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
37 #ifndef FIO_NO_HAVE_SHM_H
54 static pthread_t disk_util_thread;
55 static struct fio_mutex *disk_thread_mutex;
56 static struct fio_mutex *startup_mutex;
57 static struct fio_mutex *writeout_mutex;
58 static struct flist_head *cgroup_list;
59 static char *cgroup_mnt;
60 static int exit_value;
61 static volatile int fio_abort;
63 struct io_log *agg_io_log[DDIR_RWDIR_CNT];
66 unsigned int thread_number = 0;
67 unsigned int stat_number = 0;
68 unsigned int nr_process = 0;
69 unsigned int nr_thread = 0;
72 unsigned long done_secs = 0;
73 volatile int disk_util_exit = 0;
75 #define PAGE_ALIGN(buf) \
76 (char *) (((uintptr_t) (buf) + page_mask) & ~page_mask)
78 #define JOB_START_TIMEOUT (5 * 1000)
80 static void sig_int(int sig)
84 fio_server_got_signal(sig);
86 log_info("\nfio: terminating on signal %d\n", sig);
91 fio_terminate_threads(TERMINATE_ALL);
95 static void sig_show_status(int sig)
97 show_running_run_stats();
100 static void set_sig_handlers(void)
102 struct sigaction act;
104 memset(&act, 0, sizeof(act));
105 act.sa_handler = sig_int;
106 act.sa_flags = SA_RESTART;
107 sigaction(SIGINT, &act, NULL);
109 memset(&act, 0, sizeof(act));
110 act.sa_handler = sig_int;
111 act.sa_flags = SA_RESTART;
112 sigaction(SIGTERM, &act, NULL);
114 /* Windows uses SIGBREAK as a quit signal from other applications */
116 memset(&act, 0, sizeof(act));
117 act.sa_handler = sig_int;
118 act.sa_flags = SA_RESTART;
119 sigaction(SIGBREAK, &act, NULL);
122 memset(&act, 0, sizeof(act));
123 act.sa_handler = sig_show_status;
124 act.sa_flags = SA_RESTART;
125 sigaction(SIGUSR1, &act, NULL);
128 memset(&act, 0, sizeof(act));
129 act.sa_handler = sig_int;
130 act.sa_flags = SA_RESTART;
131 sigaction(SIGPIPE, &act, NULL);
136 * Check if we are above the minimum rate given.
138 static int __check_min_rate(struct thread_data *td, struct timeval *now,
141 unsigned long long bytes = 0;
142 unsigned long iops = 0;
145 unsigned int ratemin = 0;
146 unsigned int rate_iops = 0;
147 unsigned int rate_iops_min = 0;
149 assert(ddir_rw(ddir));
151 if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
155 * allow a 2 second settle period in the beginning
157 if (mtime_since(&td->start, now) < 2000)
160 iops += td->this_io_blocks[ddir];
161 bytes += td->this_io_bytes[ddir];
162 ratemin += td->o.ratemin[ddir];
163 rate_iops += td->o.rate_iops[ddir];
164 rate_iops_min += td->o.rate_iops_min[ddir];
167 * if rate blocks is set, sample is running
169 if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
170 spent = mtime_since(&td->lastrate[ddir], now);
171 if (spent < td->o.ratecycle)
174 if (td->o.rate[ddir]) {
176 * check bandwidth specified rate
178 if (bytes < td->rate_bytes[ddir]) {
179 log_err("%s: min rate %u not met\n", td->o.name,
183 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
184 if (rate < ratemin ||
185 bytes < td->rate_bytes[ddir]) {
186 log_err("%s: min rate %u not met, got"
187 " %luKB/sec\n", td->o.name,
194 * checks iops specified rate
196 if (iops < rate_iops) {
197 log_err("%s: min iops rate %u not met\n",
198 td->o.name, rate_iops);
201 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
202 if (rate < rate_iops_min ||
203 iops < td->rate_blocks[ddir]) {
204 log_err("%s: min iops rate %u not met,"
205 " got %lu\n", td->o.name,
206 rate_iops_min, rate);
212 td->rate_bytes[ddir] = bytes;
213 td->rate_blocks[ddir] = iops;
214 memcpy(&td->lastrate[ddir], now, sizeof(*now));
218 static int check_min_rate(struct thread_data *td, struct timeval *now,
219 unsigned long *bytes_done)
223 if (bytes_done[DDIR_READ])
224 ret |= __check_min_rate(td, now, DDIR_READ);
225 if (bytes_done[DDIR_WRITE])
226 ret |= __check_min_rate(td, now, DDIR_WRITE);
227 if (bytes_done[DDIR_TRIM])
228 ret |= __check_min_rate(td, now, DDIR_TRIM);
234 * When job exits, we can cancel the in-flight IO if we are using async
235 * io. Attempt to do so.
237 static void cleanup_pending_aio(struct thread_data *td)
239 struct flist_head *entry, *n;
244 * get immediately available events, if any
246 r = io_u_queued_complete(td, 0, NULL);
251 * now cancel remaining active events
253 if (td->io_ops->cancel) {
254 flist_for_each_safe(entry, n, &td->io_u_busylist) {
255 io_u = flist_entry(entry, struct io_u, list);
258 * if the io_u isn't in flight, then that generally
259 * means someone leaked an io_u. complain but fix
260 * it up, so we don't stall here.
262 if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
263 log_err("fio: non-busy IO on busy list\n");
266 r = td->io_ops->cancel(td, io_u);
274 r = io_u_queued_complete(td, td->cur_depth, NULL);
278 * Helper to handle the final sync of a file. Works just like the normal
279 * io path, just does everything sync.
281 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
283 struct io_u *io_u = __get_io_u(td);
289 io_u->ddir = DDIR_SYNC;
292 if (td_io_prep(td, io_u)) {
298 ret = td_io_queue(td, io_u);
300 td_verror(td, io_u->error, "td_io_queue");
303 } else if (ret == FIO_Q_QUEUED) {
304 if (io_u_queued_complete(td, 1, NULL) < 0)
306 } else if (ret == FIO_Q_COMPLETED) {
308 td_verror(td, io_u->error, "td_io_queue");
312 if (io_u_sync_complete(td, io_u, NULL) < 0)
314 } else if (ret == FIO_Q_BUSY) {
315 if (td_io_commit(td))
323 static inline void __update_tv_cache(struct thread_data *td)
325 fio_gettime(&td->tv_cache, NULL);
328 static inline void update_tv_cache(struct thread_data *td)
330 if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
331 __update_tv_cache(td);
334 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
336 if (in_ramp_time(td))
340 if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
346 static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
351 if (ret < 0 || td->error) {
353 enum error_type_bit eb;
358 eb = td_error_type(ddir, err);
359 if (!(td->o.continue_on_error & (1 << eb)))
362 if (td_non_fatal_error(td, eb, err)) {
364 * Continue with the I/Os in case of
367 update_error_count(td, err);
371 } else if (td->o.fill_device && err == ENOSPC) {
373 * We expect to hit this error if
374 * fill_device option is set.
381 * Stop the I/O in case of a fatal
384 update_error_count(td, err);
393 * The main verify engine. Runs over the writes we previously submitted,
394 * reads the blocks back in, and checks the crc/md5 of the data.
396 static void do_verify(struct thread_data *td)
403 dprint(FD_VERIFY, "starting loop\n");
406 * sync io first and invalidate cache, to make sure we really
409 for_each_file(td, f, i) {
410 if (!fio_file_open(f))
412 if (fio_io_sync(td, f))
414 if (file_invalidate_cache(td, f))
421 td_set_runstate(td, TD_VERIFYING);
424 while (!td->terminate) {
429 if (runtime_exceeded(td, &td->tv_cache)) {
430 __update_tv_cache(td);
431 if (runtime_exceeded(td, &td->tv_cache)) {
437 if (flow_threshold_exceeded(td))
440 io_u = __get_io_u(td);
444 if (get_next_verify(td, io_u)) {
449 if (td_io_prep(td, io_u)) {
454 if (td->o.verify_async)
455 io_u->end_io = verify_io_u_async;
457 io_u->end_io = verify_io_u;
459 ret = td_io_queue(td, io_u);
461 case FIO_Q_COMPLETED:
464 clear_io_u(td, io_u);
465 } else if (io_u->resid) {
466 int bytes = io_u->xfer_buflen - io_u->resid;
472 td_verror(td, EIO, "full resid");
477 io_u->xfer_buflen = io_u->resid;
478 io_u->xfer_buf += bytes;
479 io_u->offset += bytes;
481 if (ddir_rw(io_u->ddir))
482 td->ts.short_io_u[io_u->ddir]++;
485 if (io_u->offset == f->real_file_size)
488 requeue_io_u(td, &io_u);
491 ret = io_u_sync_complete(td, io_u, NULL);
499 requeue_io_u(td, &io_u);
500 ret2 = td_io_commit(td);
506 td_verror(td, -ret, "td_io_queue");
510 if (break_on_this_error(td, io_u->ddir, &ret))
514 * if we can queue more, do so. but check if there are
515 * completed io_u's first. Note that we can get BUSY even
516 * without IO queued, if the system is resource starved.
518 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
519 if (full || !td->o.iodepth_batch_complete) {
520 min_events = min(td->o.iodepth_batch_complete,
523 * if the queue is full, we MUST reap at least 1 event
525 if (full && !min_events)
530 * Reap required number of io units, if any,
531 * and do the verification on them through
532 * the callback handler
534 if (io_u_queued_complete(td, min_events, NULL) < 0) {
538 } while (full && (td->cur_depth > td->o.iodepth_low));
545 min_events = td->cur_depth;
548 ret = io_u_queued_complete(td, min_events, NULL);
550 cleanup_pending_aio(td);
552 td_set_runstate(td, TD_RUNNING);
554 dprint(FD_VERIFY, "exiting loop\n");
557 static int io_bytes_exceeded(struct thread_data *td)
559 unsigned long long bytes;
562 bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE];
563 else if (td_write(td))
564 bytes = td->this_io_bytes[DDIR_WRITE];
565 else if (td_read(td))
566 bytes = td->this_io_bytes[DDIR_READ];
568 bytes = td->this_io_bytes[DDIR_TRIM];
570 return bytes >= td->o.size;
574 * Main IO worker function. It retrieves io_u's to process and queues
575 * and reaps them, checking for rate and errors along the way.
577 static void do_io(struct thread_data *td)
582 if (in_ramp_time(td))
583 td_set_runstate(td, TD_RAMP);
585 td_set_runstate(td, TD_RUNNING);
587 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
588 (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td) ||
590 struct timeval comp_time;
591 unsigned long bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
597 if (td->terminate || td->done)
602 if (runtime_exceeded(td, &td->tv_cache)) {
603 __update_tv_cache(td);
604 if (runtime_exceeded(td, &td->tv_cache)) {
610 if (flow_threshold_exceeded(td))
620 * Add verification end_io handler if:
621 * - Asked to verify (!td_rw(td))
622 * - Or the io_u is from our verify list (mixed write/ver)
624 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
625 ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
626 if (td->o.verify_async)
627 io_u->end_io = verify_io_u_async;
629 io_u->end_io = verify_io_u;
630 td_set_runstate(td, TD_VERIFYING);
631 } else if (in_ramp_time(td))
632 td_set_runstate(td, TD_RAMP);
634 td_set_runstate(td, TD_RUNNING);
636 ret = td_io_queue(td, io_u);
638 case FIO_Q_COMPLETED:
641 clear_io_u(td, io_u);
642 } else if (io_u->resid) {
643 int bytes = io_u->xfer_buflen - io_u->resid;
644 struct fio_file *f = io_u->file;
650 td_verror(td, EIO, "full resid");
655 io_u->xfer_buflen = io_u->resid;
656 io_u->xfer_buf += bytes;
657 io_u->offset += bytes;
659 if (ddir_rw(io_u->ddir))
660 td->ts.short_io_u[io_u->ddir]++;
662 if (io_u->offset == f->real_file_size)
665 requeue_io_u(td, &io_u);
668 if (__should_check_rate(td, DDIR_READ) ||
669 __should_check_rate(td, DDIR_WRITE) ||
670 __should_check_rate(td, DDIR_TRIM))
671 fio_gettime(&comp_time, NULL);
673 ret = io_u_sync_complete(td, io_u, bytes_done);
680 * if the engine doesn't have a commit hook,
681 * the io_u is really queued. if it does have such
682 * a hook, it has to call io_u_queued() itself.
684 if (td->io_ops->commit == NULL)
685 io_u_queued(td, io_u);
688 requeue_io_u(td, &io_u);
689 ret2 = td_io_commit(td);
699 if (break_on_this_error(td, ddir, &ret))
703 * See if we need to complete some commands. Note that we
704 * can get BUSY even without IO queued, if the system is
707 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
708 if (full || !td->o.iodepth_batch_complete) {
709 min_evts = min(td->o.iodepth_batch_complete,
712 * if the queue is full, we MUST reap at least 1 event
714 if (full && !min_evts)
717 if (__should_check_rate(td, DDIR_READ) ||
718 __should_check_rate(td, DDIR_WRITE) ||
719 __should_check_rate(td, DDIR_TRIM))
720 fio_gettime(&comp_time, NULL);
723 ret = io_u_queued_complete(td, min_evts, bytes_done);
727 } while (full && (td->cur_depth > td->o.iodepth_low));
732 if (!ddir_rw_sum(bytes_done) && !(td->io_ops->flags & FIO_NOIO))
735 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
736 if (check_min_rate(td, &comp_time, bytes_done)) {
737 if (exitall_on_terminate)
738 fio_terminate_threads(td->groupid);
739 td_verror(td, EIO, "check_min_rate");
744 if (td->o.thinktime) {
745 unsigned long long b;
747 b = ddir_rw_sum(td->io_blocks);
748 if (!(b % td->o.thinktime_blocks)) {
751 if (td->o.thinktime_spin)
752 usec_spin(td->o.thinktime_spin);
754 left = td->o.thinktime - td->o.thinktime_spin;
756 usec_sleep(td, left);
761 if (td->trim_entries)
762 log_err("fio: %d trim entries leaked?\n", td->trim_entries);
764 if (td->o.fill_device && td->error == ENOSPC) {
773 ret = io_u_queued_complete(td, i, NULL);
774 if (td->o.fill_device && td->error == ENOSPC)
778 if (should_fsync(td) && td->o.end_fsync) {
779 td_set_runstate(td, TD_FSYNCING);
781 for_each_file(td, f, i) {
782 if (!fio_file_open(f))
788 cleanup_pending_aio(td);
791 * stop job if we failed doing any IO
793 if (!ddir_rw_sum(td->this_io_bytes))
797 static void cleanup_io_u(struct thread_data *td)
799 struct flist_head *entry, *n;
802 flist_for_each_safe(entry, n, &td->io_u_freelist) {
803 io_u = flist_entry(entry, struct io_u, list);
805 flist_del(&io_u->list);
806 fio_memfree(io_u, sizeof(*io_u));
812 static int init_io_u(struct thread_data *td)
815 unsigned int max_bs, min_write;
816 int cl_align, i, max_units;
820 max_units = td->o.iodepth;
821 max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
822 max_bs = max(td->o.max_bs[DDIR_TRIM], max_bs);
823 min_write = td->o.min_bs[DDIR_WRITE];
824 td->orig_buffer_size = (unsigned long long) max_bs
825 * (unsigned long long) max_units;
827 if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
830 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
833 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
834 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
837 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
838 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
842 if (data_xfer && allocate_io_mem(td))
845 if (td->o.odirect || td->o.mem_align ||
846 (td->io_ops->flags & FIO_RAWIO))
847 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
851 cl_align = os_cache_line_size();
853 for (i = 0; i < max_units; i++) {
859 ptr = fio_memalign(cl_align, sizeof(*io_u));
861 log_err("fio: unable to allocate aligned memory\n");
866 memset(io_u, 0, sizeof(*io_u));
867 INIT_FLIST_HEAD(&io_u->list);
868 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
872 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
875 io_u_fill_buffer(td, io_u, min_write, max_bs);
876 if (td_write(td) && td->o.verify_pattern_bytes) {
878 * Fill the buffer with the pattern if we are
879 * going to be doing writes.
881 fill_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
886 io_u->flags = IO_U_F_FREE;
887 flist_add(&io_u->list, &td->io_u_freelist);
894 static int switch_ioscheduler(struct thread_data *td)
896 char tmp[256], tmp2[128];
900 if (td->io_ops->flags & FIO_DISKLESSIO)
903 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
905 f = fopen(tmp, "r+");
907 if (errno == ENOENT) {
908 log_err("fio: os or kernel doesn't support IO scheduler"
912 td_verror(td, errno, "fopen iosched");
919 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
920 if (ferror(f) || ret != 1) {
921 td_verror(td, errno, "fwrite");
929 * Read back and check that the selected scheduler is now the default.
931 ret = fread(tmp, 1, sizeof(tmp), f);
932 if (ferror(f) || ret < 0) {
933 td_verror(td, errno, "fread");
938 sprintf(tmp2, "[%s]", td->o.ioscheduler);
939 if (!strstr(tmp, tmp2)) {
940 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
941 td_verror(td, EINVAL, "iosched_switch");
950 static int keep_running(struct thread_data *td)
954 if (td->o.time_based)
961 if (ddir_rw_sum(td->io_bytes) < td->o.size)
967 static int exec_string(const char *string)
969 int ret, newlen = strlen(string) + 1 + 8;
972 str = malloc(newlen);
973 sprintf(str, "sh -c %s", string);
977 log_err("fio: exec of cmd <%s> failed\n", str);
984 * Entry point for the thread based jobs. The process based jobs end up
985 * here as well, after a little setup.
987 static void *thread_main(void *data)
989 unsigned long long elapsed;
990 struct thread_data *td = data;
991 pthread_condattr_t attr;
994 if (!td->o.use_thread) {
1000 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1002 INIT_FLIST_HEAD(&td->io_u_freelist);
1003 INIT_FLIST_HEAD(&td->io_u_busylist);
1004 INIT_FLIST_HEAD(&td->io_u_requeues);
1005 INIT_FLIST_HEAD(&td->io_log_list);
1006 INIT_FLIST_HEAD(&td->io_hist_list);
1007 INIT_FLIST_HEAD(&td->verify_list);
1008 INIT_FLIST_HEAD(&td->trim_list);
1009 pthread_mutex_init(&td->io_u_lock, NULL);
1010 td->io_hist_tree = RB_ROOT;
1012 pthread_condattr_init(&attr);
1013 pthread_cond_init(&td->verify_cond, &attr);
1014 pthread_cond_init(&td->free_cond, &attr);
1016 td_set_runstate(td, TD_INITIALIZED);
1017 dprint(FD_MUTEX, "up startup_mutex\n");
1018 fio_mutex_up(startup_mutex);
1019 dprint(FD_MUTEX, "wait on td->mutex\n");
1020 fio_mutex_down(td->mutex);
1021 dprint(FD_MUTEX, "done waiting on td->mutex\n");
1024 * the ->mutex mutex is now no longer used, close it to avoid
1025 * eating a file descriptor
1027 fio_mutex_remove(td->mutex);
1030 * A new gid requires privilege, so we need to do this before setting
1033 if (td->o.gid != -1U && setgid(td->o.gid)) {
1034 td_verror(td, errno, "setgid");
1037 if (td->o.uid != -1U && setuid(td->o.uid)) {
1038 td_verror(td, errno, "setuid");
1043 * If we have a gettimeofday() thread, make sure we exclude that
1044 * thread from this job
1047 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1050 * Set affinity first, in case it has an impact on the memory
1053 if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1054 td_verror(td, errno, "cpu_set_affinity");
1058 #ifdef FIO_HAVE_LIBNUMA
1059 /* numa node setup */
1060 if (td->o.numa_cpumask_set || td->o.numa_memmask_set) {
1063 if (numa_available() < 0) {
1064 td_verror(td, errno, "Does not support NUMA API\n");
1068 if (td->o.numa_cpumask_set) {
1069 ret = numa_run_on_node_mask(td->o.numa_cpunodesmask);
1071 td_verror(td, errno, \
1072 "numa_run_on_node_mask failed\n");
1077 if (td->o.numa_memmask_set) {
1079 switch (td->o.numa_mem_mode) {
1080 case MPOL_INTERLEAVE:
1081 numa_set_interleave_mask(td->o.numa_memnodesmask);
1084 numa_set_membind(td->o.numa_memnodesmask);
1087 numa_set_localalloc();
1089 case MPOL_PREFERRED:
1090 numa_set_preferred(td->o.numa_mem_prefer_node);
1102 * May alter parameters that init_io_u() will use, so we need to
1111 if (td->o.verify_async && verify_async_init(td))
1114 if (td->ioprio_set) {
1115 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1116 td_verror(td, errno, "ioprio_set");
1121 if (td->o.cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1125 if (nice(td->o.nice) == -1 && errno != 0) {
1126 td_verror(td, errno, "nice");
1130 if (td->o.ioscheduler && switch_ioscheduler(td))
1133 if (!td->o.create_serialize && setup_files(td))
1139 if (init_random_map(td))
1142 if (td->o.exec_prerun) {
1143 if (exec_string(td->o.exec_prerun))
1147 if (td->o.pre_read) {
1148 if (pre_read_files(td) < 0)
1152 fio_gettime(&td->epoch, NULL);
1153 getrusage(RUSAGE_SELF, &td->ru_start);
1156 while (keep_running(td)) {
1157 fio_gettime(&td->start, NULL);
1158 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1159 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1160 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1162 if (td->o.ratemin[DDIR_READ] || td->o.ratemin[DDIR_WRITE] ||
1163 td->o.ratemin[DDIR_TRIM]) {
1164 memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1165 sizeof(td->bw_sample_time));
1166 memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1167 sizeof(td->bw_sample_time));
1168 memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1169 sizeof(td->bw_sample_time));
1175 prune_io_piece_log(td);
1181 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1182 elapsed = utime_since_now(&td->start);
1183 td->ts.runtime[DDIR_READ] += elapsed;
1185 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1186 elapsed = utime_since_now(&td->start);
1187 td->ts.runtime[DDIR_WRITE] += elapsed;
1189 if (td_trim(td) && td->io_bytes[DDIR_TRIM]) {
1190 elapsed = utime_since_now(&td->start);
1191 td->ts.runtime[DDIR_TRIM] += elapsed;
1194 if (td->error || td->terminate)
1197 if (!td->o.do_verify ||
1198 td->o.verify == VERIFY_NONE ||
1199 (td->io_ops->flags & FIO_UNIDIR))
1204 fio_gettime(&td->start, NULL);
1208 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1210 if (td->error || td->terminate)
1214 update_rusage_stat(td);
1215 td->ts.runtime[DDIR_READ] = (td->ts.runtime[DDIR_READ] + 999) / 1000;
1216 td->ts.runtime[DDIR_WRITE] = (td->ts.runtime[DDIR_WRITE] + 999) / 1000;
1217 td->ts.runtime[DDIR_TRIM] = (td->ts.runtime[DDIR_TRIM] + 999) / 1000;
1218 td->ts.total_run_time = mtime_since_now(&td->epoch);
1219 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1220 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1221 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1223 fio_mutex_down(writeout_mutex);
1225 if (td->o.bw_log_file) {
1226 finish_log_named(td, td->bw_log,
1227 td->o.bw_log_file, "bw");
1229 finish_log(td, td->bw_log, "bw");
1232 if (td->o.lat_log_file) {
1233 finish_log_named(td, td->lat_log,
1234 td->o.lat_log_file, "lat");
1236 finish_log(td, td->lat_log, "lat");
1239 if (td->o.lat_log_file) {
1240 finish_log_named(td, td->slat_log,
1241 td->o.lat_log_file, "slat");
1243 finish_log(td, td->slat_log, "slat");
1246 if (td->o.lat_log_file) {
1247 finish_log_named(td, td->clat_log,
1248 td->o.lat_log_file, "clat");
1250 finish_log(td, td->clat_log, "clat");
1253 if (td->o.iops_log_file) {
1254 finish_log_named(td, td->iops_log,
1255 td->o.iops_log_file, "iops");
1257 finish_log(td, td->iops_log, "iops");
1260 fio_mutex_up(writeout_mutex);
1261 if (td->o.exec_postrun)
1262 exec_string(td->o.exec_postrun);
1264 if (exitall_on_terminate)
1265 fio_terminate_threads(td->groupid);
1269 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1272 if (td->o.verify_async)
1273 verify_async_exit(td);
1275 close_and_free_files(td);
1278 cgroup_shutdown(td, &cgroup_mnt);
1280 if (td->o.cpumask_set) {
1281 int ret = fio_cpuset_exit(&td->o.cpumask);
1283 td_verror(td, ret, "fio_cpuset_exit");
1287 * do this very late, it will log file closing as well
1289 if (td->o.write_iolog_file)
1290 write_iolog_close(td);
1292 td_set_runstate(td, TD_EXITED);
1293 return (void *) (uintptr_t) td->error;
1298 * We cannot pass the td data into a forked process, so attach the td and
1299 * pass it to the thread worker.
1301 static int fork_main(int shmid, int offset)
1303 struct thread_data *td;
1307 data = shmat(shmid, NULL, 0);
1308 if (data == (void *) -1) {
1316 * HP-UX inherits shm mappings?
1321 td = data + offset * sizeof(struct thread_data);
1322 ret = thread_main(td);
1324 return (int) (uintptr_t) ret;
1328 * Run over the job map and reap the threads that have exited, if any.
1330 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1331 unsigned int *m_rate)
1333 struct thread_data *td;
1334 unsigned int cputhreads, realthreads, pending;
1338 * reap exited threads (TD_EXITED -> TD_REAPED)
1340 realthreads = pending = cputhreads = 0;
1341 for_each_td(td, i) {
1345 * ->io_ops is NULL for a thread that has closed its
1348 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1357 if (td->runstate == TD_REAPED)
1359 if (td->o.use_thread) {
1360 if (td->runstate == TD_EXITED) {
1361 td_set_runstate(td, TD_REAPED);
1368 if (td->runstate == TD_EXITED)
1372 * check if someone quit or got killed in an unusual way
1374 ret = waitpid(td->pid, &status, flags);
1376 if (errno == ECHILD) {
1377 log_err("fio: pid=%d disappeared %d\n",
1378 (int) td->pid, td->runstate);
1380 td_set_runstate(td, TD_REAPED);
1384 } else if (ret == td->pid) {
1385 if (WIFSIGNALED(status)) {
1386 int sig = WTERMSIG(status);
1389 log_err("fio: pid=%d, got signal=%d\n",
1390 (int) td->pid, sig);
1392 td_set_runstate(td, TD_REAPED);
1395 if (WIFEXITED(status)) {
1396 if (WEXITSTATUS(status) && !td->error)
1397 td->error = WEXITSTATUS(status);
1399 td_set_runstate(td, TD_REAPED);
1405 * thread is not dead, continue
1411 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1412 (*t_rate) -= ddir_rw_sum(td->o.rate);
1419 done_secs += mtime_since_now(&td->epoch) / 1000;
1422 if (*nr_running == cputhreads && !pending && realthreads)
1423 fio_terminate_threads(TERMINATE_ALL);
1427 * Main function for kicking off and reaping jobs, as needed.
1429 static void run_threads(void)
1431 struct thread_data *td;
1432 unsigned long spent;
1433 unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
1435 if (fio_pin_memory())
1438 if (fio_gtod_offload && fio_start_gtod_thread())
1443 if (output_format == FIO_OUTPUT_NORMAL) {
1444 log_info("Starting ");
1446 log_info("%d thread%s", nr_thread,
1447 nr_thread > 1 ? "s" : "");
1451 log_info("%d process%s", nr_process,
1452 nr_process > 1 ? "es" : "");
1458 todo = thread_number;
1461 m_rate = t_rate = 0;
1463 for_each_td(td, i) {
1464 print_status_init(td->thread_number - 1);
1466 if (!td->o.create_serialize)
1470 * do file setup here so it happens sequentially,
1471 * we don't want X number of threads getting their
1472 * client data interspersed on disk
1474 if (setup_files(td)) {
1477 log_err("fio: pid=%d, err=%d/%s\n",
1478 (int) td->pid, td->error, td->verror);
1479 td_set_runstate(td, TD_REAPED);
1486 * for sharing to work, each job must always open
1487 * its own files. so close them, if we opened them
1490 for_each_file(td, f, j) {
1491 if (fio_file_open(f))
1492 td_io_close_file(td, f);
1500 struct thread_data *map[REAL_MAX_JOBS];
1501 struct timeval this_start;
1502 int this_jobs = 0, left;
1505 * create threads (TD_NOT_CREATED -> TD_CREATED)
1507 for_each_td(td, i) {
1508 if (td->runstate != TD_NOT_CREATED)
1512 * never got a chance to start, killed by other
1513 * thread for some reason
1515 if (td->terminate) {
1520 if (td->o.start_delay) {
1521 spent = mtime_since_genesis();
1523 if (td->o.start_delay * 1000 > spent)
1527 if (td->o.stonewall && (nr_started || nr_running)) {
1528 dprint(FD_PROCESS, "%s: stonewall wait\n",
1536 * Set state to created. Thread will transition
1537 * to TD_INITIALIZED when it's done setting up.
1539 td_set_runstate(td, TD_CREATED);
1540 map[this_jobs++] = td;
1543 if (td->o.use_thread) {
1546 dprint(FD_PROCESS, "will pthread_create\n");
1547 ret = pthread_create(&td->thread, NULL,
1550 log_err("pthread_create: %s\n",
1555 ret = pthread_detach(td->thread);
1557 log_err("pthread_detach: %s",
1561 dprint(FD_PROCESS, "will fork\n");
1564 int ret = fork_main(shm_id, i);
1567 } else if (i == fio_debug_jobno)
1568 *fio_debug_jobp = pid;
1570 dprint(FD_MUTEX, "wait on startup_mutex\n");
1571 if (fio_mutex_down_timeout(startup_mutex, 10)) {
1572 log_err("fio: job startup hung? exiting.\n");
1573 fio_terminate_threads(TERMINATE_ALL);
1578 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1582 * Wait for the started threads to transition to
1585 fio_gettime(&this_start, NULL);
1587 while (left && !fio_abort) {
1588 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1593 for (i = 0; i < this_jobs; i++) {
1597 if (td->runstate == TD_INITIALIZED) {
1600 } else if (td->runstate >= TD_EXITED) {
1604 nr_running++; /* work-around... */
1610 log_err("fio: %d job%s failed to start\n", left,
1611 left > 1 ? "s" : "");
1612 for (i = 0; i < this_jobs; i++) {
1616 kill(td->pid, SIGTERM);
1622 * start created threads (TD_INITIALIZED -> TD_RUNNING).
1624 for_each_td(td, i) {
1625 if (td->runstate != TD_INITIALIZED)
1628 if (in_ramp_time(td))
1629 td_set_runstate(td, TD_RAMP);
1631 td_set_runstate(td, TD_RUNNING);
1634 m_rate += ddir_rw_sum(td->o.ratemin);
1635 t_rate += ddir_rw_sum(td->o.rate);
1637 fio_mutex_up(td->mutex);
1640 reap_threads(&nr_running, &t_rate, &m_rate);
1644 fio_server_idle_loop();
1650 while (nr_running) {
1651 reap_threads(&nr_running, &t_rate, &m_rate);
1654 fio_server_idle_loop();
1663 void wait_for_disk_thread_exit(void)
1665 fio_mutex_down(disk_thread_mutex);
1668 static void free_disk_util(void)
1670 disk_util_start_exit();
1671 wait_for_disk_thread_exit();
1672 disk_util_prune_entries();
1675 static void *disk_thread_main(void *data)
1679 fio_mutex_up(startup_mutex);
1681 while (threads && !ret) {
1682 usleep(DISK_UTIL_MSEC * 1000);
1685 ret = update_io_ticks();
1688 print_thread_status();
1691 fio_mutex_up(disk_thread_mutex);
1695 static int create_disk_util_thread(void)
1701 disk_thread_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
1703 ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
1705 fio_mutex_remove(disk_thread_mutex);
1706 log_err("Can't create disk util thread: %s\n", strerror(ret));
1710 ret = pthread_detach(disk_util_thread);
1712 fio_mutex_remove(disk_thread_mutex);
1713 log_err("Can't detatch disk util thread: %s\n", strerror(ret));
1717 dprint(FD_MUTEX, "wait on startup_mutex\n");
1718 fio_mutex_down(startup_mutex);
1719 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1723 int fio_backend(void)
1725 struct thread_data *td;
1729 if (load_profile(exec_profile))
1732 exec_profile = NULL;
1738 setup_log(&agg_io_log[DDIR_READ], 0);
1739 setup_log(&agg_io_log[DDIR_WRITE], 0);
1740 setup_log(&agg_io_log[DDIR_TRIM], 0);
1743 startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
1744 if (startup_mutex == NULL)
1746 writeout_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1747 if (writeout_mutex == NULL)
1751 create_disk_util_thread();
1753 cgroup_list = smalloc(sizeof(*cgroup_list));
1754 INIT_FLIST_HEAD(cgroup_list);
1761 __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1762 __finish_log(agg_io_log[DDIR_WRITE],
1763 "agg-write_bw.log");
1764 __finish_log(agg_io_log[DDIR_TRIM],
1765 "agg-write_bw.log");
1770 fio_options_free(td);
1773 cgroup_kill(cgroup_list);
1777 fio_mutex_remove(startup_mutex);
1778 fio_mutex_remove(writeout_mutex);
1779 fio_mutex_remove(disk_thread_mutex);