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
53 #include "lib/getrusage.h"
56 static pthread_t disk_util_thread;
57 static struct fio_mutex *disk_thread_mutex;
58 static struct fio_mutex *startup_mutex;
59 static struct fio_mutex *writeout_mutex;
60 static struct flist_head *cgroup_list;
61 static char *cgroup_mnt;
62 static int exit_value;
63 static volatile int fio_abort;
65 struct io_log *agg_io_log[DDIR_RWDIR_CNT];
68 unsigned int thread_number = 0;
69 unsigned int stat_number = 0;
70 unsigned int nr_process = 0;
71 unsigned int nr_thread = 0;
74 unsigned long done_secs = 0;
75 volatile int disk_util_exit = 0;
77 #define PAGE_ALIGN(buf) \
78 (char *) (((uintptr_t) (buf) + page_mask) & ~page_mask)
80 #define JOB_START_TIMEOUT (5 * 1000)
82 static void sig_int(int sig)
86 fio_server_got_signal(sig);
88 log_info("\nfio: terminating on signal %d\n", sig);
93 fio_terminate_threads(TERMINATE_ALL);
97 static void sig_show_status(int sig)
99 show_running_run_stats();
102 static void set_sig_handlers(void)
104 struct sigaction act;
106 memset(&act, 0, sizeof(act));
107 act.sa_handler = sig_int;
108 act.sa_flags = SA_RESTART;
109 sigaction(SIGINT, &act, NULL);
111 memset(&act, 0, sizeof(act));
112 act.sa_handler = sig_int;
113 act.sa_flags = SA_RESTART;
114 sigaction(SIGTERM, &act, NULL);
116 /* Windows uses SIGBREAK as a quit signal from other applications */
118 memset(&act, 0, sizeof(act));
119 act.sa_handler = sig_int;
120 act.sa_flags = SA_RESTART;
121 sigaction(SIGBREAK, &act, NULL);
124 memset(&act, 0, sizeof(act));
125 act.sa_handler = sig_show_status;
126 act.sa_flags = SA_RESTART;
127 sigaction(SIGUSR1, &act, NULL);
130 memset(&act, 0, sizeof(act));
131 act.sa_handler = sig_int;
132 act.sa_flags = SA_RESTART;
133 sigaction(SIGPIPE, &act, NULL);
138 * Check if we are above the minimum rate given.
140 static int __check_min_rate(struct thread_data *td, struct timeval *now,
143 unsigned long long bytes = 0;
144 unsigned long iops = 0;
147 unsigned int ratemin = 0;
148 unsigned int rate_iops = 0;
149 unsigned int rate_iops_min = 0;
151 assert(ddir_rw(ddir));
153 if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
157 * allow a 2 second settle period in the beginning
159 if (mtime_since(&td->start, now) < 2000)
162 iops += td->this_io_blocks[ddir];
163 bytes += td->this_io_bytes[ddir];
164 ratemin += td->o.ratemin[ddir];
165 rate_iops += td->o.rate_iops[ddir];
166 rate_iops_min += td->o.rate_iops_min[ddir];
169 * if rate blocks is set, sample is running
171 if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
172 spent = mtime_since(&td->lastrate[ddir], now);
173 if (spent < td->o.ratecycle)
176 if (td->o.rate[ddir]) {
178 * check bandwidth specified rate
180 if (bytes < td->rate_bytes[ddir]) {
181 log_err("%s: min rate %u not met\n", td->o.name,
185 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
186 if (rate < ratemin ||
187 bytes < td->rate_bytes[ddir]) {
188 log_err("%s: min rate %u not met, got"
189 " %luKB/sec\n", td->o.name,
196 * checks iops specified rate
198 if (iops < rate_iops) {
199 log_err("%s: min iops rate %u not met\n",
200 td->o.name, rate_iops);
203 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
204 if (rate < rate_iops_min ||
205 iops < td->rate_blocks[ddir]) {
206 log_err("%s: min iops rate %u not met,"
207 " got %lu\n", td->o.name,
208 rate_iops_min, rate);
214 td->rate_bytes[ddir] = bytes;
215 td->rate_blocks[ddir] = iops;
216 memcpy(&td->lastrate[ddir], now, sizeof(*now));
220 static int check_min_rate(struct thread_data *td, struct timeval *now,
221 uint64_t *bytes_done)
225 if (bytes_done[DDIR_READ])
226 ret |= __check_min_rate(td, now, DDIR_READ);
227 if (bytes_done[DDIR_WRITE])
228 ret |= __check_min_rate(td, now, DDIR_WRITE);
229 if (bytes_done[DDIR_TRIM])
230 ret |= __check_min_rate(td, now, DDIR_TRIM);
236 * When job exits, we can cancel the in-flight IO if we are using async
237 * io. Attempt to do so.
239 static void cleanup_pending_aio(struct thread_data *td)
241 struct flist_head *entry, *n;
246 * get immediately available events, if any
248 r = io_u_queued_complete(td, 0, NULL);
253 * now cancel remaining active events
255 if (td->io_ops->cancel) {
256 flist_for_each_safe(entry, n, &td->io_u_busylist) {
257 io_u = flist_entry(entry, struct io_u, list);
260 * if the io_u isn't in flight, then that generally
261 * means someone leaked an io_u. complain but fix
262 * it up, so we don't stall here.
264 if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
265 log_err("fio: non-busy IO on busy list\n");
268 r = td->io_ops->cancel(td, io_u);
276 r = io_u_queued_complete(td, td->cur_depth, NULL);
280 * Helper to handle the final sync of a file. Works just like the normal
281 * io path, just does everything sync.
283 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
285 struct io_u *io_u = __get_io_u(td);
291 io_u->ddir = DDIR_SYNC;
294 if (td_io_prep(td, io_u)) {
300 ret = td_io_queue(td, io_u);
302 td_verror(td, io_u->error, "td_io_queue");
305 } else if (ret == FIO_Q_QUEUED) {
306 if (io_u_queued_complete(td, 1, NULL) < 0)
308 } else if (ret == FIO_Q_COMPLETED) {
310 td_verror(td, io_u->error, "td_io_queue");
314 if (io_u_sync_complete(td, io_u, NULL) < 0)
316 } else if (ret == FIO_Q_BUSY) {
317 if (td_io_commit(td))
325 static inline void __update_tv_cache(struct thread_data *td)
327 fio_gettime(&td->tv_cache, NULL);
330 static inline void update_tv_cache(struct thread_data *td)
332 if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
333 __update_tv_cache(td);
336 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
338 if (in_ramp_time(td))
342 if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
348 static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
353 if (ret < 0 || td->error) {
355 enum error_type_bit eb;
360 eb = td_error_type(ddir, err);
361 if (!(td->o.continue_on_error & (1 << eb)))
364 if (td_non_fatal_error(td, eb, err)) {
366 * Continue with the I/Os in case of
369 update_error_count(td, err);
373 } else if (td->o.fill_device && err == ENOSPC) {
375 * We expect to hit this error if
376 * fill_device option is set.
383 * Stop the I/O in case of a fatal
386 update_error_count(td, err);
395 * The main verify engine. Runs over the writes we previously submitted,
396 * reads the blocks back in, and checks the crc/md5 of the data.
398 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
400 uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
406 dprint(FD_VERIFY, "starting loop\n");
409 * sync io first and invalidate cache, to make sure we really
412 for_each_file(td, f, i) {
413 if (!fio_file_open(f))
415 if (fio_io_sync(td, f))
417 if (file_invalidate_cache(td, f))
424 td_set_runstate(td, TD_VERIFYING);
427 while (!td->terminate) {
433 if (runtime_exceeded(td, &td->tv_cache)) {
434 __update_tv_cache(td);
435 if (runtime_exceeded(td, &td->tv_cache)) {
441 if (flow_threshold_exceeded(td))
444 if (!td->o.experimental_verify) {
445 io_u = __get_io_u(td);
449 if (get_next_verify(td, io_u)) {
454 if (td_io_prep(td, io_u)) {
459 if (ddir_rw_sum(bytes_done) + td->o.rw_min_bs > verify_bytes)
462 while ((io_u = get_io_u(td)) != NULL) {
464 * We are only interested in the places where
465 * we wrote or trimmed IOs. Turn those into
466 * reads for verification purposes.
468 if (io_u->ddir == DDIR_READ) {
470 * Pretend we issued it for rwmix
473 td->io_issues[DDIR_READ]++;
476 } else if (io_u->ddir == DDIR_TRIM) {
477 io_u->ddir = DDIR_READ;
478 io_u->flags |= IO_U_F_TRIMMED;
480 } else if (io_u->ddir == DDIR_WRITE) {
481 io_u->ddir = DDIR_READ;
493 if (td->o.verify_async)
494 io_u->end_io = verify_io_u_async;
496 io_u->end_io = verify_io_u;
500 ret = td_io_queue(td, io_u);
502 case FIO_Q_COMPLETED:
505 clear_io_u(td, io_u);
506 } else if (io_u->resid) {
507 int bytes = io_u->xfer_buflen - io_u->resid;
513 td_verror(td, EIO, "full resid");
518 io_u->xfer_buflen = io_u->resid;
519 io_u->xfer_buf += bytes;
520 io_u->offset += bytes;
522 if (ddir_rw(io_u->ddir))
523 td->ts.short_io_u[io_u->ddir]++;
526 if (io_u->offset == f->real_file_size)
529 requeue_io_u(td, &io_u);
532 ret = io_u_sync_complete(td, io_u, bytes_done);
540 requeue_io_u(td, &io_u);
541 ret2 = td_io_commit(td);
547 td_verror(td, -ret, "td_io_queue");
551 if (break_on_this_error(td, ddir, &ret))
555 * if we can queue more, do so. but check if there are
556 * completed io_u's first. Note that we can get BUSY even
557 * without IO queued, if the system is resource starved.
559 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
560 if (full || !td->o.iodepth_batch_complete) {
561 min_events = min(td->o.iodepth_batch_complete,
564 * if the queue is full, we MUST reap at least 1 event
566 if (full && !min_events)
571 * Reap required number of io units, if any,
572 * and do the verification on them through
573 * the callback handler
575 if (io_u_queued_complete(td, min_events, bytes_done) < 0) {
579 } while (full && (td->cur_depth > td->o.iodepth_low));
586 min_events = td->cur_depth;
589 ret = io_u_queued_complete(td, min_events, NULL);
591 cleanup_pending_aio(td);
593 td_set_runstate(td, TD_RUNNING);
595 dprint(FD_VERIFY, "exiting loop\n");
598 static int io_bytes_exceeded(struct thread_data *td)
600 unsigned long long bytes;
603 bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE];
604 else if (td_write(td))
605 bytes = td->this_io_bytes[DDIR_WRITE];
606 else if (td_read(td))
607 bytes = td->this_io_bytes[DDIR_READ];
609 bytes = td->this_io_bytes[DDIR_TRIM];
611 return bytes >= td->o.size;
615 * Main IO worker function. It retrieves io_u's to process and queues
616 * and reaps them, checking for rate and errors along the way.
618 * Returns number of bytes written and trimmed.
620 static uint64_t do_io(struct thread_data *td)
622 uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
626 if (in_ramp_time(td))
627 td_set_runstate(td, TD_RAMP);
629 td_set_runstate(td, TD_RUNNING);
631 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
632 (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td) ||
634 struct timeval comp_time;
640 if (td->terminate || td->done)
645 if (runtime_exceeded(td, &td->tv_cache)) {
646 __update_tv_cache(td);
647 if (runtime_exceeded(td, &td->tv_cache)) {
653 if (flow_threshold_exceeded(td))
663 * Add verification end_io handler if:
664 * - Asked to verify (!td_rw(td))
665 * - Or the io_u is from our verify list (mixed write/ver)
667 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
668 ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
669 if (td->o.verify_async)
670 io_u->end_io = verify_io_u_async;
672 io_u->end_io = verify_io_u;
673 td_set_runstate(td, TD_VERIFYING);
674 } else if (in_ramp_time(td))
675 td_set_runstate(td, TD_RAMP);
677 td_set_runstate(td, TD_RUNNING);
679 ret = td_io_queue(td, io_u);
681 case FIO_Q_COMPLETED:
684 clear_io_u(td, io_u);
685 } else if (io_u->resid) {
686 int bytes = io_u->xfer_buflen - io_u->resid;
687 struct fio_file *f = io_u->file;
693 td_verror(td, EIO, "full resid");
698 io_u->xfer_buflen = io_u->resid;
699 io_u->xfer_buf += bytes;
700 io_u->offset += bytes;
702 if (ddir_rw(io_u->ddir))
703 td->ts.short_io_u[io_u->ddir]++;
705 if (io_u->offset == f->real_file_size)
708 requeue_io_u(td, &io_u);
711 if (__should_check_rate(td, DDIR_READ) ||
712 __should_check_rate(td, DDIR_WRITE) ||
713 __should_check_rate(td, DDIR_TRIM))
714 fio_gettime(&comp_time, NULL);
716 ret = io_u_sync_complete(td, io_u, bytes_done);
723 * if the engine doesn't have a commit hook,
724 * the io_u is really queued. if it does have such
725 * a hook, it has to call io_u_queued() itself.
727 if (td->io_ops->commit == NULL)
728 io_u_queued(td, io_u);
731 requeue_io_u(td, &io_u);
732 ret2 = td_io_commit(td);
742 if (break_on_this_error(td, ddir, &ret))
746 * See if we need to complete some commands. Note that we
747 * can get BUSY even without IO queued, if the system is
750 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
751 if (full || !td->o.iodepth_batch_complete) {
752 min_evts = min(td->o.iodepth_batch_complete,
755 * if the queue is full, we MUST reap at least 1 event
757 if (full && !min_evts)
760 if (__should_check_rate(td, DDIR_READ) ||
761 __should_check_rate(td, DDIR_WRITE) ||
762 __should_check_rate(td, DDIR_TRIM))
763 fio_gettime(&comp_time, NULL);
766 ret = io_u_queued_complete(td, min_evts, bytes_done);
770 } while (full && (td->cur_depth > td->o.iodepth_low));
775 if (!ddir_rw_sum(bytes_done) && !(td->io_ops->flags & FIO_NOIO))
778 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
779 if (check_min_rate(td, &comp_time, bytes_done)) {
780 if (exitall_on_terminate)
781 fio_terminate_threads(td->groupid);
782 td_verror(td, EIO, "check_min_rate");
787 if (td->o.thinktime) {
788 unsigned long long b;
790 b = ddir_rw_sum(td->io_blocks);
791 if (!(b % td->o.thinktime_blocks)) {
794 if (td->o.thinktime_spin)
795 usec_spin(td->o.thinktime_spin);
797 left = td->o.thinktime - td->o.thinktime_spin;
799 usec_sleep(td, left);
804 if (td->trim_entries)
805 log_err("fio: %d trim entries leaked?\n", td->trim_entries);
807 if (td->o.fill_device && td->error == ENOSPC) {
816 ret = io_u_queued_complete(td, i, NULL);
817 if (td->o.fill_device && td->error == ENOSPC)
821 if (should_fsync(td) && td->o.end_fsync) {
822 td_set_runstate(td, TD_FSYNCING);
824 for_each_file(td, f, i) {
825 if (!fio_file_open(f))
831 cleanup_pending_aio(td);
834 * stop job if we failed doing any IO
836 if (!ddir_rw_sum(td->this_io_bytes))
839 return bytes_done[DDIR_WRITE] + bytes_done[DDIR_TRIM];
842 static void cleanup_io_u(struct thread_data *td)
844 struct flist_head *entry, *n;
847 flist_for_each_safe(entry, n, &td->io_u_freelist) {
848 io_u = flist_entry(entry, struct io_u, list);
850 flist_del(&io_u->list);
852 if (td->io_ops->io_u_free)
853 td->io_ops->io_u_free(td, io_u);
855 fio_memfree(io_u, sizeof(*io_u));
861 static int init_io_u(struct thread_data *td)
864 unsigned int max_bs, min_write;
865 int cl_align, i, max_units;
869 max_units = td->o.iodepth;
870 max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
871 max_bs = max(td->o.max_bs[DDIR_TRIM], max_bs);
872 min_write = td->o.min_bs[DDIR_WRITE];
873 td->orig_buffer_size = (unsigned long long) max_bs
874 * (unsigned long long) max_units;
876 if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
879 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
882 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
883 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
886 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
887 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
891 if (data_xfer && allocate_io_mem(td))
894 if (td->o.odirect || td->o.mem_align ||
895 (td->io_ops->flags & FIO_RAWIO))
896 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
900 cl_align = os_cache_line_size();
902 for (i = 0; i < max_units; i++) {
908 ptr = fio_memalign(cl_align, sizeof(*io_u));
910 log_err("fio: unable to allocate aligned memory\n");
915 memset(io_u, 0, sizeof(*io_u));
916 INIT_FLIST_HEAD(&io_u->list);
917 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
921 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
924 io_u_fill_buffer(td, io_u, min_write, max_bs);
925 if (td_write(td) && td->o.verify_pattern_bytes) {
927 * Fill the buffer with the pattern if we are
928 * going to be doing writes.
930 fill_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
935 io_u->flags = IO_U_F_FREE;
936 flist_add(&io_u->list, &td->io_u_freelist);
938 if (td->io_ops->io_u_init) {
939 int ret = td->io_ops->io_u_init(td, io_u);
942 log_err("fio: failed to init engine data: %d\n", ret);
953 static int switch_ioscheduler(struct thread_data *td)
955 char tmp[256], tmp2[128];
959 if (td->io_ops->flags & FIO_DISKLESSIO)
962 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
964 f = fopen(tmp, "r+");
966 if (errno == ENOENT) {
967 log_err("fio: os or kernel doesn't support IO scheduler"
971 td_verror(td, errno, "fopen iosched");
978 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
979 if (ferror(f) || ret != 1) {
980 td_verror(td, errno, "fwrite");
988 * Read back and check that the selected scheduler is now the default.
990 ret = fread(tmp, 1, sizeof(tmp), f);
991 if (ferror(f) || ret < 0) {
992 td_verror(td, errno, "fread");
997 sprintf(tmp2, "[%s]", td->o.ioscheduler);
998 if (!strstr(tmp, tmp2)) {
999 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1000 td_verror(td, EINVAL, "iosched_switch");
1009 static int keep_running(struct thread_data *td)
1013 if (td->o.time_based)
1020 if (ddir_rw_sum(td->io_bytes) < td->o.size)
1026 static int exec_string(const char *string)
1028 int ret, newlen = strlen(string) + 1 + 8;
1031 str = malloc(newlen);
1032 sprintf(str, "sh -c %s", string);
1036 log_err("fio: exec of cmd <%s> failed\n", str);
1043 * Entry point for the thread based jobs. The process based jobs end up
1044 * here as well, after a little setup.
1046 static void *thread_main(void *data)
1048 unsigned long long elapsed;
1049 struct thread_data *td = data;
1050 pthread_condattr_t attr;
1053 if (!td->o.use_thread) {
1059 fio_local_clock_init(td->o.use_thread);
1061 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1063 INIT_FLIST_HEAD(&td->io_u_freelist);
1064 INIT_FLIST_HEAD(&td->io_u_busylist);
1065 INIT_FLIST_HEAD(&td->io_u_requeues);
1066 INIT_FLIST_HEAD(&td->io_log_list);
1067 INIT_FLIST_HEAD(&td->io_hist_list);
1068 INIT_FLIST_HEAD(&td->verify_list);
1069 INIT_FLIST_HEAD(&td->trim_list);
1070 INIT_FLIST_HEAD(&td->next_rand_list);
1071 pthread_mutex_init(&td->io_u_lock, NULL);
1072 td->io_hist_tree = RB_ROOT;
1074 pthread_condattr_init(&attr);
1075 pthread_cond_init(&td->verify_cond, &attr);
1076 pthread_cond_init(&td->free_cond, &attr);
1078 td_set_runstate(td, TD_INITIALIZED);
1079 dprint(FD_MUTEX, "up startup_mutex\n");
1080 fio_mutex_up(startup_mutex);
1081 dprint(FD_MUTEX, "wait on td->mutex\n");
1082 fio_mutex_down(td->mutex);
1083 dprint(FD_MUTEX, "done waiting on td->mutex\n");
1086 * the ->mutex mutex is now no longer used, close it to avoid
1087 * eating a file descriptor
1089 fio_mutex_remove(td->mutex);
1092 * A new gid requires privilege, so we need to do this before setting
1095 if (td->o.gid != -1U && setgid(td->o.gid)) {
1096 td_verror(td, errno, "setgid");
1099 if (td->o.uid != -1U && setuid(td->o.uid)) {
1100 td_verror(td, errno, "setuid");
1105 * If we have a gettimeofday() thread, make sure we exclude that
1106 * thread from this job
1109 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1112 * Set affinity first, in case it has an impact on the memory
1115 if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1116 td_verror(td, errno, "cpu_set_affinity");
1120 #ifdef CONFIG_LIBNUMA
1121 /* numa node setup */
1122 if (td->o.numa_cpumask_set || td->o.numa_memmask_set) {
1125 if (numa_available() < 0) {
1126 td_verror(td, errno, "Does not support NUMA API\n");
1130 if (td->o.numa_cpumask_set) {
1131 ret = numa_run_on_node_mask(td->o.numa_cpunodesmask);
1133 td_verror(td, errno, \
1134 "numa_run_on_node_mask failed\n");
1139 if (td->o.numa_memmask_set) {
1141 switch (td->o.numa_mem_mode) {
1142 case MPOL_INTERLEAVE:
1143 numa_set_interleave_mask(td->o.numa_memnodesmask);
1146 numa_set_membind(td->o.numa_memnodesmask);
1149 numa_set_localalloc();
1151 case MPOL_PREFERRED:
1152 numa_set_preferred(td->o.numa_mem_prefer_node);
1164 * May alter parameters that init_io_u() will use, so we need to
1173 if (td->o.verify_async && verify_async_init(td))
1176 if (td->ioprio_set) {
1177 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1178 td_verror(td, errno, "ioprio_set");
1183 if (td->o.cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1187 if (nice(td->o.nice) == -1 && errno != 0) {
1188 td_verror(td, errno, "nice");
1192 if (td->o.ioscheduler && switch_ioscheduler(td))
1195 if (!td->o.create_serialize && setup_files(td))
1201 if (init_random_map(td))
1204 if (td->o.exec_prerun) {
1205 if (exec_string(td->o.exec_prerun))
1209 if (td->o.pre_read) {
1210 if (pre_read_files(td) < 0)
1214 fio_gettime(&td->epoch, NULL);
1215 fio_getrusage(&td->ru_start);
1217 while (keep_running(td)) {
1218 uint64_t verify_bytes;
1220 fio_gettime(&td->start, NULL);
1221 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1222 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1223 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1225 if (td->o.ratemin[DDIR_READ] || td->o.ratemin[DDIR_WRITE] ||
1226 td->o.ratemin[DDIR_TRIM]) {
1227 memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1228 sizeof(td->bw_sample_time));
1229 memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1230 sizeof(td->bw_sample_time));
1231 memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1232 sizeof(td->bw_sample_time));
1238 prune_io_piece_log(td);
1240 verify_bytes = do_io(td);
1244 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1245 elapsed = utime_since_now(&td->start);
1246 td->ts.runtime[DDIR_READ] += elapsed;
1248 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1249 elapsed = utime_since_now(&td->start);
1250 td->ts.runtime[DDIR_WRITE] += elapsed;
1252 if (td_trim(td) && td->io_bytes[DDIR_TRIM]) {
1253 elapsed = utime_since_now(&td->start);
1254 td->ts.runtime[DDIR_TRIM] += elapsed;
1257 if (td->error || td->terminate)
1260 if (!td->o.do_verify ||
1261 td->o.verify == VERIFY_NONE ||
1262 (td->io_ops->flags & FIO_UNIDIR))
1267 fio_gettime(&td->start, NULL);
1269 do_verify(td, verify_bytes);
1271 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1273 if (td->error || td->terminate)
1277 update_rusage_stat(td);
1278 td->ts.runtime[DDIR_READ] = (td->ts.runtime[DDIR_READ] + 999) / 1000;
1279 td->ts.runtime[DDIR_WRITE] = (td->ts.runtime[DDIR_WRITE] + 999) / 1000;
1280 td->ts.runtime[DDIR_TRIM] = (td->ts.runtime[DDIR_TRIM] + 999) / 1000;
1281 td->ts.total_run_time = mtime_since_now(&td->epoch);
1282 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1283 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1284 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1286 fio_mutex_down(writeout_mutex);
1288 if (td->o.bw_log_file) {
1289 finish_log_named(td, td->bw_log,
1290 td->o.bw_log_file, "bw");
1292 finish_log(td, td->bw_log, "bw");
1295 if (td->o.lat_log_file) {
1296 finish_log_named(td, td->lat_log,
1297 td->o.lat_log_file, "lat");
1299 finish_log(td, td->lat_log, "lat");
1302 if (td->o.lat_log_file) {
1303 finish_log_named(td, td->slat_log,
1304 td->o.lat_log_file, "slat");
1306 finish_log(td, td->slat_log, "slat");
1309 if (td->o.lat_log_file) {
1310 finish_log_named(td, td->clat_log,
1311 td->o.lat_log_file, "clat");
1313 finish_log(td, td->clat_log, "clat");
1316 if (td->o.iops_log_file) {
1317 finish_log_named(td, td->iops_log,
1318 td->o.iops_log_file, "iops");
1320 finish_log(td, td->iops_log, "iops");
1323 fio_mutex_up(writeout_mutex);
1324 if (td->o.exec_postrun)
1325 exec_string(td->o.exec_postrun);
1327 if (exitall_on_terminate)
1328 fio_terminate_threads(td->groupid);
1332 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1335 if (td->o.verify_async)
1336 verify_async_exit(td);
1338 close_and_free_files(td);
1341 cgroup_shutdown(td, &cgroup_mnt);
1343 if (td->o.cpumask_set) {
1344 int ret = fio_cpuset_exit(&td->o.cpumask);
1346 td_verror(td, ret, "fio_cpuset_exit");
1350 * do this very late, it will log file closing as well
1352 if (td->o.write_iolog_file)
1353 write_iolog_close(td);
1355 td_set_runstate(td, TD_EXITED);
1356 return (void *) (uintptr_t) td->error;
1361 * We cannot pass the td data into a forked process, so attach the td and
1362 * pass it to the thread worker.
1364 static int fork_main(int shmid, int offset)
1366 struct thread_data *td;
1370 data = shmat(shmid, NULL, 0);
1371 if (data == (void *) -1) {
1379 * HP-UX inherits shm mappings?
1384 td = data + offset * sizeof(struct thread_data);
1385 ret = thread_main(td);
1387 return (int) (uintptr_t) ret;
1391 * Run over the job map and reap the threads that have exited, if any.
1393 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1394 unsigned int *m_rate)
1396 struct thread_data *td;
1397 unsigned int cputhreads, realthreads, pending;
1401 * reap exited threads (TD_EXITED -> TD_REAPED)
1403 realthreads = pending = cputhreads = 0;
1404 for_each_td(td, i) {
1408 * ->io_ops is NULL for a thread that has closed its
1411 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1420 if (td->runstate == TD_REAPED)
1422 if (td->o.use_thread) {
1423 if (td->runstate == TD_EXITED) {
1424 td_set_runstate(td, TD_REAPED);
1431 if (td->runstate == TD_EXITED)
1435 * check if someone quit or got killed in an unusual way
1437 ret = waitpid(td->pid, &status, flags);
1439 if (errno == ECHILD) {
1440 log_err("fio: pid=%d disappeared %d\n",
1441 (int) td->pid, td->runstate);
1443 td_set_runstate(td, TD_REAPED);
1447 } else if (ret == td->pid) {
1448 if (WIFSIGNALED(status)) {
1449 int sig = WTERMSIG(status);
1451 if (sig != SIGTERM && sig != SIGUSR2)
1452 log_err("fio: pid=%d, got signal=%d\n",
1453 (int) td->pid, sig);
1455 td_set_runstate(td, TD_REAPED);
1458 if (WIFEXITED(status)) {
1459 if (WEXITSTATUS(status) && !td->error)
1460 td->error = WEXITSTATUS(status);
1462 td_set_runstate(td, TD_REAPED);
1468 * thread is not dead, continue
1474 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1475 (*t_rate) -= ddir_rw_sum(td->o.rate);
1482 done_secs += mtime_since_now(&td->epoch) / 1000;
1485 if (*nr_running == cputhreads && !pending && realthreads)
1486 fio_terminate_threads(TERMINATE_ALL);
1490 * Main function for kicking off and reaping jobs, as needed.
1492 static void run_threads(void)
1494 struct thread_data *td;
1495 unsigned long spent;
1496 unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
1498 if (fio_pin_memory())
1501 if (fio_gtod_offload && fio_start_gtod_thread())
1504 fio_idle_prof_init();
1508 if (output_format == FIO_OUTPUT_NORMAL) {
1509 log_info("Starting ");
1511 log_info("%d thread%s", nr_thread,
1512 nr_thread > 1 ? "s" : "");
1516 log_info("%d process%s", nr_process,
1517 nr_process > 1 ? "es" : "");
1523 todo = thread_number;
1526 m_rate = t_rate = 0;
1528 for_each_td(td, i) {
1529 print_status_init(td->thread_number - 1);
1531 if (!td->o.create_serialize)
1535 * do file setup here so it happens sequentially,
1536 * we don't want X number of threads getting their
1537 * client data interspersed on disk
1539 if (setup_files(td)) {
1542 log_err("fio: pid=%d, err=%d/%s\n",
1543 (int) td->pid, td->error, td->verror);
1544 td_set_runstate(td, TD_REAPED);
1551 * for sharing to work, each job must always open
1552 * its own files. so close them, if we opened them
1555 for_each_file(td, f, j) {
1556 if (fio_file_open(f))
1557 td_io_close_file(td, f);
1562 /* start idle threads before io threads start to run */
1563 fio_idle_prof_start();
1568 struct thread_data *map[REAL_MAX_JOBS];
1569 struct timeval this_start;
1570 int this_jobs = 0, left;
1573 * create threads (TD_NOT_CREATED -> TD_CREATED)
1575 for_each_td(td, i) {
1576 if (td->runstate != TD_NOT_CREATED)
1580 * never got a chance to start, killed by other
1581 * thread for some reason
1583 if (td->terminate) {
1588 if (td->o.start_delay) {
1589 spent = mtime_since_genesis();
1591 if (td->o.start_delay * 1000 > spent)
1595 if (td->o.stonewall && (nr_started || nr_running)) {
1596 dprint(FD_PROCESS, "%s: stonewall wait\n",
1604 * Set state to created. Thread will transition
1605 * to TD_INITIALIZED when it's done setting up.
1607 td_set_runstate(td, TD_CREATED);
1608 map[this_jobs++] = td;
1611 if (td->o.use_thread) {
1614 dprint(FD_PROCESS, "will pthread_create\n");
1615 ret = pthread_create(&td->thread, NULL,
1618 log_err("pthread_create: %s\n",
1623 ret = pthread_detach(td->thread);
1625 log_err("pthread_detach: %s",
1629 dprint(FD_PROCESS, "will fork\n");
1632 int ret = fork_main(shm_id, i);
1635 } else if (i == fio_debug_jobno)
1636 *fio_debug_jobp = pid;
1638 dprint(FD_MUTEX, "wait on startup_mutex\n");
1639 if (fio_mutex_down_timeout(startup_mutex, 10)) {
1640 log_err("fio: job startup hung? exiting.\n");
1641 fio_terminate_threads(TERMINATE_ALL);
1646 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1650 * Wait for the started threads to transition to
1653 fio_gettime(&this_start, NULL);
1655 while (left && !fio_abort) {
1656 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1661 for (i = 0; i < this_jobs; i++) {
1665 if (td->runstate == TD_INITIALIZED) {
1668 } else if (td->runstate >= TD_EXITED) {
1672 nr_running++; /* work-around... */
1678 log_err("fio: %d job%s failed to start\n", left,
1679 left > 1 ? "s" : "");
1680 for (i = 0; i < this_jobs; i++) {
1684 kill(td->pid, SIGTERM);
1690 * start created threads (TD_INITIALIZED -> TD_RUNNING).
1692 for_each_td(td, i) {
1693 if (td->runstate != TD_INITIALIZED)
1696 if (in_ramp_time(td))
1697 td_set_runstate(td, TD_RAMP);
1699 td_set_runstate(td, TD_RUNNING);
1702 m_rate += ddir_rw_sum(td->o.ratemin);
1703 t_rate += ddir_rw_sum(td->o.rate);
1705 fio_mutex_up(td->mutex);
1708 reap_threads(&nr_running, &t_rate, &m_rate);
1712 fio_server_idle_loop();
1718 while (nr_running) {
1719 reap_threads(&nr_running, &t_rate, &m_rate);
1722 fio_server_idle_loop();
1727 fio_idle_prof_stop();
1733 void wait_for_disk_thread_exit(void)
1735 fio_mutex_down(disk_thread_mutex);
1738 static void free_disk_util(void)
1740 disk_util_start_exit();
1741 wait_for_disk_thread_exit();
1742 disk_util_prune_entries();
1745 static void *disk_thread_main(void *data)
1749 fio_mutex_up(startup_mutex);
1751 while (threads && !ret) {
1752 usleep(DISK_UTIL_MSEC * 1000);
1755 ret = update_io_ticks();
1758 print_thread_status();
1761 fio_mutex_up(disk_thread_mutex);
1765 static int create_disk_util_thread(void)
1771 disk_thread_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
1773 ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
1775 fio_mutex_remove(disk_thread_mutex);
1776 log_err("Can't create disk util thread: %s\n", strerror(ret));
1780 ret = pthread_detach(disk_util_thread);
1782 fio_mutex_remove(disk_thread_mutex);
1783 log_err("Can't detatch disk util thread: %s\n", strerror(ret));
1787 dprint(FD_MUTEX, "wait on startup_mutex\n");
1788 fio_mutex_down(startup_mutex);
1789 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1793 int fio_backend(void)
1795 struct thread_data *td;
1799 if (load_profile(exec_profile))
1802 exec_profile = NULL;
1808 setup_log(&agg_io_log[DDIR_READ], 0);
1809 setup_log(&agg_io_log[DDIR_WRITE], 0);
1810 setup_log(&agg_io_log[DDIR_TRIM], 0);
1813 startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
1814 if (startup_mutex == NULL)
1816 writeout_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1817 if (writeout_mutex == NULL)
1821 create_disk_util_thread();
1823 cgroup_list = smalloc(sizeof(*cgroup_list));
1824 INIT_FLIST_HEAD(cgroup_list);
1831 __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1832 __finish_log(agg_io_log[DDIR_WRITE],
1833 "agg-write_bw.log");
1834 __finish_log(agg_io_log[DDIR_TRIM],
1835 "agg-write_bw.log");
1840 fio_options_free(td);
1843 cgroup_kill(cgroup_list);
1847 fio_mutex_remove(startup_mutex);
1848 fio_mutex_remove(writeout_mutex);
1849 fio_mutex_remove(disk_thread_mutex);