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
40 #ifndef FIO_NO_HAVE_SHM_H
53 #include "lib/getrusage.h"
57 static pthread_t disk_util_thread;
58 static struct fio_mutex *disk_thread_mutex;
59 static struct fio_mutex *startup_mutex;
60 static struct flist_head *cgroup_list;
61 static char *cgroup_mnt;
62 static int exit_value;
63 static volatile int fio_abort;
64 static unsigned int nr_process = 0;
65 static unsigned int nr_thread = 0;
67 struct io_log *agg_io_log[DDIR_RWDIR_CNT];
70 unsigned int thread_number = 0;
71 unsigned int stat_number = 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,
186 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
190 if (rate < ratemin ||
191 bytes < td->rate_bytes[ddir]) {
192 log_err("%s: min rate %u not met, got"
193 " %luKB/sec\n", td->o.name,
200 * checks iops specified rate
202 if (iops < rate_iops) {
203 log_err("%s: min iops rate %u not met\n",
204 td->o.name, rate_iops);
207 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
208 if (rate < rate_iops_min ||
209 iops < td->rate_blocks[ddir]) {
210 log_err("%s: min iops rate %u not met,"
211 " got %lu\n", td->o.name,
212 rate_iops_min, rate);
218 td->rate_bytes[ddir] = bytes;
219 td->rate_blocks[ddir] = iops;
220 memcpy(&td->lastrate[ddir], now, sizeof(*now));
224 static int check_min_rate(struct thread_data *td, struct timeval *now,
225 uint64_t *bytes_done)
229 if (bytes_done[DDIR_READ])
230 ret |= __check_min_rate(td, now, DDIR_READ);
231 if (bytes_done[DDIR_WRITE])
232 ret |= __check_min_rate(td, now, DDIR_WRITE);
233 if (bytes_done[DDIR_TRIM])
234 ret |= __check_min_rate(td, now, DDIR_TRIM);
240 * When job exits, we can cancel the in-flight IO if we are using async
241 * io. Attempt to do so.
243 static void cleanup_pending_aio(struct thread_data *td)
248 * get immediately available events, if any
250 r = io_u_queued_complete(td, 0, NULL);
255 * now cancel remaining active events
257 if (td->io_ops->cancel) {
261 io_u_qiter(&td->io_u_all, io_u, i) {
262 if (io_u->flags & IO_U_F_FLIGHT) {
263 r = td->io_ops->cancel(td, io_u);
271 r = io_u_queued_complete(td, td->cur_depth, NULL);
275 * Helper to handle the final sync of a file. Works just like the normal
276 * io path, just does everything sync.
278 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
280 struct io_u *io_u = __get_io_u(td);
286 io_u->ddir = DDIR_SYNC;
289 if (td_io_prep(td, io_u)) {
295 ret = td_io_queue(td, io_u);
297 td_verror(td, io_u->error, "td_io_queue");
300 } else if (ret == FIO_Q_QUEUED) {
301 if (io_u_queued_complete(td, 1, NULL) < 0)
303 } else if (ret == FIO_Q_COMPLETED) {
305 td_verror(td, io_u->error, "td_io_queue");
309 if (io_u_sync_complete(td, io_u, NULL) < 0)
311 } else if (ret == FIO_Q_BUSY) {
312 if (td_io_commit(td))
320 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
324 if (fio_file_open(f))
325 return fio_io_sync(td, f);
327 if (td_io_open_file(td, f))
330 ret = fio_io_sync(td, f);
331 td_io_close_file(td, f);
335 static inline void __update_tv_cache(struct thread_data *td)
337 fio_gettime(&td->tv_cache, NULL);
340 static inline void update_tv_cache(struct thread_data *td)
342 if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
343 __update_tv_cache(td);
346 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
348 if (in_ramp_time(td))
352 if (utime_since(&td->epoch, t) >= td->o.timeout)
358 static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
363 if (ret < 0 || td->error) {
365 enum error_type_bit eb;
370 eb = td_error_type(ddir, err);
371 if (!(td->o.continue_on_error & (1 << eb)))
374 if (td_non_fatal_error(td, eb, err)) {
376 * Continue with the I/Os in case of
379 update_error_count(td, err);
383 } else if (td->o.fill_device && err == ENOSPC) {
385 * We expect to hit this error if
386 * fill_device option is set.
393 * Stop the I/O in case of a fatal
396 update_error_count(td, err);
404 static void check_update_rusage(struct thread_data *td)
406 if (td->update_rusage) {
407 td->update_rusage = 0;
408 update_rusage_stat(td);
409 fio_mutex_up(td->rusage_sem);
414 * The main verify engine. Runs over the writes we previously submitted,
415 * reads the blocks back in, and checks the crc/md5 of the data.
417 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
419 uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
425 dprint(FD_VERIFY, "starting loop\n");
428 * sync io first and invalidate cache, to make sure we really
431 for_each_file(td, f, i) {
432 if (!fio_file_open(f))
434 if (fio_io_sync(td, f))
436 if (file_invalidate_cache(td, f))
440 check_update_rusage(td);
445 td_set_runstate(td, TD_VERIFYING);
448 while (!td->terminate) {
453 check_update_rusage(td);
455 if (runtime_exceeded(td, &td->tv_cache)) {
456 __update_tv_cache(td);
457 if (runtime_exceeded(td, &td->tv_cache)) {
463 if (flow_threshold_exceeded(td))
466 if (!td->o.experimental_verify) {
467 io_u = __get_io_u(td);
471 if (get_next_verify(td, io_u)) {
476 if (td_io_prep(td, io_u)) {
481 if (ddir_rw_sum(bytes_done) + td->o.rw_min_bs > verify_bytes)
484 while ((io_u = get_io_u(td)) != NULL) {
492 * We are only interested in the places where
493 * we wrote or trimmed IOs. Turn those into
494 * reads for verification purposes.
496 if (io_u->ddir == DDIR_READ) {
498 * Pretend we issued it for rwmix
501 td->io_issues[DDIR_READ]++;
504 } else if (io_u->ddir == DDIR_TRIM) {
505 io_u->ddir = DDIR_READ;
506 io_u->flags |= IO_U_F_TRIMMED;
508 } else if (io_u->ddir == DDIR_WRITE) {
509 io_u->ddir = DDIR_READ;
521 if (td->o.verify_async)
522 io_u->end_io = verify_io_u_async;
524 io_u->end_io = verify_io_u;
528 ret = td_io_queue(td, io_u);
530 case FIO_Q_COMPLETED:
533 clear_io_u(td, io_u);
534 } else if (io_u->resid) {
535 int bytes = io_u->xfer_buflen - io_u->resid;
541 td_verror(td, EIO, "full resid");
546 io_u->xfer_buflen = io_u->resid;
547 io_u->xfer_buf += bytes;
548 io_u->offset += bytes;
550 if (ddir_rw(io_u->ddir))
551 td->ts.short_io_u[io_u->ddir]++;
554 if (io_u->offset == f->real_file_size)
557 requeue_io_u(td, &io_u);
560 ret = io_u_sync_complete(td, io_u, bytes_done);
568 requeue_io_u(td, &io_u);
569 ret2 = td_io_commit(td);
575 td_verror(td, -ret, "td_io_queue");
579 if (break_on_this_error(td, ddir, &ret))
583 * if we can queue more, do so. but check if there are
584 * completed io_u's first. Note that we can get BUSY even
585 * without IO queued, if the system is resource starved.
588 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
589 if (full || !td->o.iodepth_batch_complete) {
590 min_events = min(td->o.iodepth_batch_complete,
593 * if the queue is full, we MUST reap at least 1 event
595 if (full && !min_events)
600 * Reap required number of io units, if any,
601 * and do the verification on them through
602 * the callback handler
604 if (io_u_queued_complete(td, min_events, bytes_done) < 0) {
608 } while (full && (td->cur_depth > td->o.iodepth_low));
614 check_update_rusage(td);
617 min_events = td->cur_depth;
620 ret = io_u_queued_complete(td, min_events, NULL);
622 cleanup_pending_aio(td);
624 td_set_runstate(td, TD_RUNNING);
626 dprint(FD_VERIFY, "exiting loop\n");
629 static unsigned int exceeds_number_ios(struct thread_data *td)
631 unsigned long long number_ios;
633 if (!td->o.number_ios)
636 number_ios = ddir_rw_sum(td->this_io_blocks);
637 number_ios += td->io_u_queued + td->io_u_in_flight;
639 return number_ios >= td->o.number_ios;
642 static int io_bytes_exceeded(struct thread_data *td)
644 unsigned long long bytes;
647 bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE];
648 else if (td_write(td))
649 bytes = td->this_io_bytes[DDIR_WRITE];
650 else if (td_read(td))
651 bytes = td->this_io_bytes[DDIR_READ];
653 bytes = td->this_io_bytes[DDIR_TRIM];
655 return bytes >= td->o.size || exceeds_number_ios(td);
659 * Main IO worker function. It retrieves io_u's to process and queues
660 * and reaps them, checking for rate and errors along the way.
662 * Returns number of bytes written and trimmed.
664 static uint64_t do_io(struct thread_data *td)
666 uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
669 uint64_t total_bytes, bytes_issued = 0;
671 if (in_ramp_time(td))
672 td_set_runstate(td, TD_RAMP);
674 td_set_runstate(td, TD_RUNNING);
679 * If verify_backlog is enabled, we'll run the verify in this
680 * handler as well. For that case, we may need up to twice the
683 total_bytes = td->o.size;
684 if (td->o.verify != VERIFY_NONE &&
685 (td_write(td) && td->o.verify_backlog))
686 total_bytes += td->o.size;
688 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
689 (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td) ||
691 struct timeval comp_time;
697 check_update_rusage(td);
699 if (td->terminate || td->done)
704 if (runtime_exceeded(td, &td->tv_cache)) {
705 __update_tv_cache(td);
706 if (runtime_exceeded(td, &td->tv_cache)) {
712 if (flow_threshold_exceeded(td))
715 if (bytes_issued >= total_bytes)
719 if (IS_ERR_OR_NULL(io_u)) {
720 int err = PTR_ERR(io_u);
727 if (td->o.latency_target)
735 * Add verification end_io handler if:
736 * - Asked to verify (!td_rw(td))
737 * - Or the io_u is from our verify list (mixed write/ver)
739 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
740 ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
742 if (!td->o.verify_pattern_bytes) {
743 io_u->rand_seed = __rand(&td->__verify_state);
744 if (sizeof(int) != sizeof(long *))
745 io_u->rand_seed *= __rand(&td->__verify_state);
748 if (td->o.verify_async)
749 io_u->end_io = verify_io_u_async;
751 io_u->end_io = verify_io_u;
752 td_set_runstate(td, TD_VERIFYING);
753 } else if (in_ramp_time(td))
754 td_set_runstate(td, TD_RAMP);
756 td_set_runstate(td, TD_RUNNING);
759 * Always log IO before it's issued, so we know the specific
760 * order of it. The logged unit will track when the IO has
763 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
765 td->o.verify != VERIFY_NONE &&
766 !td->o.experimental_verify)
767 log_io_piece(td, io_u);
769 ret = td_io_queue(td, io_u);
771 case FIO_Q_COMPLETED:
774 clear_io_u(td, io_u);
775 } else if (io_u->resid) {
776 int bytes = io_u->xfer_buflen - io_u->resid;
777 struct fio_file *f = io_u->file;
779 bytes_issued += bytes;
784 td_verror(td, EIO, "full resid");
789 io_u->xfer_buflen = io_u->resid;
790 io_u->xfer_buf += bytes;
791 io_u->offset += bytes;
793 if (ddir_rw(io_u->ddir))
794 td->ts.short_io_u[io_u->ddir]++;
796 if (io_u->offset == f->real_file_size)
799 requeue_io_u(td, &io_u);
802 if (__should_check_rate(td, DDIR_READ) ||
803 __should_check_rate(td, DDIR_WRITE) ||
804 __should_check_rate(td, DDIR_TRIM))
805 fio_gettime(&comp_time, NULL);
807 ret = io_u_sync_complete(td, io_u, bytes_done);
810 bytes_issued += io_u->xfer_buflen;
815 * if the engine doesn't have a commit hook,
816 * the io_u is really queued. if it does have such
817 * a hook, it has to call io_u_queued() itself.
819 if (td->io_ops->commit == NULL)
820 io_u_queued(td, io_u);
821 bytes_issued += io_u->xfer_buflen;
824 requeue_io_u(td, &io_u);
825 ret2 = td_io_commit(td);
835 if (break_on_this_error(td, ddir, &ret))
839 * See if we need to complete some commands. Note that we
840 * can get BUSY even without IO queued, if the system is
844 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
845 if (full || !td->o.iodepth_batch_complete) {
846 min_evts = min(td->o.iodepth_batch_complete,
849 * if the queue is full, we MUST reap at least 1 event
851 if (full && !min_evts)
854 if (__should_check_rate(td, DDIR_READ) ||
855 __should_check_rate(td, DDIR_WRITE) ||
856 __should_check_rate(td, DDIR_TRIM))
857 fio_gettime(&comp_time, NULL);
860 ret = io_u_queued_complete(td, min_evts, bytes_done);
864 } while (full && (td->cur_depth > td->o.iodepth_low));
869 if (!ddir_rw_sum(bytes_done) && !(td->io_ops->flags & FIO_NOIO))
872 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
873 if (check_min_rate(td, &comp_time, bytes_done)) {
874 if (exitall_on_terminate)
875 fio_terminate_threads(td->groupid);
876 td_verror(td, EIO, "check_min_rate");
880 if (!in_ramp_time(td) && td->o.latency_target)
881 lat_target_check(td);
883 if (td->o.thinktime) {
884 unsigned long long b;
886 b = ddir_rw_sum(td->io_blocks);
887 if (!(b % td->o.thinktime_blocks)) {
892 if (td->o.thinktime_spin)
893 usec_spin(td->o.thinktime_spin);
895 left = td->o.thinktime - td->o.thinktime_spin;
897 usec_sleep(td, left);
902 check_update_rusage(td);
904 if (td->trim_entries)
905 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
907 if (td->o.fill_device && td->error == ENOSPC) {
916 ret = io_u_queued_complete(td, i, bytes_done);
917 if (td->o.fill_device && td->error == ENOSPC)
921 if (should_fsync(td) && td->o.end_fsync) {
922 td_set_runstate(td, TD_FSYNCING);
924 for_each_file(td, f, i) {
925 if (!fio_file_fsync(td, f))
928 log_err("fio: end_fsync failed for file %s\n",
933 cleanup_pending_aio(td);
936 * stop job if we failed doing any IO
938 if (!ddir_rw_sum(td->this_io_bytes))
941 return bytes_done[DDIR_WRITE] + bytes_done[DDIR_TRIM];
944 static void cleanup_io_u(struct thread_data *td)
948 while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
950 if (td->io_ops->io_u_free)
951 td->io_ops->io_u_free(td, io_u);
953 fio_memfree(io_u, sizeof(*io_u));
958 io_u_rexit(&td->io_u_requeues);
959 io_u_qexit(&td->io_u_freelist);
960 io_u_qexit(&td->io_u_all);
963 static int init_io_u(struct thread_data *td)
966 unsigned int max_bs, min_write;
967 int cl_align, i, max_units;
968 int data_xfer = 1, err;
971 max_units = td->o.iodepth;
972 max_bs = td_max_bs(td);
973 min_write = td->o.min_bs[DDIR_WRITE];
974 td->orig_buffer_size = (unsigned long long) max_bs
975 * (unsigned long long) max_units;
977 if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
981 err += io_u_rinit(&td->io_u_requeues, td->o.iodepth);
982 err += io_u_qinit(&td->io_u_freelist, td->o.iodepth);
983 err += io_u_qinit(&td->io_u_all, td->o.iodepth);
986 log_err("fio: failed setting up IO queues\n");
991 * if we may later need to do address alignment, then add any
992 * possible adjustment here so that we don't cause a buffer
993 * overflow later. this adjustment may be too much if we get
994 * lucky and the allocator gives us an aligned address.
996 if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
997 (td->io_ops->flags & FIO_RAWIO))
998 td->orig_buffer_size += page_mask + td->o.mem_align;
1000 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1003 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1004 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1007 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1008 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1012 if (data_xfer && allocate_io_mem(td))
1015 if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1016 (td->io_ops->flags & FIO_RAWIO))
1017 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
1019 p = td->orig_buffer;
1021 cl_align = os_cache_line_size();
1023 for (i = 0; i < max_units; i++) {
1029 ptr = fio_memalign(cl_align, sizeof(*io_u));
1031 log_err("fio: unable to allocate aligned memory\n");
1036 memset(io_u, 0, sizeof(*io_u));
1037 INIT_FLIST_HEAD(&io_u->verify_list);
1038 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1042 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1045 io_u_fill_buffer(td, io_u, min_write, max_bs);
1046 if (td_write(td) && td->o.verify_pattern_bytes) {
1048 * Fill the buffer with the pattern if we are
1049 * going to be doing writes.
1051 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1056 io_u->flags = IO_U_F_FREE;
1057 io_u_qpush(&td->io_u_freelist, io_u);
1060 * io_u never leaves this stack, used for iteration of all
1063 io_u_qpush(&td->io_u_all, io_u);
1065 if (td->io_ops->io_u_init) {
1066 int ret = td->io_ops->io_u_init(td, io_u);
1069 log_err("fio: failed to init engine data: %d\n", ret);
1080 static int switch_ioscheduler(struct thread_data *td)
1082 char tmp[256], tmp2[128];
1086 if (td->io_ops->flags & FIO_DISKLESSIO)
1089 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
1091 f = fopen(tmp, "r+");
1093 if (errno == ENOENT) {
1094 log_err("fio: os or kernel doesn't support IO scheduler"
1098 td_verror(td, errno, "fopen iosched");
1105 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1106 if (ferror(f) || ret != 1) {
1107 td_verror(td, errno, "fwrite");
1115 * Read back and check that the selected scheduler is now the default.
1117 memset(tmp, 0, sizeof(tmp));
1118 ret = fread(tmp, 1, sizeof(tmp) - 1, f);
1119 if (ferror(f) || ret < 0) {
1120 td_verror(td, errno, "fread");
1125 sprintf(tmp2, "[%s]", td->o.ioscheduler);
1126 if (!strstr(tmp, tmp2)) {
1127 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1128 td_verror(td, EINVAL, "iosched_switch");
1137 static int keep_running(struct thread_data *td)
1141 if (td->o.time_based)
1147 if (exceeds_number_ios(td))
1150 if (td->o.size != -1ULL && ddir_rw_sum(td->io_bytes) < td->o.size) {
1154 * If the difference is less than the minimum IO size, we
1157 diff = td->o.size - ddir_rw_sum(td->io_bytes);
1158 if (diff < td_max_bs(td))
1161 if (fio_files_done(td))
1170 static int exec_string(struct thread_options *o, const char *string, const char *mode)
1172 int ret, newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
1175 str = malloc(newlen);
1176 sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
1178 log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1181 log_err("fio: exec of cmd <%s> failed\n", str);
1188 * Dry run to compute correct state of numberio for verification.
1190 static uint64_t do_dry_run(struct thread_data *td)
1192 uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
1194 td_set_runstate(td, TD_RUNNING);
1196 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1197 (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td)) {
1201 if (td->terminate || td->done)
1204 io_u = get_io_u(td);
1208 io_u->flags |= IO_U_F_FLIGHT;
1211 if (ddir_rw(acct_ddir(io_u)))
1212 td->io_issues[acct_ddir(io_u)]++;
1213 if (ddir_rw(io_u->ddir)) {
1214 io_u_mark_depth(td, 1);
1215 td->ts.total_io_u[io_u->ddir]++;
1218 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1220 td->o.verify != VERIFY_NONE &&
1221 !td->o.experimental_verify)
1222 log_io_piece(td, io_u);
1224 ret = io_u_sync_complete(td, io_u, bytes_done);
1228 return bytes_done[DDIR_WRITE] + bytes_done[DDIR_TRIM];
1232 * Entry point for the thread based jobs. The process based jobs end up
1233 * here as well, after a little setup.
1235 static void *thread_main(void *data)
1237 unsigned long long elapsed;
1238 struct thread_data *td = data;
1239 struct thread_options *o = &td->o;
1240 pthread_condattr_t attr;
1244 if (!o->use_thread) {
1251 * fio_time_init() may not have been called yet if running as a server
1255 fio_local_clock_init(o->use_thread);
1257 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1260 fio_server_send_start(td);
1262 INIT_FLIST_HEAD(&td->io_log_list);
1263 INIT_FLIST_HEAD(&td->io_hist_list);
1264 INIT_FLIST_HEAD(&td->verify_list);
1265 INIT_FLIST_HEAD(&td->trim_list);
1266 INIT_FLIST_HEAD(&td->next_rand_list);
1267 pthread_mutex_init(&td->io_u_lock, NULL);
1268 td->io_hist_tree = RB_ROOT;
1270 pthread_condattr_init(&attr);
1271 pthread_cond_init(&td->verify_cond, &attr);
1272 pthread_cond_init(&td->free_cond, &attr);
1274 td_set_runstate(td, TD_INITIALIZED);
1275 dprint(FD_MUTEX, "up startup_mutex\n");
1276 fio_mutex_up(startup_mutex);
1277 dprint(FD_MUTEX, "wait on td->mutex\n");
1278 fio_mutex_down(td->mutex);
1279 dprint(FD_MUTEX, "done waiting on td->mutex\n");
1282 * A new gid requires privilege, so we need to do this before setting
1285 if (o->gid != -1U && setgid(o->gid)) {
1286 td_verror(td, errno, "setgid");
1289 if (o->uid != -1U && setuid(o->uid)) {
1290 td_verror(td, errno, "setuid");
1295 * If we have a gettimeofday() thread, make sure we exclude that
1296 * thread from this job
1299 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1302 * Set affinity first, in case it has an impact on the memory
1305 if (o->cpumask_set) {
1306 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1307 ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1309 log_err("fio: no CPUs set\n");
1310 log_err("fio: Try increasing number of available CPUs\n");
1311 td_verror(td, EINVAL, "cpus_split");
1315 ret = fio_setaffinity(td->pid, o->cpumask);
1317 td_verror(td, errno, "cpu_set_affinity");
1322 #ifdef CONFIG_LIBNUMA
1323 /* numa node setup */
1324 if (o->numa_cpumask_set || o->numa_memmask_set) {
1327 if (numa_available() < 0) {
1328 td_verror(td, errno, "Does not support NUMA API\n");
1332 if (o->numa_cpumask_set) {
1333 ret = numa_run_on_node_mask(o->numa_cpunodesmask);
1335 td_verror(td, errno, \
1336 "numa_run_on_node_mask failed\n");
1341 if (o->numa_memmask_set) {
1343 switch (o->numa_mem_mode) {
1344 case MPOL_INTERLEAVE:
1345 numa_set_interleave_mask(o->numa_memnodesmask);
1348 numa_set_membind(o->numa_memnodesmask);
1351 numa_set_localalloc();
1353 case MPOL_PREFERRED:
1354 numa_set_preferred(o->numa_mem_prefer_node);
1365 if (fio_pin_memory(td))
1369 * May alter parameters that init_io_u() will use, so we need to
1378 if (o->verify_async && verify_async_init(td))
1382 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1384 td_verror(td, errno, "ioprio_set");
1389 if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1393 if (nice(o->nice) == -1 && errno != 0) {
1394 td_verror(td, errno, "nice");
1398 if (o->ioscheduler && switch_ioscheduler(td))
1401 if (!o->create_serialize && setup_files(td))
1407 if (init_random_map(td))
1410 if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
1414 if (pre_read_files(td) < 0)
1418 fio_verify_init(td);
1420 fio_gettime(&td->epoch, NULL);
1421 fio_getrusage(&td->ru_start);
1423 while (keep_running(td)) {
1424 uint64_t verify_bytes;
1426 fio_gettime(&td->start, NULL);
1427 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1428 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1429 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1431 if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1432 o->ratemin[DDIR_TRIM]) {
1433 memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1434 sizeof(td->bw_sample_time));
1435 memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1436 sizeof(td->bw_sample_time));
1437 memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1438 sizeof(td->bw_sample_time));
1444 prune_io_piece_log(td);
1446 if (td->o.verify_only && (td_write(td) || td_rw(td)))
1447 verify_bytes = do_dry_run(td);
1449 verify_bytes = do_io(td);
1453 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1454 elapsed = utime_since_now(&td->start);
1455 td->ts.runtime[DDIR_READ] += elapsed;
1457 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1458 elapsed = utime_since_now(&td->start);
1459 td->ts.runtime[DDIR_WRITE] += elapsed;
1461 if (td_trim(td) && td->io_bytes[DDIR_TRIM]) {
1462 elapsed = utime_since_now(&td->start);
1463 td->ts.runtime[DDIR_TRIM] += elapsed;
1466 if (td->error || td->terminate)
1469 if (!o->do_verify ||
1470 o->verify == VERIFY_NONE ||
1471 (td->io_ops->flags & FIO_UNIDIR))
1476 fio_gettime(&td->start, NULL);
1478 do_verify(td, verify_bytes);
1480 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1482 if (td->error || td->terminate)
1486 update_rusage_stat(td);
1487 td->ts.runtime[DDIR_READ] = (td->ts.runtime[DDIR_READ] + 999) / 1000;
1488 td->ts.runtime[DDIR_WRITE] = (td->ts.runtime[DDIR_WRITE] + 999) / 1000;
1489 td->ts.runtime[DDIR_TRIM] = (td->ts.runtime[DDIR_TRIM] + 999) / 1000;
1490 td->ts.total_run_time = mtime_since_now(&td->epoch);
1491 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1492 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1493 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1495 fio_unpin_memory(td);
1497 fio_writeout_logs(td);
1499 if (o->exec_postrun)
1500 exec_string(o, o->exec_postrun, (const char *)"postrun");
1502 if (exitall_on_terminate)
1503 fio_terminate_threads(td->groupid);
1507 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1510 if (o->verify_async)
1511 verify_async_exit(td);
1513 close_and_free_files(td);
1516 cgroup_shutdown(td, &cgroup_mnt);
1518 if (o->cpumask_set) {
1519 int ret = fio_cpuset_exit(&o->cpumask);
1521 td_verror(td, ret, "fio_cpuset_exit");
1525 * do this very late, it will log file closing as well
1527 if (o->write_iolog_file)
1528 write_iolog_close(td);
1530 fio_mutex_remove(td->rusage_sem);
1531 td->rusage_sem = NULL;
1533 fio_mutex_remove(td->mutex);
1536 td_set_runstate(td, TD_EXITED);
1537 return (void *) (uintptr_t) td->error;
1542 * We cannot pass the td data into a forked process, so attach the td and
1543 * pass it to the thread worker.
1545 static int fork_main(int shmid, int offset)
1547 struct thread_data *td;
1551 data = shmat(shmid, NULL, 0);
1552 if (data == (void *) -1) {
1560 * HP-UX inherits shm mappings?
1565 td = data + offset * sizeof(struct thread_data);
1566 ret = thread_main(td);
1568 return (int) (uintptr_t) ret;
1572 * Run over the job map and reap the threads that have exited, if any.
1574 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1575 unsigned int *m_rate)
1577 struct thread_data *td;
1578 unsigned int cputhreads, realthreads, pending;
1582 * reap exited threads (TD_EXITED -> TD_REAPED)
1584 realthreads = pending = cputhreads = 0;
1585 for_each_td(td, i) {
1589 * ->io_ops is NULL for a thread that has closed its
1592 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1601 if (td->runstate == TD_REAPED)
1603 if (td->o.use_thread) {
1604 if (td->runstate == TD_EXITED) {
1605 td_set_runstate(td, TD_REAPED);
1612 if (td->runstate == TD_EXITED)
1616 * check if someone quit or got killed in an unusual way
1618 ret = waitpid(td->pid, &status, flags);
1620 if (errno == ECHILD) {
1621 log_err("fio: pid=%d disappeared %d\n",
1622 (int) td->pid, td->runstate);
1624 td_set_runstate(td, TD_REAPED);
1628 } else if (ret == td->pid) {
1629 if (WIFSIGNALED(status)) {
1630 int sig = WTERMSIG(status);
1632 if (sig != SIGTERM && sig != SIGUSR2)
1633 log_err("fio: pid=%d, got signal=%d\n",
1634 (int) td->pid, sig);
1636 td_set_runstate(td, TD_REAPED);
1639 if (WIFEXITED(status)) {
1640 if (WEXITSTATUS(status) && !td->error)
1641 td->error = WEXITSTATUS(status);
1643 td_set_runstate(td, TD_REAPED);
1649 * thread is not dead, continue
1655 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1656 (*t_rate) -= ddir_rw_sum(td->o.rate);
1663 done_secs += mtime_since_now(&td->epoch) / 1000;
1664 profile_td_exit(td);
1667 if (*nr_running == cputhreads && !pending && realthreads)
1668 fio_terminate_threads(TERMINATE_ALL);
1671 static void do_usleep(unsigned int usecs)
1673 check_for_running_stats();
1678 * Main function for kicking off and reaping jobs, as needed.
1680 static void run_threads(void)
1682 struct thread_data *td;
1683 unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
1686 if (fio_gtod_offload && fio_start_gtod_thread())
1689 fio_idle_prof_init();
1693 nr_thread = nr_process = 0;
1694 for_each_td(td, i) {
1695 if (td->o.use_thread)
1701 if (output_format == FIO_OUTPUT_NORMAL) {
1702 log_info("Starting ");
1704 log_info("%d thread%s", nr_thread,
1705 nr_thread > 1 ? "s" : "");
1709 log_info("%d process%s", nr_process,
1710 nr_process > 1 ? "es" : "");
1716 todo = thread_number;
1719 m_rate = t_rate = 0;
1721 for_each_td(td, i) {
1722 print_status_init(td->thread_number - 1);
1724 if (!td->o.create_serialize)
1728 * do file setup here so it happens sequentially,
1729 * we don't want X number of threads getting their
1730 * client data interspersed on disk
1732 if (setup_files(td)) {
1735 log_err("fio: pid=%d, err=%d/%s\n",
1736 (int) td->pid, td->error, td->verror);
1737 td_set_runstate(td, TD_REAPED);
1744 * for sharing to work, each job must always open
1745 * its own files. so close them, if we opened them
1748 for_each_file(td, f, j) {
1749 if (fio_file_open(f))
1750 td_io_close_file(td, f);
1755 /* start idle threads before io threads start to run */
1756 fio_idle_prof_start();
1761 struct thread_data *map[REAL_MAX_JOBS];
1762 struct timeval this_start;
1763 int this_jobs = 0, left;
1766 * create threads (TD_NOT_CREATED -> TD_CREATED)
1768 for_each_td(td, i) {
1769 if (td->runstate != TD_NOT_CREATED)
1773 * never got a chance to start, killed by other
1774 * thread for some reason
1776 if (td->terminate) {
1781 if (td->o.start_delay) {
1782 spent = utime_since_genesis();
1784 if (td->o.start_delay > spent)
1788 if (td->o.stonewall && (nr_started || nr_running)) {
1789 dprint(FD_PROCESS, "%s: stonewall wait\n",
1796 td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED);
1797 td->update_rusage = 0;
1800 * Set state to created. Thread will transition
1801 * to TD_INITIALIZED when it's done setting up.
1803 td_set_runstate(td, TD_CREATED);
1804 map[this_jobs++] = td;
1807 if (td->o.use_thread) {
1810 dprint(FD_PROCESS, "will pthread_create\n");
1811 ret = pthread_create(&td->thread, NULL,
1814 log_err("pthread_create: %s\n",
1819 ret = pthread_detach(td->thread);
1821 log_err("pthread_detach: %s",
1825 dprint(FD_PROCESS, "will fork\n");
1828 int ret = fork_main(shm_id, i);
1831 } else if (i == fio_debug_jobno)
1832 *fio_debug_jobp = pid;
1834 dprint(FD_MUTEX, "wait on startup_mutex\n");
1835 if (fio_mutex_down_timeout(startup_mutex, 10)) {
1836 log_err("fio: job startup hung? exiting.\n");
1837 fio_terminate_threads(TERMINATE_ALL);
1842 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1846 * Wait for the started threads to transition to
1849 fio_gettime(&this_start, NULL);
1851 while (left && !fio_abort) {
1852 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1857 for (i = 0; i < this_jobs; i++) {
1861 if (td->runstate == TD_INITIALIZED) {
1864 } else if (td->runstate >= TD_EXITED) {
1868 nr_running++; /* work-around... */
1874 log_err("fio: %d job%s failed to start\n", left,
1875 left > 1 ? "s" : "");
1876 for (i = 0; i < this_jobs; i++) {
1880 kill(td->pid, SIGTERM);
1886 * start created threads (TD_INITIALIZED -> TD_RUNNING).
1888 for_each_td(td, i) {
1889 if (td->runstate != TD_INITIALIZED)
1892 if (in_ramp_time(td))
1893 td_set_runstate(td, TD_RAMP);
1895 td_set_runstate(td, TD_RUNNING);
1898 m_rate += ddir_rw_sum(td->o.ratemin);
1899 t_rate += ddir_rw_sum(td->o.rate);
1901 fio_mutex_up(td->mutex);
1904 reap_threads(&nr_running, &t_rate, &m_rate);
1910 while (nr_running) {
1911 reap_threads(&nr_running, &t_rate, &m_rate);
1915 fio_idle_prof_stop();
1920 void wait_for_disk_thread_exit(void)
1922 fio_mutex_down(disk_thread_mutex);
1925 static void free_disk_util(void)
1927 disk_util_start_exit();
1928 wait_for_disk_thread_exit();
1929 disk_util_prune_entries();
1932 static void *disk_thread_main(void *data)
1936 fio_mutex_up(startup_mutex);
1938 while (threads && !ret) {
1939 usleep(DISK_UTIL_MSEC * 1000);
1942 ret = update_io_ticks();
1945 print_thread_status();
1948 fio_mutex_up(disk_thread_mutex);
1952 static int create_disk_util_thread(void)
1958 disk_thread_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
1960 ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
1962 fio_mutex_remove(disk_thread_mutex);
1963 log_err("Can't create disk util thread: %s\n", strerror(ret));
1967 ret = pthread_detach(disk_util_thread);
1969 fio_mutex_remove(disk_thread_mutex);
1970 log_err("Can't detatch disk util thread: %s\n", strerror(ret));
1974 dprint(FD_MUTEX, "wait on startup_mutex\n");
1975 fio_mutex_down(startup_mutex);
1976 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1980 int fio_backend(void)
1982 struct thread_data *td;
1986 if (load_profile(exec_profile))
1989 exec_profile = NULL;
1995 setup_log(&agg_io_log[DDIR_READ], 0, IO_LOG_TYPE_BW);
1996 setup_log(&agg_io_log[DDIR_WRITE], 0, IO_LOG_TYPE_BW);
1997 setup_log(&agg_io_log[DDIR_TRIM], 0, IO_LOG_TYPE_BW);
2000 startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
2001 if (startup_mutex == NULL)
2006 create_disk_util_thread();
2008 cgroup_list = smalloc(sizeof(*cgroup_list));
2009 INIT_FLIST_HEAD(cgroup_list);
2016 __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
2017 __finish_log(agg_io_log[DDIR_WRITE],
2018 "agg-write_bw.log");
2019 __finish_log(agg_io_log[DDIR_TRIM],
2020 "agg-write_bw.log");
2025 fio_options_free(td);
2028 cgroup_kill(cgroup_list);
2032 fio_mutex_remove(startup_mutex);
2033 fio_mutex_remove(disk_thread_mutex);