no longer used, close it to avoid
1013 * eating a file descriptor
1015 fio_mutex_remove(td->mutex);
1018 * A new gid requires privilege, so we need to do this before setting
1021 if (td->o.gid != -1U && setgid(td->o.gid)) {
1022 td_verror(td, errno, "setgid");
1025 if (td->o.uid != -1U && setuid(td->o.uid)) {
1026 td_verror(td, errno, "setuid");
1031 * If we have a gettimeofday() thread, make sure we exclude that
1032 * thread from this job
1035 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1038 * Set affinity first, in case it has an impact on the memory
1041 if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1042 td_verror(td, errno, "cpu_set_affinity");
1047 * May alter parameters that init_io_u() will use, so we need to
1056 if (td->o.verify_async && verify_async_init(td))
1059 if (td->ioprio_set) {
1060 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1061 td_verror(td, errno, "ioprio_set");
1066 if (td->o.cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1070 if (nice(td->o.nice) == -1 && errno != 0) {
1071 td_verror(td, errno, "nice");
1075 if (td->o.ioscheduler && switch_ioscheduler(td))
1078 if (!td->o.create_serialize && setup_files(td))
1084 if (init_random_map(td))
1087 if (td->o.exec_prerun) {
1088 if (exec_string(td->o.exec_prerun))
1092 if (td->o.pre_read) {
1093 if (pre_read_files(td) < 0)
1097 fio_gettime(&td->epoch, NULL);
1098 getrusage(RUSAGE_SELF, &td->ru_start);
1101 while (keep_running(td)) {
1102 fio_gettime(&td->start, NULL);
1103 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1104 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1105 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1107 if (td->o.ratemin[DDIR_READ] || td->o.ratemin[DDIR_WRITE] ||
1108 td->o.ratemin[DDIR_TRIM]) {
1109 memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1110 sizeof(td->bw_sample_time));
1111 memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1112 sizeof(td->bw_sample_time));
1113 memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1114 sizeof(td->bw_sample_time));
1120 prune_io_piece_log(td);
1126 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1127 elapsed = utime_since_now(&td->start);
1128 td->ts.runtime[DDIR_READ] += elapsed;
1130 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1131 elapsed = utime_since_now(&td->start);
1132 td->ts.runtime[DDIR_WRITE] += elapsed;
1134 if (td_trim(td) && td->io_bytes[DDIR_TRIM]) {
1135 elapsed = utime_since_now(&td->start);
1136 td->ts.runtime[DDIR_TRIM] += elapsed;
1139 if (td->error || td->terminate)
1142 if (!td->o.do_verify ||
1143 td->o.verify == VERIFY_NONE ||
1144 (td->io_ops->flags & FIO_UNIDIR))
1149 fio_gettime(&td->start, NULL);
1153 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1155 if (td->error || td->terminate)
1159 update_rusage_stat(td);
1160 td->ts.runtime[DDIR_READ] = (td->ts.runtime[DDIR_READ] + 999) / 1000;
1161 td->ts.runtime[DDIR_WRITE] = (td->ts.runtime[DDIR_WRITE] + 999) / 1000;
1162 td->ts.runtime[DDIR_TRIM] = (td->ts.runtime[DDIR_TRIM] + 999) / 1000;
1163 td->ts.total_run_time = mtime_since_now(&td->epoch);
1164 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1165 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1166 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1168 fio_mutex_down(writeout_mutex);
1170 if (td->o.bw_log_file) {
1171 finish_log_named(td, td->bw_log,
1172 td->o.bw_log_file, "bw");
1174 finish_log(td, td->bw_log, "bw");
1177 if (td->o.lat_log_file) {
1178 finish_log_named(td, td->lat_log,
1179 td->o.lat_log_file, "lat");
1181 finish_log(td, td->lat_log, "lat");
1184 if (td->o.lat_log_file) {
1185 finish_log_named(td, td->slat_log,
1186 td->o.lat_log_file, "slat");
1188 finish_log(td, td->slat_log, "slat");
1191 if (td->o.lat_log_file) {
1192 finish_log_named(td, td->clat_log,
1193 td->o.lat_log_file, "clat");
1195 finish_log(td, td->clat_log, "clat");
1198 if (td->o.iops_log_file) {
1199 finish_log_named(td, td->iops_log,
1200 td->o.iops_log_file, "iops");
1202 finish_log(td, td->iops_log, "iops");
1205 fio_mutex_up(writeout_mutex);
1206 if (td->o.exec_postrun)
1207 exec_string(td->o.exec_postrun);
1209 if (exitall_on_terminate)
1210 fio_terminate_threads(td->groupid);
1214 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1217 if (td->o.verify_async)
1218 verify_async_exit(td);
1220 close_and_free_files(td);
1223 cgroup_shutdown(td, &cgroup_mnt);
1225 if (td->o.cpumask_set) {
1226 int ret = fio_cpuset_exit(&td->o.cpumask);
1228 td_verror(td, ret, "fio_cpuset_exit");
1232 * do this very late, it will log file closing as well
1234 if (td->o.write_iolog_file)
1235 write_iolog_close(td);
1237 td_set_runstate(td, TD_EXITED);
1238 return (void *) (uintptr_t) td->error;
1243 * We cannot pass the td data into a forked process, so attach the td and
1244 * pass it to the thread worker.
1246 static int fork_main(int shmid, int offset)
1248 struct thread_data *td;
1252 data = shmat(shmid, NULL, 0);
1253 if (data == (void *) -1) {
1261 * HP-UX inherits shm mappings?
1266 td = data + offset * sizeof(struct thread_data);
1267 ret = thread_main(td);
1269 return (int) (uintptr_t) ret;
1273 * Run over the job map and reap the threads that have exited, if any.
1275 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1276 unsigned int *m_rate)
1278 struct thread_data *td;
1279 unsigned int cputhreads, realthreads, pending;
1283 * reap exited threads (TD_EXITED -> TD_REAPED)
1285 realthreads = pending = cputhreads = 0;
1286 for_each_td(td, i) {
1290 * ->io_ops is NULL for a thread that has closed its
1293 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1302 if (td->runstate == TD_REAPED)
1304 if (td->o.use_thread) {
1305 if (td->runstate == TD_EXITED) {
1306 td_set_runstate(td, TD_REAPED);
1313 if (td->runstate == TD_EXITED)
1317 * check if someone quit or got killed in an unusual way
1319 ret = waitpid(td->pid, &status, flags);
1321 if (errno == ECHILD) {
1322 log_err("fio: pid=%d disappeared %d\n",
1323 (int) td->pid, td->runstate);
1325 td_set_runstate(td, TD_REAPED);
1329 } else if (ret == td->pid) {
1330 if (WIFSIGNALED(status)) {
1331 int sig = WTERMSIG(status);
1334 log_err("fio: pid=%d, got signal=%d\n",
1335 (int) td->pid, sig);
1337 td_set_runstate(td, TD_REAPED);
1340 if (WIFEXITED(status)) {
1341 if (WEXITSTATUS(status) && !td->error)
1342 td->error = WEXITSTATUS(status);
1344 td_set_runstate(td, TD_REAPED);
1350 * thread is not dead, continue
1356 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1357 (*t_rate) -= ddir_rw_sum(td->o.rate);
1364 done_secs += mtime_since_now(&td->epoch) / 1000;
1367 if (*nr_running == cputhreads && !pending && realthreads)
1368 fio_terminate_threads(TERMINATE_ALL);
1372 * Main function for kicking off and reaping jobs, as needed.
1374 static void run_threads(void)
1376 struct thread_data *td;
1377 unsigned long spent;
1378 unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
1380 if (fio_pin_memory())
1383 if (fio_gtod_offload && fio_start_gtod_thread())
1388 if (output_format == FIO_OUTPUT_NORMAL) {
1389 log_info("Starting ");
1391 log_info("%d thread%s", nr_thread,
1392 nr_thread > 1 ? "s" : "");
1396 log_info("%d process%s", nr_process,
1397 nr_process > 1 ? "es" : "");
1403 todo = thread_number;
1406 m_rate = t_rate = 0;
1408 for_each_td(td, i) {
1409 print_status_init(td->thread_number - 1);
1411 if (!td->o.create_serialize)
1415 * do file setup here so it happens sequentially,
1416 * we don't want X number of threads getting their
1417 * client data interspersed on disk
1419 if (setup_files(td)) {
1422 log_err("fio: pid=%d, err=%d/%s\n",
1423 (int) td->pid, td->error, td->verror);
1424 td_set_runstate(td, TD_REAPED);
1431 * for sharing to work, each job must always open
1432 * its own files. so close them, if we opened them
1435 for_each_file(td, f, j) {
1436 if (fio_file_open(f))
1437 td_io_close_file(td, f);
1445 struct thread_data *map[REAL_MAX_JOBS];
1446 struct timeval this_start;
1447 int this_jobs = 0, left;
1450 * create threads (TD_NOT_CREATED -> TD_CREATED)
1452 for_each_td(td, i) {
1453 if (td->runstate != TD_NOT_CREATED)
1457 * never got a chance to start, killed by other
1458 * thread for some reason
1460 if (td->terminate) {
1465 if (td->o.start_delay) {
1466 spent = mtime_since_genesis();
1468 if (td->o.start_delay * 1000 > spent)
1472 if (td->o.stonewall && (nr_started || nr_running)) {
1473 dprint(FD_PROCESS, "%s: stonewall wait\n",
1481 * Set state to created. Thread will transition
1482 * to TD_INITIALIZED when it's done setting up.
1484 td_set_runstate(td, TD_CREATED);
1485 map[this_jobs++] = td;
1488 if (td->o.use_thread) {
1491 dprint(FD_PROCESS, "will pthread_create\n");
1492 ret = pthread_create(&td->thread, NULL,
1495 log_err("pthread_create: %s\n",
1500 ret = pthread_detach(td->thread);
1502 log_err("pthread_detach: %s",
1506 dprint(FD_PROCESS, "will fork\n");
1509 int ret = fork_main(shm_id, i);
1512 } else if (i == fio_debug_jobno)
1513 *fio_debug_jobp = pid;
1515 dprint(FD_MUTEX, "wait on startup_mutex\n");
1516 if (fio_mutex_down_timeout(startup_mutex, 10)) {
1517 log_err("fio: job startup hung? exiting.\n");
1518 fio_terminate_threads(TERMINATE_ALL);
1523 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1527 * Wait for the started threads to transition to
1530 fio_gettime(&this_start, NULL);
1532 while (left && !fio_abort) {
1533 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1538 for (i = 0; i < this_jobs; i++) {
1542 if (td->runstate == TD_INITIALIZED) {
1545 } else if (td->runstate >= TD_EXITED) {
1549 nr_running++; /* work-around... */
1555 log_err("fio: %d job%s failed to start\n", left,
1556 left > 1 ? "s" : "");
1557 for (i = 0; i < this_jobs; i++) {
1561 kill(td->pid, SIGTERM);
1567 * start created threads (TD_INITIALIZED -> TD_RUNNING).
1569 for_each_td(td, i) {
1570 if (td->runstate != TD_INITIALIZED)
1573 if (in_ramp_time(td))
1574 td_set_runstate(td, TD_RAMP);
1576 td_set_runstate(td, TD_RUNNING);
1579 m_rate += ddir_rw_sum(td->o.ratemin);
1580 t_rate += ddir_rw_sum(td->o.rate);
1582 fio_mutex_up(td->mutex);
1585 reap_threads(&nr_running, &t_rate, &m_rate);
1589 fio_server_idle_loop();
1595 while (nr_running) {
1596 reap_threads(&nr_running, &t_rate, &m_rate);
1599 fio_server_idle_loop();
1608 void wait_for_disk_thread_exit(void)
1610 fio_mutex_down(disk_thread_mutex);
1613 static void *disk_thread_main(void *data)
1617 fio_mutex_up(startup_mutex);
1619 while (threads && !ret) {
1620 usleep(DISK_UTIL_MSEC * 1000);
1623 ret = update_io_ticks();
1626 print_thread_status();
1629 fio_mutex_up(disk_thread_mutex);
1633 static int create_disk_util_thread(void)
1639 disk_thread_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
1641 ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
1643 fio_mutex_remove(disk_thread_mutex);
1644 log_err("Can't create disk util thread: %s\n", strerror(ret));
1648 ret = pthread_detach(disk_util_thread);
1650 fio_mutex_remove(disk_thread_mutex);
1651 log_err("Can't detatch disk util thread: %s\n", strerror(ret));
1655 dprint(FD_MUTEX, "wait on startup_mutex\n");
1656 fio_mutex_down(startup_mutex);
1657 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1661 int fio_backend(void)
1663 struct thread_data *td;
1667 if (load_profile(exec_profile))
1670 exec_profile = NULL;
1676 setup_log(&agg_io_log[DDIR_READ], 0);
1677 setup_log(&agg_io_log[DDIR_WRITE], 0);
1678 setup_log(&agg_io_log[DDIR_TRIM], 0);
1681 startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
1682 if (startup_mutex == NULL)
1684 writeout_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1685 if (writeout_mutex == NULL)
1689 create_disk_util_thread();
1691 cgroup_list = smalloc(sizeof(*cgroup_list));
1692 INIT_FLIST_HEAD(cgroup_list);
1699 __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1700 __finish_log(agg_io_log[DDIR_WRITE],
1701 "agg-write_bw.log");
1702 __finish_log(agg_io_log[DDIR_TRIM],
1703 "agg-write_bw.log");
1708 fio_options_free(td);
1711 cgroup_kill(cgroup_list);
1715 fio_mutex_remove(startup_mutex);
1716 fio_mutex_remove(writeout_mutex);
1717 fio_mutex_remove(disk_thread_mutex);