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);
208 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
212 if (rate < rate_iops_min ||
213 iops < td->rate_blocks[ddir]) {
214 log_err("%s: min iops rate %u not met,"
215 " got %lu\n", td->o.name,
216 rate_iops_min, rate);
222 td->rate_bytes[ddir] = bytes;
223 td->rate_blocks[ddir] = iops;
224 memcpy(&td->lastrate[ddir], now, sizeof(*now));
228 static int check_min_rate(struct thread_data *td, struct timeval *now,
229 uint64_t *bytes_done)
233 if (bytes_done[DDIR_READ])
234 ret |= __check_min_rate(td, now, DDIR_READ);
235 if (bytes_done[DDIR_WRITE])
236 ret |= __check_min_rate(td, now, DDIR_WRITE);
237 if (bytes_done[DDIR_TRIM])
238 ret |= __check_min_rate(td, now, DDIR_TRIM);
244 * When job exits, we can cancel the in-flight IO if we are using async
245 * io. Attempt to do so.
247 static void cleanup_pending_aio(struct thread_data *td)
252 * get immediately available events, if any
254 r = io_u_queued_complete(td, 0, NULL);
259 * now cancel remaining active events
261 if (td->io_ops->cancel) {
265 io_u_qiter(&td->io_u_all, io_u, i) {
266 if (io_u->flags & IO_U_F_FLIGHT) {
267 r = td->io_ops->cancel(td, io_u);
275 r = io_u_queued_complete(td, td->cur_depth, NULL);
279 * Helper to handle the final sync of a file. Works just like the normal
280 * io path, just does everything sync.
282 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
284 struct io_u *io_u = __get_io_u(td);
290 io_u->ddir = DDIR_SYNC;
293 if (td_io_prep(td, io_u)) {
299 ret = td_io_queue(td, io_u);
301 td_verror(td, io_u->error, "td_io_queue");
304 } else if (ret == FIO_Q_QUEUED) {
305 if (io_u_queued_complete(td, 1, NULL) < 0)
307 } else if (ret == FIO_Q_COMPLETED) {
309 td_verror(td, io_u->error, "td_io_queue");
313 if (io_u_sync_complete(td, io_u, NULL) < 0)
315 } else if (ret == FIO_Q_BUSY) {
316 if (td_io_commit(td))
324 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
328 if (fio_file_open(f))
329 return fio_io_sync(td, f);
331 if (td_io_open_file(td, f))
334 ret = fio_io_sync(td, f);
335 td_io_close_file(td, f);
339 static inline void __update_tv_cache(struct thread_data *td)
341 fio_gettime(&td->tv_cache, NULL);
344 static inline void update_tv_cache(struct thread_data *td)
346 if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
347 __update_tv_cache(td);
350 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
352 if (in_ramp_time(td))
356 if (utime_since(&td->epoch, t) >= td->o.timeout)
362 static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
367 if (ret < 0 || td->error) {
369 enum error_type_bit eb;
374 eb = td_error_type(ddir, err);
375 if (!(td->o.continue_on_error & (1 << eb)))
378 if (td_non_fatal_error(td, eb, err)) {
380 * Continue with the I/Os in case of
383 update_error_count(td, err);
387 } else if (td->o.fill_device && err == ENOSPC) {
389 * We expect to hit this error if
390 * fill_device option is set.
397 * Stop the I/O in case of a fatal
400 update_error_count(td, err);
408 static void check_update_rusage(struct thread_data *td)
410 if (td->update_rusage) {
411 td->update_rusage = 0;
412 update_rusage_stat(td);
413 fio_mutex_up(td->rusage_sem);
418 * The main verify engine. Runs over the writes we previously submitted,
419 * reads the blocks back in, and checks the crc/md5 of the data.
421 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
423 uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
429 dprint(FD_VERIFY, "starting loop\n");
432 * sync io first and invalidate cache, to make sure we really
435 for_each_file(td, f, i) {
436 if (!fio_file_open(f))
438 if (fio_io_sync(td, f))
440 if (file_invalidate_cache(td, f))
444 check_update_rusage(td);
449 td_set_runstate(td, TD_VERIFYING);
452 while (!td->terminate) {
457 check_update_rusage(td);
459 if (runtime_exceeded(td, &td->tv_cache)) {
460 __update_tv_cache(td);
461 if (runtime_exceeded(td, &td->tv_cache)) {
467 if (flow_threshold_exceeded(td))
470 if (!td->o.experimental_verify) {
471 io_u = __get_io_u(td);
475 if (get_next_verify(td, io_u)) {
480 if (td_io_prep(td, io_u)) {
485 if (ddir_rw_sum(bytes_done) + td->o.rw_min_bs > verify_bytes)
488 while ((io_u = get_io_u(td)) != NULL) {
496 * We are only interested in the places where
497 * we wrote or trimmed IOs. Turn those into
498 * reads for verification purposes.
500 if (io_u->ddir == DDIR_READ) {
502 * Pretend we issued it for rwmix
505 td->io_issues[DDIR_READ]++;
508 } else if (io_u->ddir == DDIR_TRIM) {
509 io_u->ddir = DDIR_READ;
510 io_u->flags |= IO_U_F_TRIMMED;
512 } else if (io_u->ddir == DDIR_WRITE) {
513 io_u->ddir = DDIR_READ;
525 if (td->o.verify_async)
526 io_u->end_io = verify_io_u_async;
528 io_u->end_io = verify_io_u;
532 ret = td_io_queue(td, io_u);
534 case FIO_Q_COMPLETED:
537 clear_io_u(td, io_u);
538 } else if (io_u->resid) {
539 int bytes = io_u->xfer_buflen - io_u->resid;
545 td_verror(td, EIO, "full resid");
550 io_u->xfer_buflen = io_u->resid;
551 io_u->xfer_buf += bytes;
552 io_u->offset += bytes;
554 if (ddir_rw(io_u->ddir))
555 td->ts.short_io_u[io_u->ddir]++;
558 if (io_u->offset == f->real_file_size)
561 requeue_io_u(td, &io_u);
564 ret = io_u_sync_complete(td, io_u, bytes_done);
572 requeue_io_u(td, &io_u);
573 ret2 = td_io_commit(td);
579 td_verror(td, -ret, "td_io_queue");
583 if (break_on_this_error(td, ddir, &ret))
587 * if we can queue more, do so. but check if there are
588 * completed io_u's first. Note that we can get BUSY even
589 * without IO queued, if the system is resource starved.
592 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
593 if (full || !td->o.iodepth_batch_complete) {
594 min_events = min(td->o.iodepth_batch_complete,
597 * if the queue is full, we MUST reap at least 1 event
599 if (full && !min_events)
604 * Reap required number of io units, if any,
605 * and do the verification on them through
606 * the callback handler
608 if (io_u_queued_complete(td, min_events, bytes_done) < 0) {
612 } while (full && (td->cur_depth > td->o.iodepth_low));
618 check_update_rusage(td);
621 min_events = td->cur_depth;
624 ret = io_u_queued_complete(td, min_events, NULL);
626 cleanup_pending_aio(td);
628 td_set_runstate(td, TD_RUNNING);
630 dprint(FD_VERIFY, "exiting loop\n");
633 static unsigned int exceeds_number_ios(struct thread_data *td)
635 unsigned long long number_ios;
637 if (!td->o.number_ios)
640 number_ios = ddir_rw_sum(td->this_io_blocks);
641 number_ios += td->io_u_queued + td->io_u_in_flight;
643 return number_ios >= td->o.number_ios;
646 static int io_bytes_exceeded(struct thread_data *td)
648 unsigned long long bytes;
651 bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE];
652 else if (td_write(td))
653 bytes = td->this_io_bytes[DDIR_WRITE];
654 else if (td_read(td))
655 bytes = td->this_io_bytes[DDIR_READ];
657 bytes = td->this_io_bytes[DDIR_TRIM];
659 return bytes >= td->o.size || exceeds_number_ios(td);
663 * Main IO worker function. It retrieves io_u's to process and queues
664 * and reaps them, checking for rate and errors along the way.
666 * Returns number of bytes written and trimmed.
668 static uint64_t do_io(struct thread_data *td)
670 uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
673 uint64_t total_bytes, bytes_issued = 0;
675 if (in_ramp_time(td))
676 td_set_runstate(td, TD_RAMP);
678 td_set_runstate(td, TD_RUNNING);
683 * If verify_backlog is enabled, we'll run the verify in this
684 * handler as well. For that case, we may need up to twice the
687 total_bytes = td->o.size;
688 if (td->o.verify != VERIFY_NONE &&
689 (td_write(td) && td->o.verify_backlog))
690 total_bytes += td->o.size;
692 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
693 (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td) ||
695 struct timeval comp_time;
701 check_update_rusage(td);
703 if (td->terminate || td->done)
708 if (runtime_exceeded(td, &td->tv_cache)) {
709 __update_tv_cache(td);
710 if (runtime_exceeded(td, &td->tv_cache)) {
716 if (flow_threshold_exceeded(td))
719 if (bytes_issued >= total_bytes)
723 if (IS_ERR_OR_NULL(io_u)) {
724 int err = PTR_ERR(io_u);
731 if (td->o.latency_target)
739 * Add verification end_io handler if:
740 * - Asked to verify (!td_rw(td))
741 * - Or the io_u is from our verify list (mixed write/ver)
743 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
744 ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
746 if (!td->o.verify_pattern_bytes) {
747 io_u->rand_seed = __rand(&td->__verify_state);
748 if (sizeof(int) != sizeof(long *))
749 io_u->rand_seed *= __rand(&td->__verify_state);
752 if (td->o.verify_async)
753 io_u->end_io = verify_io_u_async;
755 io_u->end_io = verify_io_u;
756 td_set_runstate(td, TD_VERIFYING);
757 } else if (in_ramp_time(td))
758 td_set_runstate(td, TD_RAMP);
760 td_set_runstate(td, TD_RUNNING);
763 * Always log IO before it's issued, so we know the specific
764 * order of it. The logged unit will track when the IO has
767 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
769 td->o.verify != VERIFY_NONE &&
770 !td->o.experimental_verify)
771 log_io_piece(td, io_u);
773 ret = td_io_queue(td, io_u);
775 case FIO_Q_COMPLETED:
778 clear_io_u(td, io_u);
779 } else if (io_u->resid) {
780 int bytes = io_u->xfer_buflen - io_u->resid;
781 struct fio_file *f = io_u->file;
783 bytes_issued += bytes;
788 td_verror(td, EIO, "full resid");
793 io_u->xfer_buflen = io_u->resid;
794 io_u->xfer_buf += bytes;
795 io_u->offset += bytes;
797 if (ddir_rw(io_u->ddir))
798 td->ts.short_io_u[io_u->ddir]++;
800 if (io_u->offset == f->real_file_size)
803 requeue_io_u(td, &io_u);
806 if (__should_check_rate(td, DDIR_READ) ||
807 __should_check_rate(td, DDIR_WRITE) ||
808 __should_check_rate(td, DDIR_TRIM))
809 fio_gettime(&comp_time, NULL);
811 ret = io_u_sync_complete(td, io_u, bytes_done);
814 bytes_issued += io_u->xfer_buflen;
819 * if the engine doesn't have a commit hook,
820 * the io_u is really queued. if it does have such
821 * a hook, it has to call io_u_queued() itself.
823 if (td->io_ops->commit == NULL)
824 io_u_queued(td, io_u);
825 bytes_issued += io_u->xfer_buflen;
828 requeue_io_u(td, &io_u);
829 ret2 = td_io_commit(td);
839 if (break_on_this_error(td, ddir, &ret))
843 * See if we need to complete some commands. Note that we
844 * can get BUSY even without IO queued, if the system is
848 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
849 if (full || !td->o.iodepth_batch_complete) {
850 min_evts = min(td->o.iodepth_batch_complete,
853 * if the queue is full, we MUST reap at least 1 event
855 if (full && !min_evts)
858 if (__should_check_rate(td, DDIR_READ) ||
859 __should_check_rate(td, DDIR_WRITE) ||
860 __should_check_rate(td, DDIR_TRIM))
861 fio_gettime(&comp_time, NULL);
864 ret = io_u_queued_complete(td, min_evts, bytes_done);
868 } while (full && (td->cur_depth > td->o.iodepth_low));
873 if (!ddir_rw_sum(bytes_done) && !(td->io_ops->flags & FIO_NOIO))
876 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
877 if (check_min_rate(td, &comp_time, bytes_done)) {
878 if (exitall_on_terminate)
879 fio_terminate_threads(td->groupid);
880 td_verror(td, EIO, "check_min_rate");
884 if (!in_ramp_time(td) && td->o.latency_target)
885 lat_target_check(td);
887 if (td->o.thinktime) {
888 unsigned long long b;
890 b = ddir_rw_sum(td->io_blocks);
891 if (!(b % td->o.thinktime_blocks)) {
896 if (td->o.thinktime_spin)
897 usec_spin(td->o.thinktime_spin);
899 left = td->o.thinktime - td->o.thinktime_spin;
901 usec_sleep(td, left);
906 check_update_rusage(td);
908 if (td->trim_entries)
909 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
911 if (td->o.fill_device && td->error == ENOSPC) {
920 ret = io_u_queued_complete(td, i, bytes_done);
921 if (td->o.fill_device && td->error == ENOSPC)
925 if (should_fsync(td) && td->o.end_fsync) {
926 td_set_runstate(td, TD_FSYNCING);
928 for_each_file(td, f, i) {
929 if (!fio_file_fsync(td, f))
932 log_err("fio: end_fsync failed for file %s\n",
937 cleanup_pending_aio(td);
940 * stop job if we failed doing any IO
942 if (!ddir_rw_sum(td->this_io_bytes))
945 return bytes_done[DDIR_WRITE] + bytes_done[DDIR_TRIM];
948 static void cleanup_io_u(struct thread_data *td)
952 while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
954 if (td->io_ops->io_u_free)
955 td->io_ops->io_u_free(td, io_u);
957 fio_memfree(io_u, sizeof(*io_u));
962 io_u_rexit(&td->io_u_requeues);
963 io_u_qexit(&td->io_u_freelist);
964 io_u_qexit(&td->io_u_all);
967 static int init_io_u(struct thread_data *td)
970 unsigned int max_bs, min_write;
971 int cl_align, i, max_units;
972 int data_xfer = 1, err;
975 max_units = td->o.iodepth;
976 max_bs = td_max_bs(td);
977 min_write = td->o.min_bs[DDIR_WRITE];
978 td->orig_buffer_size = (unsigned long long) max_bs
979 * (unsigned long long) max_units;
981 if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
985 err += io_u_rinit(&td->io_u_requeues, td->o.iodepth);
986 err += io_u_qinit(&td->io_u_freelist, td->o.iodepth);
987 err += io_u_qinit(&td->io_u_all, td->o.iodepth);
990 log_err("fio: failed setting up IO queues\n");
995 * if we may later need to do address alignment, then add any
996 * possible adjustment here so that we don't cause a buffer
997 * overflow later. this adjustment may be too much if we get
998 * lucky and the allocator gives us an aligned address.
1000 if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1001 (td->io_ops->flags & FIO_RAWIO))
1002 td->orig_buffer_size += page_mask + td->o.mem_align;
1004 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1007 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1008 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1011 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1012 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1016 if (data_xfer && allocate_io_mem(td))
1019 if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1020 (td->io_ops->flags & FIO_RAWIO))
1021 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
1023 p = td->orig_buffer;
1025 cl_align = os_cache_line_size();
1027 for (i = 0; i < max_units; i++) {
1033 ptr = fio_memalign(cl_align, sizeof(*io_u));
1035 log_err("fio: unable to allocate aligned memory\n");
1040 memset(io_u, 0, sizeof(*io_u));
1041 INIT_FLIST_HEAD(&io_u->verify_list);
1042 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1046 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1049 io_u_fill_buffer(td, io_u, min_write, max_bs);
1050 if (td_write(td) && td->o.verify_pattern_bytes) {
1052 * Fill the buffer with the pattern if we are
1053 * going to be doing writes.
1055 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1060 io_u->flags = IO_U_F_FREE;
1061 io_u_qpush(&td->io_u_freelist, io_u);
1064 * io_u never leaves this stack, used for iteration of all
1067 io_u_qpush(&td->io_u_all, io_u);
1069 if (td->io_ops->io_u_init) {
1070 int ret = td->io_ops->io_u_init(td, io_u);
1073 log_err("fio: failed to init engine data: %d\n", ret);
1084 static int switch_ioscheduler(struct thread_data *td)
1086 char tmp[256], tmp2[128];
1090 if (td->io_ops->flags & FIO_DISKLESSIO)
1093 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
1095 f = fopen(tmp, "r+");
1097 if (errno == ENOENT) {
1098 log_err("fio: os or kernel doesn't support IO scheduler"
1102 td_verror(td, errno, "fopen iosched");
1109 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1110 if (ferror(f) || ret != 1) {
1111 td_verror(td, errno, "fwrite");
1119 * Read back and check that the selected scheduler is now the default.
1121 ret = fread(tmp, sizeof(tmp), 1, f);
1122 if (ferror(f) || ret < 0) {
1123 td_verror(td, errno, "fread");
1127 tmp[sizeof(tmp) - 1] = '\0';
1130 sprintf(tmp2, "[%s]", td->o.ioscheduler);
1131 if (!strstr(tmp, tmp2)) {
1132 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1133 td_verror(td, EINVAL, "iosched_switch");
1142 static int keep_running(struct thread_data *td)
1146 if (td->o.time_based)
1152 if (exceeds_number_ios(td))
1155 if (td->o.size != -1ULL && ddir_rw_sum(td->io_bytes) < td->o.size) {
1159 * If the difference is less than the minimum IO size, we
1162 diff = td->o.size - ddir_rw_sum(td->io_bytes);
1163 if (diff < td_max_bs(td))
1166 if (fio_files_done(td))
1175 static int exec_string(struct thread_options *o, const char *string, const char *mode)
1177 int ret, newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
1180 str = malloc(newlen);
1181 sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
1183 log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1186 log_err("fio: exec of cmd <%s> failed\n", str);
1193 * Dry run to compute correct state of numberio for verification.
1195 static uint64_t do_dry_run(struct thread_data *td)
1197 uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
1199 td_set_runstate(td, TD_RUNNING);
1201 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1202 (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td)) {
1206 if (td->terminate || td->done)
1209 io_u = get_io_u(td);
1213 io_u->flags |= IO_U_F_FLIGHT;
1216 if (ddir_rw(acct_ddir(io_u)))
1217 td->io_issues[acct_ddir(io_u)]++;
1218 if (ddir_rw(io_u->ddir)) {
1219 io_u_mark_depth(td, 1);
1220 td->ts.total_io_u[io_u->ddir]++;
1223 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1225 td->o.verify != VERIFY_NONE &&
1226 !td->o.experimental_verify)
1227 log_io_piece(td, io_u);
1229 ret = io_u_sync_complete(td, io_u, bytes_done);
1233 return bytes_done[DDIR_WRITE] + bytes_done[DDIR_TRIM];
1237 * Entry point for the thread based jobs. The process based jobs end up
1238 * here as well, after a little setup.
1240 static void *thread_main(void *data)
1242 unsigned long long elapsed;
1243 struct thread_data *td = data;
1244 struct thread_options *o = &td->o;
1245 pthread_condattr_t attr;
1249 if (!o->use_thread) {
1256 * fio_time_init() may not have been called yet if running as a server
1260 fio_local_clock_init(o->use_thread);
1262 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1265 fio_server_send_start(td);
1267 INIT_FLIST_HEAD(&td->io_log_list);
1268 INIT_FLIST_HEAD(&td->io_hist_list);
1269 INIT_FLIST_HEAD(&td->verify_list);
1270 INIT_FLIST_HEAD(&td->trim_list);
1271 INIT_FLIST_HEAD(&td->next_rand_list);
1272 pthread_mutex_init(&td->io_u_lock, NULL);
1273 td->io_hist_tree = RB_ROOT;
1275 pthread_condattr_init(&attr);
1276 pthread_cond_init(&td->verify_cond, &attr);
1277 pthread_cond_init(&td->free_cond, &attr);
1279 td_set_runstate(td, TD_INITIALIZED);
1280 dprint(FD_MUTEX, "up startup_mutex\n");
1281 fio_mutex_up(startup_mutex);
1282 dprint(FD_MUTEX, "wait on td->mutex\n");
1283 fio_mutex_down(td->mutex);
1284 dprint(FD_MUTEX, "done waiting on td->mutex\n");
1287 * A new gid requires privilege, so we need to do this before setting
1290 if (o->gid != -1U && setgid(o->gid)) {
1291 td_verror(td, errno, "setgid");
1294 if (o->uid != -1U && setuid(o->uid)) {
1295 td_verror(td, errno, "setuid");
1300 * If we have a gettimeofday() thread, make sure we exclude that
1301 * thread from this job
1304 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1307 * Set affinity first, in case it has an impact on the memory
1310 if (o->cpumask_set) {
1311 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1312 ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1314 log_err("fio: no CPUs set\n");
1315 log_err("fio: Try increasing number of available CPUs\n");
1316 td_verror(td, EINVAL, "cpus_split");
1320 ret = fio_setaffinity(td->pid, o->cpumask);
1322 td_verror(td, errno, "cpu_set_affinity");
1327 #ifdef CONFIG_LIBNUMA
1328 /* numa node setup */
1329 if (o->numa_cpumask_set || o->numa_memmask_set) {
1332 if (numa_available() < 0) {
1333 td_verror(td, errno, "Does not support NUMA API\n");
1337 if (o->numa_cpumask_set) {
1338 ret = numa_run_on_node_mask(o->numa_cpunodesmask);
1340 td_verror(td, errno, \
1341 "numa_run_on_node_mask failed\n");
1346 if (o->numa_memmask_set) {
1348 switch (o->numa_mem_mode) {
1349 case MPOL_INTERLEAVE:
1350 numa_set_interleave_mask(o->numa_memnodesmask);
1353 numa_set_membind(o->numa_memnodesmask);
1356 numa_set_localalloc();
1358 case MPOL_PREFERRED:
1359 numa_set_preferred(o->numa_mem_prefer_node);
1370 if (fio_pin_memory(td))
1374 * May alter parameters that init_io_u() will use, so we need to
1383 if (o->verify_async && verify_async_init(td))
1387 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1389 td_verror(td, errno, "ioprio_set");
1394 if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1398 if (nice(o->nice) == -1 && errno != 0) {
1399 td_verror(td, errno, "nice");
1403 if (o->ioscheduler && switch_ioscheduler(td))
1406 if (!o->create_serialize && setup_files(td))
1412 if (init_random_map(td))
1415 if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
1419 if (pre_read_files(td) < 0)
1423 fio_verify_init(td);
1425 fio_gettime(&td->epoch, NULL);
1426 fio_getrusage(&td->ru_start);
1428 while (keep_running(td)) {
1429 uint64_t verify_bytes;
1431 fio_gettime(&td->start, NULL);
1432 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1433 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1434 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1436 if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1437 o->ratemin[DDIR_TRIM]) {
1438 memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1439 sizeof(td->bw_sample_time));
1440 memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1441 sizeof(td->bw_sample_time));
1442 memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1443 sizeof(td->bw_sample_time));
1449 prune_io_piece_log(td);
1451 if (td->o.verify_only && (td_write(td) || td_rw(td)))
1452 verify_bytes = do_dry_run(td);
1454 verify_bytes = do_io(td);
1458 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1459 elapsed = utime_since_now(&td->start);
1460 td->ts.runtime[DDIR_READ] += elapsed;
1462 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1463 elapsed = utime_since_now(&td->start);
1464 td->ts.runtime[DDIR_WRITE] += elapsed;
1466 if (td_trim(td) && td->io_bytes[DDIR_TRIM]) {
1467 elapsed = utime_since_now(&td->start);
1468 td->ts.runtime[DDIR_TRIM] += elapsed;
1471 if (td->error || td->terminate)
1474 if (!o->do_verify ||
1475 o->verify == VERIFY_NONE ||
1476 (td->io_ops->flags & FIO_UNIDIR))
1481 fio_gettime(&td->start, NULL);
1483 do_verify(td, verify_bytes);
1485 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1487 if (td->error || td->terminate)
1491 update_rusage_stat(td);
1492 td->ts.runtime[DDIR_READ] = (td->ts.runtime[DDIR_READ] + 999) / 1000;
1493 td->ts.runtime[DDIR_WRITE] = (td->ts.runtime[DDIR_WRITE] + 999) / 1000;
1494 td->ts.runtime[DDIR_TRIM] = (td->ts.runtime[DDIR_TRIM] + 999) / 1000;
1495 td->ts.total_run_time = mtime_since_now(&td->epoch);
1496 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1497 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1498 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1500 fio_unpin_memory(td);
1502 fio_writeout_logs(td);
1504 if (o->exec_postrun)
1505 exec_string(o, o->exec_postrun, (const char *)"postrun");
1507 if (exitall_on_terminate)
1508 fio_terminate_threads(td->groupid);
1512 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1515 if (o->verify_async)
1516 verify_async_exit(td);
1518 close_and_free_files(td);
1521 cgroup_shutdown(td, &cgroup_mnt);
1523 if (o->cpumask_set) {
1524 int ret = fio_cpuset_exit(&o->cpumask);
1526 td_verror(td, ret, "fio_cpuset_exit");
1530 * do this very late, it will log file closing as well
1532 if (o->write_iolog_file)
1533 write_iolog_close(td);
1535 fio_mutex_remove(td->rusage_sem);
1536 td->rusage_sem = NULL;
1538 fio_mutex_remove(td->mutex);
1541 td_set_runstate(td, TD_EXITED);
1542 return (void *) (uintptr_t) td->error;
1547 * We cannot pass the td data into a forked process, so attach the td and
1548 * pass it to the thread worker.
1550 static int fork_main(int shmid, int offset)
1552 struct thread_data *td;
1556 data = shmat(shmid, NULL, 0);
1557 if (data == (void *) -1) {
1565 * HP-UX inherits shm mappings?
1570 td = data + offset * sizeof(struct thread_data);
1571 ret = thread_main(td);
1573 return (int) (uintptr_t) ret;
1577 * Run over the job map and reap the threads that have exited, if any.
1579 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1580 unsigned int *m_rate)
1582 struct thread_data *td;
1583 unsigned int cputhreads, realthreads, pending;
1587 * reap exited threads (TD_EXITED -> TD_REAPED)
1589 realthreads = pending = cputhreads = 0;
1590 for_each_td(td, i) {
1594 * ->io_ops is NULL for a thread that has closed its
1597 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1606 if (td->runstate == TD_REAPED)
1608 if (td->o.use_thread) {
1609 if (td->runstate == TD_EXITED) {
1610 td_set_runstate(td, TD_REAPED);
1617 if (td->runstate == TD_EXITED)
1621 * check if someone quit or got killed in an unusual way
1623 ret = waitpid(td->pid, &status, flags);
1625 if (errno == ECHILD) {
1626 log_err("fio: pid=%d disappeared %d\n",
1627 (int) td->pid, td->runstate);
1629 td_set_runstate(td, TD_REAPED);
1633 } else if (ret == td->pid) {
1634 if (WIFSIGNALED(status)) {
1635 int sig = WTERMSIG(status);
1637 if (sig != SIGTERM && sig != SIGUSR2)
1638 log_err("fio: pid=%d, got signal=%d\n",
1639 (int) td->pid, sig);
1641 td_set_runstate(td, TD_REAPED);
1644 if (WIFEXITED(status)) {
1645 if (WEXITSTATUS(status) && !td->error)
1646 td->error = WEXITSTATUS(status);
1648 td_set_runstate(td, TD_REAPED);
1654 * thread is not dead, continue
1660 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1661 (*t_rate) -= ddir_rw_sum(td->o.rate);
1668 done_secs += mtime_since_now(&td->epoch) / 1000;
1669 profile_td_exit(td);
1672 if (*nr_running == cputhreads && !pending && realthreads)
1673 fio_terminate_threads(TERMINATE_ALL);
1676 static void do_usleep(unsigned int usecs)
1678 check_for_running_stats();
1683 * Main function for kicking off and reaping jobs, as needed.
1685 static void run_threads(void)
1687 struct thread_data *td;
1688 unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
1691 if (fio_gtod_offload && fio_start_gtod_thread())
1694 fio_idle_prof_init();
1698 nr_thread = nr_process = 0;
1699 for_each_td(td, i) {
1700 if (td->o.use_thread)
1706 if (output_format == FIO_OUTPUT_NORMAL) {
1707 log_info("Starting ");
1709 log_info("%d thread%s", nr_thread,
1710 nr_thread > 1 ? "s" : "");
1714 log_info("%d process%s", nr_process,
1715 nr_process > 1 ? "es" : "");
1721 todo = thread_number;
1724 m_rate = t_rate = 0;
1726 for_each_td(td, i) {
1727 print_status_init(td->thread_number - 1);
1729 if (!td->o.create_serialize)
1733 * do file setup here so it happens sequentially,
1734 * we don't want X number of threads getting their
1735 * client data interspersed on disk
1737 if (setup_files(td)) {
1740 log_err("fio: pid=%d, err=%d/%s\n",
1741 (int) td->pid, td->error, td->verror);
1742 td_set_runstate(td, TD_REAPED);
1749 * for sharing to work, each job must always open
1750 * its own files. so close them, if we opened them
1753 for_each_file(td, f, j) {
1754 if (fio_file_open(f))
1755 td_io_close_file(td, f);
1760 /* start idle threads before io threads start to run */
1761 fio_idle_prof_start();
1766 struct thread_data *map[REAL_MAX_JOBS];
1767 struct timeval this_start;
1768 int this_jobs = 0, left;
1771 * create threads (TD_NOT_CREATED -> TD_CREATED)
1773 for_each_td(td, i) {
1774 if (td->runstate != TD_NOT_CREATED)
1778 * never got a chance to start, killed by other
1779 * thread for some reason
1781 if (td->terminate) {
1786 if (td->o.start_delay) {
1787 spent = utime_since_genesis();
1789 if (td->o.start_delay > spent)
1793 if (td->o.stonewall && (nr_started || nr_running)) {
1794 dprint(FD_PROCESS, "%s: stonewall wait\n",
1801 td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED);
1802 td->update_rusage = 0;
1805 * Set state to created. Thread will transition
1806 * to TD_INITIALIZED when it's done setting up.
1808 td_set_runstate(td, TD_CREATED);
1809 map[this_jobs++] = td;
1812 if (td->o.use_thread) {
1815 dprint(FD_PROCESS, "will pthread_create\n");
1816 ret = pthread_create(&td->thread, NULL,
1819 log_err("pthread_create: %s\n",
1824 ret = pthread_detach(td->thread);
1826 log_err("pthread_detach: %s",
1830 dprint(FD_PROCESS, "will fork\n");
1833 int ret = fork_main(shm_id, i);
1836 } else if (i == fio_debug_jobno)
1837 *fio_debug_jobp = pid;
1839 dprint(FD_MUTEX, "wait on startup_mutex\n");
1840 if (fio_mutex_down_timeout(startup_mutex, 10)) {
1841 log_err("fio: job startup hung? exiting.\n");
1842 fio_terminate_threads(TERMINATE_ALL);
1847 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1851 * Wait for the started threads to transition to
1854 fio_gettime(&this_start, NULL);
1856 while (left && !fio_abort) {
1857 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1862 for (i = 0; i < this_jobs; i++) {
1866 if (td->runstate == TD_INITIALIZED) {
1869 } else if (td->runstate >= TD_EXITED) {
1873 nr_running++; /* work-around... */
1879 log_err("fio: %d job%s failed to start\n", left,
1880 left > 1 ? "s" : "");
1881 for (i = 0; i < this_jobs; i++) {
1885 kill(td->pid, SIGTERM);
1891 * start created threads (TD_INITIALIZED -> TD_RUNNING).
1893 for_each_td(td, i) {
1894 if (td->runstate != TD_INITIALIZED)
1897 if (in_ramp_time(td))
1898 td_set_runstate(td, TD_RAMP);
1900 td_set_runstate(td, TD_RUNNING);
1903 m_rate += ddir_rw_sum(td->o.ratemin);
1904 t_rate += ddir_rw_sum(td->o.rate);
1906 fio_mutex_up(td->mutex);
1909 reap_threads(&nr_running, &t_rate, &m_rate);
1915 while (nr_running) {
1916 reap_threads(&nr_running, &t_rate, &m_rate);
1920 fio_idle_prof_stop();
1925 void wait_for_disk_thread_exit(void)
1927 fio_mutex_down(disk_thread_mutex);
1930 static void free_disk_util(void)
1932 disk_util_start_exit();
1933 wait_for_disk_thread_exit();
1934 disk_util_prune_entries();
1937 static void *disk_thread_main(void *data)
1941 fio_mutex_up(startup_mutex);
1943 while (threads && !ret) {
1944 usleep(DISK_UTIL_MSEC * 1000);
1947 ret = update_io_ticks();
1950 print_thread_status();
1953 fio_mutex_up(disk_thread_mutex);
1957 static int create_disk_util_thread(void)
1963 disk_thread_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
1965 ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
1967 fio_mutex_remove(disk_thread_mutex);
1968 log_err("Can't create disk util thread: %s\n", strerror(ret));
1972 ret = pthread_detach(disk_util_thread);
1974 fio_mutex_remove(disk_thread_mutex);
1975 log_err("Can't detatch disk util thread: %s\n", strerror(ret));
1979 dprint(FD_MUTEX, "wait on startup_mutex\n");
1980 fio_mutex_down(startup_mutex);
1981 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1985 int fio_backend(void)
1987 struct thread_data *td;
1991 if (load_profile(exec_profile))
1994 exec_profile = NULL;
2000 setup_log(&agg_io_log[DDIR_READ], 0, IO_LOG_TYPE_BW);
2001 setup_log(&agg_io_log[DDIR_WRITE], 0, IO_LOG_TYPE_BW);
2002 setup_log(&agg_io_log[DDIR_TRIM], 0, IO_LOG_TYPE_BW);
2005 startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
2006 if (startup_mutex == NULL)
2011 create_disk_util_thread();
2013 cgroup_list = smalloc(sizeof(*cgroup_list));
2014 INIT_FLIST_HEAD(cgroup_list);
2021 __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
2022 __finish_log(agg_io_log[DDIR_WRITE],
2023 "agg-write_bw.log");
2024 __finish_log(agg_io_log[DDIR_TRIM],
2025 "agg-write_bw.log");
2030 fio_options_free(td);
2033 cgroup_kill(cgroup_list);
2037 fio_mutex_remove(startup_mutex);
2038 fio_mutex_remove(disk_thread_mutex);