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 fio_mutex *writeout_mutex;
61 static struct flist_head *cgroup_list;
62 static char *cgroup_mnt;
63 static int exit_value;
64 static volatile int fio_abort;
65 static unsigned int nr_process = 0;
66 static unsigned int nr_thread = 0;
68 struct io_log *agg_io_log[DDIR_RWDIR_CNT];
71 unsigned int thread_number = 0;
72 unsigned int stat_number = 0;
75 unsigned long done_secs = 0;
76 volatile int disk_util_exit = 0;
78 #define PAGE_ALIGN(buf) \
79 (char *) (((uintptr_t) (buf) + page_mask) & ~page_mask)
81 #define JOB_START_TIMEOUT (5 * 1000)
83 static void sig_int(int sig)
87 fio_server_got_signal(sig);
89 log_info("\nfio: terminating on signal %d\n", sig);
94 fio_terminate_threads(TERMINATE_ALL);
98 static void sig_show_status(int sig)
100 show_running_run_stats();
103 static void set_sig_handlers(void)
105 struct sigaction act;
107 memset(&act, 0, sizeof(act));
108 act.sa_handler = sig_int;
109 act.sa_flags = SA_RESTART;
110 sigaction(SIGINT, &act, NULL);
112 memset(&act, 0, sizeof(act));
113 act.sa_handler = sig_int;
114 act.sa_flags = SA_RESTART;
115 sigaction(SIGTERM, &act, NULL);
117 /* Windows uses SIGBREAK as a quit signal from other applications */
119 memset(&act, 0, sizeof(act));
120 act.sa_handler = sig_int;
121 act.sa_flags = SA_RESTART;
122 sigaction(SIGBREAK, &act, NULL);
125 memset(&act, 0, sizeof(act));
126 act.sa_handler = sig_show_status;
127 act.sa_flags = SA_RESTART;
128 sigaction(SIGUSR1, &act, NULL);
131 memset(&act, 0, sizeof(act));
132 act.sa_handler = sig_int;
133 act.sa_flags = SA_RESTART;
134 sigaction(SIGPIPE, &act, NULL);
139 * Check if we are above the minimum rate given.
141 static int __check_min_rate(struct thread_data *td, struct timeval *now,
144 unsigned long long bytes = 0;
145 unsigned long iops = 0;
148 unsigned int ratemin = 0;
149 unsigned int rate_iops = 0;
150 unsigned int rate_iops_min = 0;
152 assert(ddir_rw(ddir));
154 if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
158 * allow a 2 second settle period in the beginning
160 if (mtime_since(&td->start, now) < 2000)
163 iops += td->this_io_blocks[ddir];
164 bytes += td->this_io_bytes[ddir];
165 ratemin += td->o.ratemin[ddir];
166 rate_iops += td->o.rate_iops[ddir];
167 rate_iops_min += td->o.rate_iops_min[ddir];
170 * if rate blocks is set, sample is running
172 if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
173 spent = mtime_since(&td->lastrate[ddir], now);
174 if (spent < td->o.ratecycle)
177 if (td->o.rate[ddir]) {
179 * check bandwidth specified rate
181 if (bytes < td->rate_bytes[ddir]) {
182 log_err("%s: min rate %u not met\n", td->o.name,
186 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
187 if (rate < ratemin ||
188 bytes < td->rate_bytes[ddir]) {
189 log_err("%s: min rate %u not met, got"
190 " %luKB/sec\n", td->o.name,
197 * checks iops specified rate
199 if (iops < rate_iops) {
200 log_err("%s: min iops rate %u not met\n",
201 td->o.name, rate_iops);
204 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
205 if (rate < rate_iops_min ||
206 iops < td->rate_blocks[ddir]) {
207 log_err("%s: min iops rate %u not met,"
208 " got %lu\n", td->o.name,
209 rate_iops_min, rate);
215 td->rate_bytes[ddir] = bytes;
216 td->rate_blocks[ddir] = iops;
217 memcpy(&td->lastrate[ddir], now, sizeof(*now));
221 static int check_min_rate(struct thread_data *td, struct timeval *now,
222 uint64_t *bytes_done)
226 if (bytes_done[DDIR_READ])
227 ret |= __check_min_rate(td, now, DDIR_READ);
228 if (bytes_done[DDIR_WRITE])
229 ret |= __check_min_rate(td, now, DDIR_WRITE);
230 if (bytes_done[DDIR_TRIM])
231 ret |= __check_min_rate(td, now, DDIR_TRIM);
237 * When job exits, we can cancel the in-flight IO if we are using async
238 * io. Attempt to do so.
240 static void cleanup_pending_aio(struct thread_data *td)
245 * get immediately available events, if any
247 r = io_u_queued_complete(td, 0, NULL);
252 * now cancel remaining active events
254 if (td->io_ops->cancel) {
258 io_u_qiter(&td->io_u_all, io_u, i) {
259 if (io_u->flags & IO_U_F_FLIGHT) {
260 r = td->io_ops->cancel(td, io_u);
268 r = io_u_queued_complete(td, td->cur_depth, NULL);
272 * Helper to handle the final sync of a file. Works just like the normal
273 * io path, just does everything sync.
275 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
277 struct io_u *io_u = __get_io_u(td);
283 io_u->ddir = DDIR_SYNC;
286 if (td_io_prep(td, io_u)) {
292 ret = td_io_queue(td, io_u);
294 td_verror(td, io_u->error, "td_io_queue");
297 } else if (ret == FIO_Q_QUEUED) {
298 if (io_u_queued_complete(td, 1, NULL) < 0)
300 } else if (ret == FIO_Q_COMPLETED) {
302 td_verror(td, io_u->error, "td_io_queue");
306 if (io_u_sync_complete(td, io_u, NULL) < 0)
308 } else if (ret == FIO_Q_BUSY) {
309 if (td_io_commit(td))
317 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
321 if (fio_file_open(f))
322 return fio_io_sync(td, f);
324 if (td_io_open_file(td, f))
327 ret = fio_io_sync(td, f);
328 td_io_close_file(td, f);
332 static inline void __update_tv_cache(struct thread_data *td)
334 fio_gettime(&td->tv_cache, NULL);
337 static inline void update_tv_cache(struct thread_data *td)
339 if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
340 __update_tv_cache(td);
343 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
345 if (in_ramp_time(td))
349 if (utime_since(&td->epoch, t) >= td->o.timeout)
355 static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
360 if (ret < 0 || td->error) {
362 enum error_type_bit eb;
367 eb = td_error_type(ddir, err);
368 if (!(td->o.continue_on_error & (1 << eb)))
371 if (td_non_fatal_error(td, eb, err)) {
373 * Continue with the I/Os in case of
376 update_error_count(td, err);
380 } else if (td->o.fill_device && err == ENOSPC) {
382 * We expect to hit this error if
383 * fill_device option is set.
390 * Stop the I/O in case of a fatal
393 update_error_count(td, err);
401 static void check_update_rusage(struct thread_data *td)
403 if (td->update_rusage) {
404 td->update_rusage = 0;
405 update_rusage_stat(td);
406 fio_mutex_up(td->rusage_sem);
411 * The main verify engine. Runs over the writes we previously submitted,
412 * reads the blocks back in, and checks the crc/md5 of the data.
414 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
416 uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
422 dprint(FD_VERIFY, "starting loop\n");
425 * sync io first and invalidate cache, to make sure we really
428 for_each_file(td, f, i) {
429 if (!fio_file_open(f))
431 if (fio_io_sync(td, f))
433 if (file_invalidate_cache(td, f))
437 check_update_rusage(td);
442 td_set_runstate(td, TD_VERIFYING);
445 while (!td->terminate) {
450 check_update_rusage(td);
452 if (runtime_exceeded(td, &td->tv_cache)) {
453 __update_tv_cache(td);
454 if (runtime_exceeded(td, &td->tv_cache)) {
460 if (flow_threshold_exceeded(td))
463 if (!td->o.experimental_verify) {
464 io_u = __get_io_u(td);
468 if (get_next_verify(td, io_u)) {
473 if (td_io_prep(td, io_u)) {
478 if (ddir_rw_sum(bytes_done) + td->o.rw_min_bs > verify_bytes)
481 while ((io_u = get_io_u(td)) != NULL) {
489 * We are only interested in the places where
490 * we wrote or trimmed IOs. Turn those into
491 * reads for verification purposes.
493 if (io_u->ddir == DDIR_READ) {
495 * Pretend we issued it for rwmix
498 td->io_issues[DDIR_READ]++;
501 } else if (io_u->ddir == DDIR_TRIM) {
502 io_u->ddir = DDIR_READ;
503 io_u->flags |= IO_U_F_TRIMMED;
505 } else if (io_u->ddir == DDIR_WRITE) {
506 io_u->ddir = DDIR_READ;
518 if (td->o.verify_async)
519 io_u->end_io = verify_io_u_async;
521 io_u->end_io = verify_io_u;
525 ret = td_io_queue(td, io_u);
527 case FIO_Q_COMPLETED:
530 clear_io_u(td, io_u);
531 } else if (io_u->resid) {
532 int bytes = io_u->xfer_buflen - io_u->resid;
538 td_verror(td, EIO, "full resid");
543 io_u->xfer_buflen = io_u->resid;
544 io_u->xfer_buf += bytes;
545 io_u->offset += bytes;
547 if (ddir_rw(io_u->ddir))
548 td->ts.short_io_u[io_u->ddir]++;
551 if (io_u->offset == f->real_file_size)
554 requeue_io_u(td, &io_u);
557 ret = io_u_sync_complete(td, io_u, bytes_done);
565 requeue_io_u(td, &io_u);
566 ret2 = td_io_commit(td);
572 td_verror(td, -ret, "td_io_queue");
576 if (break_on_this_error(td, ddir, &ret))
580 * if we can queue more, do so. but check if there are
581 * completed io_u's first. Note that we can get BUSY even
582 * without IO queued, if the system is resource starved.
585 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
586 if (full || !td->o.iodepth_batch_complete) {
587 min_events = min(td->o.iodepth_batch_complete,
590 * if the queue is full, we MUST reap at least 1 event
592 if (full && !min_events)
597 * Reap required number of io units, if any,
598 * and do the verification on them through
599 * the callback handler
601 if (io_u_queued_complete(td, min_events, bytes_done) < 0) {
605 } while (full && (td->cur_depth > td->o.iodepth_low));
611 check_update_rusage(td);
614 min_events = td->cur_depth;
617 ret = io_u_queued_complete(td, min_events, NULL);
619 cleanup_pending_aio(td);
621 td_set_runstate(td, TD_RUNNING);
623 dprint(FD_VERIFY, "exiting loop\n");
626 static unsigned int exceeds_number_ios(struct thread_data *td)
628 unsigned long long number_ios;
630 if (!td->o.number_ios)
633 number_ios = ddir_rw_sum(td->this_io_blocks);
634 number_ios += td->io_u_queued + td->io_u_in_flight;
636 return number_ios >= td->o.number_ios;
639 static int io_bytes_exceeded(struct thread_data *td)
641 unsigned long long bytes;
644 bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE];
645 else if (td_write(td))
646 bytes = td->this_io_bytes[DDIR_WRITE];
647 else if (td_read(td))
648 bytes = td->this_io_bytes[DDIR_READ];
650 bytes = td->this_io_bytes[DDIR_TRIM];
652 return bytes >= td->o.size || exceeds_number_ios(td);
656 * Main IO worker function. It retrieves io_u's to process and queues
657 * and reaps them, checking for rate and errors along the way.
659 * Returns number of bytes written and trimmed.
661 static uint64_t do_io(struct thread_data *td)
663 uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
666 uint64_t total_bytes, bytes_issued = 0;
668 if (in_ramp_time(td))
669 td_set_runstate(td, TD_RAMP);
671 td_set_runstate(td, TD_RUNNING);
676 * If verify_backlog is enabled, we'll run the verify in this
677 * handler as well. For that case, we may need up to twice the
680 total_bytes = td->o.size;
681 if (td->o.verify != VERIFY_NONE &&
682 (td_write(td) && td->o.verify_backlog))
683 total_bytes += td->o.size;
685 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
686 (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td) ||
688 struct timeval comp_time;
694 check_update_rusage(td);
696 if (td->terminate || td->done)
701 if (runtime_exceeded(td, &td->tv_cache)) {
702 __update_tv_cache(td);
703 if (runtime_exceeded(td, &td->tv_cache)) {
709 if (flow_threshold_exceeded(td))
712 if (bytes_issued >= total_bytes)
716 if (IS_ERR_OR_NULL(io_u)) {
717 int err = PTR_ERR(io_u);
724 if (td->o.latency_target)
732 * Add verification end_io handler if:
733 * - Asked to verify (!td_rw(td))
734 * - Or the io_u is from our verify list (mixed write/ver)
736 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
737 ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
739 if (!td->o.verify_pattern_bytes) {
740 io_u->rand_seed = __rand(&td->__verify_state);
741 if (sizeof(int) != sizeof(long *))
742 io_u->rand_seed *= __rand(&td->__verify_state);
745 if (td->o.verify_async)
746 io_u->end_io = verify_io_u_async;
748 io_u->end_io = verify_io_u;
749 td_set_runstate(td, TD_VERIFYING);
750 } else if (in_ramp_time(td))
751 td_set_runstate(td, TD_RAMP);
753 td_set_runstate(td, TD_RUNNING);
756 * Always log IO before it's issued, so we know the specific
757 * order of it. The logged unit will track when the IO has
760 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
762 td->o.verify != VERIFY_NONE &&
763 !td->o.experimental_verify)
764 log_io_piece(td, io_u);
766 ret = td_io_queue(td, io_u);
768 case FIO_Q_COMPLETED:
771 clear_io_u(td, io_u);
772 } else if (io_u->resid) {
773 int bytes = io_u->xfer_buflen - io_u->resid;
774 struct fio_file *f = io_u->file;
776 bytes_issued += bytes;
781 td_verror(td, EIO, "full resid");
786 io_u->xfer_buflen = io_u->resid;
787 io_u->xfer_buf += bytes;
788 io_u->offset += bytes;
790 if (ddir_rw(io_u->ddir))
791 td->ts.short_io_u[io_u->ddir]++;
793 if (io_u->offset == f->real_file_size)
796 requeue_io_u(td, &io_u);
799 if (__should_check_rate(td, DDIR_READ) ||
800 __should_check_rate(td, DDIR_WRITE) ||
801 __should_check_rate(td, DDIR_TRIM))
802 fio_gettime(&comp_time, NULL);
804 ret = io_u_sync_complete(td, io_u, bytes_done);
807 bytes_issued += io_u->xfer_buflen;
812 * if the engine doesn't have a commit hook,
813 * the io_u is really queued. if it does have such
814 * a hook, it has to call io_u_queued() itself.
816 if (td->io_ops->commit == NULL)
817 io_u_queued(td, io_u);
818 bytes_issued += io_u->xfer_buflen;
821 requeue_io_u(td, &io_u);
822 ret2 = td_io_commit(td);
832 if (break_on_this_error(td, ddir, &ret))
836 * See if we need to complete some commands. Note that we
837 * can get BUSY even without IO queued, if the system is
841 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
842 if (full || !td->o.iodepth_batch_complete) {
843 min_evts = min(td->o.iodepth_batch_complete,
846 * if the queue is full, we MUST reap at least 1 event
848 if (full && !min_evts)
851 if (__should_check_rate(td, DDIR_READ) ||
852 __should_check_rate(td, DDIR_WRITE) ||
853 __should_check_rate(td, DDIR_TRIM))
854 fio_gettime(&comp_time, NULL);
857 ret = io_u_queued_complete(td, min_evts, bytes_done);
861 } while (full && (td->cur_depth > td->o.iodepth_low));
866 if (!ddir_rw_sum(bytes_done) && !(td->io_ops->flags & FIO_NOIO))
869 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
870 if (check_min_rate(td, &comp_time, bytes_done)) {
871 if (exitall_on_terminate)
872 fio_terminate_threads(td->groupid);
873 td_verror(td, EIO, "check_min_rate");
877 if (!in_ramp_time(td) && td->o.latency_target)
878 lat_target_check(td);
880 if (td->o.thinktime) {
881 unsigned long long b;
883 b = ddir_rw_sum(td->io_blocks);
884 if (!(b % td->o.thinktime_blocks)) {
889 if (td->o.thinktime_spin)
890 usec_spin(td->o.thinktime_spin);
892 left = td->o.thinktime - td->o.thinktime_spin;
894 usec_sleep(td, left);
899 check_update_rusage(td);
901 if (td->trim_entries)
902 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
904 if (td->o.fill_device && td->error == ENOSPC) {
913 ret = io_u_queued_complete(td, i, bytes_done);
914 if (td->o.fill_device && td->error == ENOSPC)
918 if (should_fsync(td) && td->o.end_fsync) {
919 td_set_runstate(td, TD_FSYNCING);
921 for_each_file(td, f, i) {
922 if (!fio_file_fsync(td, f))
925 log_err("fio: end_fsync failed for file %s\n",
930 cleanup_pending_aio(td);
933 * stop job if we failed doing any IO
935 if (!ddir_rw_sum(td->this_io_bytes))
938 return bytes_done[DDIR_WRITE] + bytes_done[DDIR_TRIM];
941 static void cleanup_io_u(struct thread_data *td)
945 while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
947 if (td->io_ops->io_u_free)
948 td->io_ops->io_u_free(td, io_u);
950 fio_memfree(io_u, sizeof(*io_u));
955 io_u_rexit(&td->io_u_requeues);
956 io_u_qexit(&td->io_u_freelist);
957 io_u_qexit(&td->io_u_all);
960 static int init_io_u(struct thread_data *td)
963 unsigned int max_bs, min_write;
964 int cl_align, i, max_units;
965 int data_xfer = 1, err;
968 max_units = td->o.iodepth;
969 max_bs = td_max_bs(td);
970 min_write = td->o.min_bs[DDIR_WRITE];
971 td->orig_buffer_size = (unsigned long long) max_bs
972 * (unsigned long long) max_units;
974 if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
978 err += io_u_rinit(&td->io_u_requeues, td->o.iodepth);
979 err += io_u_qinit(&td->io_u_freelist, td->o.iodepth);
980 err += io_u_qinit(&td->io_u_all, td->o.iodepth);
983 log_err("fio: failed setting up IO queues\n");
988 * if we may later need to do address alignment, then add any
989 * possible adjustment here so that we don't cause a buffer
990 * overflow later. this adjustment may be too much if we get
991 * lucky and the allocator gives us an aligned address.
993 if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
994 (td->io_ops->flags & FIO_RAWIO))
995 td->orig_buffer_size += page_mask + td->o.mem_align;
997 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1000 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1001 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1004 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1005 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1009 if (data_xfer && allocate_io_mem(td))
1012 if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1013 (td->io_ops->flags & FIO_RAWIO))
1014 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
1016 p = td->orig_buffer;
1018 cl_align = os_cache_line_size();
1020 for (i = 0; i < max_units; i++) {
1026 ptr = fio_memalign(cl_align, sizeof(*io_u));
1028 log_err("fio: unable to allocate aligned memory\n");
1033 memset(io_u, 0, sizeof(*io_u));
1034 INIT_FLIST_HEAD(&io_u->verify_list);
1035 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1039 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1042 io_u_fill_buffer(td, io_u, min_write, max_bs);
1043 if (td_write(td) && td->o.verify_pattern_bytes) {
1045 * Fill the buffer with the pattern if we are
1046 * going to be doing writes.
1048 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1053 io_u->flags = IO_U_F_FREE;
1054 io_u_qpush(&td->io_u_freelist, io_u);
1057 * io_u never leaves this stack, used for iteration of all
1060 io_u_qpush(&td->io_u_all, io_u);
1062 if (td->io_ops->io_u_init) {
1063 int ret = td->io_ops->io_u_init(td, io_u);
1066 log_err("fio: failed to init engine data: %d\n", ret);
1077 static int switch_ioscheduler(struct thread_data *td)
1079 char tmp[256], tmp2[128];
1083 if (td->io_ops->flags & FIO_DISKLESSIO)
1086 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
1088 f = fopen(tmp, "r+");
1090 if (errno == ENOENT) {
1091 log_err("fio: os or kernel doesn't support IO scheduler"
1095 td_verror(td, errno, "fopen iosched");
1102 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1103 if (ferror(f) || ret != 1) {
1104 td_verror(td, errno, "fwrite");
1112 * Read back and check that the selected scheduler is now the default.
1114 ret = fread(tmp, 1, sizeof(tmp), f);
1115 if (ferror(f) || ret < 0) {
1116 td_verror(td, errno, "fread");
1121 sprintf(tmp2, "[%s]", td->o.ioscheduler);
1122 if (!strstr(tmp, tmp2)) {
1123 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1124 td_verror(td, EINVAL, "iosched_switch");
1133 static int keep_running(struct thread_data *td)
1137 if (td->o.time_based)
1143 if (exceeds_number_ios(td))
1146 if (td->o.size != -1ULL && ddir_rw_sum(td->io_bytes) < td->o.size) {
1150 * If the difference is less than the minimum IO size, we
1153 diff = td->o.size - ddir_rw_sum(td->io_bytes);
1154 if (diff < td_max_bs(td))
1157 if (fio_files_done(td))
1166 static int exec_string(struct thread_options *o, const char *string, const char *mode)
1168 int ret, newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
1171 str = malloc(newlen);
1172 sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
1174 log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1177 log_err("fio: exec of cmd <%s> failed\n", str);
1184 * Dry run to compute correct state of numberio for verification.
1186 static uint64_t do_dry_run(struct thread_data *td)
1188 uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
1190 td_set_runstate(td, TD_RUNNING);
1192 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1193 (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td)) {
1197 if (td->terminate || td->done)
1200 io_u = get_io_u(td);
1204 io_u->flags |= IO_U_F_FLIGHT;
1207 if (ddir_rw(acct_ddir(io_u)))
1208 td->io_issues[acct_ddir(io_u)]++;
1209 if (ddir_rw(io_u->ddir)) {
1210 io_u_mark_depth(td, 1);
1211 td->ts.total_io_u[io_u->ddir]++;
1214 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1216 td->o.verify != VERIFY_NONE &&
1217 !td->o.experimental_verify)
1218 log_io_piece(td, io_u);
1220 ret = io_u_sync_complete(td, io_u, bytes_done);
1224 return bytes_done[DDIR_WRITE] + bytes_done[DDIR_TRIM];
1228 * Entry point for the thread based jobs. The process based jobs end up
1229 * here as well, after a little setup.
1231 static void *thread_main(void *data)
1233 unsigned long long elapsed;
1234 struct thread_data *td = data;
1235 struct thread_options *o = &td->o;
1236 pthread_condattr_t attr;
1240 if (!o->use_thread) {
1247 * fio_time_init() may not have been called yet if running as a server
1251 fio_local_clock_init(o->use_thread);
1253 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1256 fio_server_send_start(td);
1258 INIT_FLIST_HEAD(&td->io_log_list);
1259 INIT_FLIST_HEAD(&td->io_hist_list);
1260 INIT_FLIST_HEAD(&td->verify_list);
1261 INIT_FLIST_HEAD(&td->trim_list);
1262 INIT_FLIST_HEAD(&td->next_rand_list);
1263 pthread_mutex_init(&td->io_u_lock, NULL);
1264 td->io_hist_tree = RB_ROOT;
1266 pthread_condattr_init(&attr);
1267 pthread_cond_init(&td->verify_cond, &attr);
1268 pthread_cond_init(&td->free_cond, &attr);
1270 td_set_runstate(td, TD_INITIALIZED);
1271 dprint(FD_MUTEX, "up startup_mutex\n");
1272 fio_mutex_up(startup_mutex);
1273 dprint(FD_MUTEX, "wait on td->mutex\n");
1274 fio_mutex_down(td->mutex);
1275 dprint(FD_MUTEX, "done waiting on td->mutex\n");
1278 * A new gid requires privilege, so we need to do this before setting
1281 if (o->gid != -1U && setgid(o->gid)) {
1282 td_verror(td, errno, "setgid");
1285 if (o->uid != -1U && setuid(o->uid)) {
1286 td_verror(td, errno, "setuid");
1291 * If we have a gettimeofday() thread, make sure we exclude that
1292 * thread from this job
1295 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1298 * Set affinity first, in case it has an impact on the memory
1301 if (o->cpumask_set) {
1302 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1303 ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1305 log_err("fio: no CPUs set\n");
1306 log_err("fio: Try increasing number of available CPUs\n");
1307 td_verror(td, EINVAL, "cpus_split");
1311 ret = fio_setaffinity(td->pid, o->cpumask);
1313 td_verror(td, errno, "cpu_set_affinity");
1318 #ifdef CONFIG_LIBNUMA
1319 /* numa node setup */
1320 if (o->numa_cpumask_set || o->numa_memmask_set) {
1323 if (numa_available() < 0) {
1324 td_verror(td, errno, "Does not support NUMA API\n");
1328 if (o->numa_cpumask_set) {
1329 ret = numa_run_on_node_mask(o->numa_cpunodesmask);
1331 td_verror(td, errno, \
1332 "numa_run_on_node_mask failed\n");
1337 if (o->numa_memmask_set) {
1339 switch (o->numa_mem_mode) {
1340 case MPOL_INTERLEAVE:
1341 numa_set_interleave_mask(o->numa_memnodesmask);
1344 numa_set_membind(o->numa_memnodesmask);
1347 numa_set_localalloc();
1349 case MPOL_PREFERRED:
1350 numa_set_preferred(o->numa_mem_prefer_node);
1361 if (fio_pin_memory(td))
1365 * May alter parameters that init_io_u() will use, so we need to
1374 if (o->verify_async && verify_async_init(td))
1378 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1380 td_verror(td, errno, "ioprio_set");
1385 if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1389 if (nice(o->nice) == -1 && errno != 0) {
1390 td_verror(td, errno, "nice");
1394 if (o->ioscheduler && switch_ioscheduler(td))
1397 if (!o->create_serialize && setup_files(td))
1403 if (init_random_map(td))
1406 if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
1410 if (pre_read_files(td) < 0)
1414 fio_verify_init(td);
1416 fio_gettime(&td->epoch, NULL);
1417 fio_getrusage(&td->ru_start);
1419 while (keep_running(td)) {
1420 uint64_t verify_bytes;
1422 fio_gettime(&td->start, NULL);
1423 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1424 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1425 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1427 if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1428 o->ratemin[DDIR_TRIM]) {
1429 memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1430 sizeof(td->bw_sample_time));
1431 memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1432 sizeof(td->bw_sample_time));
1433 memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1434 sizeof(td->bw_sample_time));
1440 prune_io_piece_log(td);
1442 if (td->o.verify_only && (td_write(td) || td_rw(td)))
1443 verify_bytes = do_dry_run(td);
1445 verify_bytes = do_io(td);
1449 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1450 elapsed = utime_since_now(&td->start);
1451 td->ts.runtime[DDIR_READ] += elapsed;
1453 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1454 elapsed = utime_since_now(&td->start);
1455 td->ts.runtime[DDIR_WRITE] += elapsed;
1457 if (td_trim(td) && td->io_bytes[DDIR_TRIM]) {
1458 elapsed = utime_since_now(&td->start);
1459 td->ts.runtime[DDIR_TRIM] += elapsed;
1462 if (td->error || td->terminate)
1465 if (!o->do_verify ||
1466 o->verify == VERIFY_NONE ||
1467 (td->io_ops->flags & FIO_UNIDIR))
1472 fio_gettime(&td->start, NULL);
1474 do_verify(td, verify_bytes);
1476 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1478 if (td->error || td->terminate)
1482 update_rusage_stat(td);
1483 td->ts.runtime[DDIR_READ] = (td->ts.runtime[DDIR_READ] + 999) / 1000;
1484 td->ts.runtime[DDIR_WRITE] = (td->ts.runtime[DDIR_WRITE] + 999) / 1000;
1485 td->ts.runtime[DDIR_TRIM] = (td->ts.runtime[DDIR_TRIM] + 999) / 1000;
1486 td->ts.total_run_time = mtime_since_now(&td->epoch);
1487 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1488 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1489 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1491 fio_unpin_memory(td);
1493 fio_mutex_down(writeout_mutex);
1496 if (o->bw_log_file) {
1497 finish_log_named(td, td->bw_log,
1498 o->bw_log_file, "bw");
1500 finish_log(td, td->bw_log, "bw");
1503 if (o->lat_log_file) {
1504 finish_log_named(td, td->lat_log,
1505 o->lat_log_file, "lat");
1507 finish_log(td, td->lat_log, "lat");
1510 if (o->lat_log_file) {
1511 finish_log_named(td, td->slat_log,
1512 o->lat_log_file, "slat");
1514 finish_log(td, td->slat_log, "slat");
1517 if (o->lat_log_file) {
1518 finish_log_named(td, td->clat_log,
1519 o->lat_log_file, "clat");
1521 finish_log(td, td->clat_log, "clat");
1524 if (o->iops_log_file) {
1525 finish_log_named(td, td->iops_log,
1526 o->iops_log_file, "iops");
1528 finish_log(td, td->iops_log, "iops");
1531 fio_mutex_up(writeout_mutex);
1532 if (o->exec_postrun)
1533 exec_string(o, o->exec_postrun, (const char *)"postrun");
1535 if (exitall_on_terminate)
1536 fio_terminate_threads(td->groupid);
1540 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1543 if (o->verify_async)
1544 verify_async_exit(td);
1546 close_and_free_files(td);
1549 cgroup_shutdown(td, &cgroup_mnt);
1551 if (o->cpumask_set) {
1552 int ret = fio_cpuset_exit(&o->cpumask);
1554 td_verror(td, ret, "fio_cpuset_exit");
1558 * do this very late, it will log file closing as well
1560 if (o->write_iolog_file)
1561 write_iolog_close(td);
1563 fio_mutex_remove(td->rusage_sem);
1564 td->rusage_sem = NULL;
1566 fio_mutex_remove(td->mutex);
1569 td_set_runstate(td, TD_EXITED);
1570 return (void *) (uintptr_t) td->error;
1575 * We cannot pass the td data into a forked process, so attach the td and
1576 * pass it to the thread worker.
1578 static int fork_main(int shmid, int offset)
1580 struct thread_data *td;
1584 data = shmat(shmid, NULL, 0);
1585 if (data == (void *) -1) {
1593 * HP-UX inherits shm mappings?
1598 td = data + offset * sizeof(struct thread_data);
1599 ret = thread_main(td);
1601 return (int) (uintptr_t) ret;
1605 * Run over the job map and reap the threads that have exited, if any.
1607 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1608 unsigned int *m_rate)
1610 struct thread_data *td;
1611 unsigned int cputhreads, realthreads, pending;
1615 * reap exited threads (TD_EXITED -> TD_REAPED)
1617 realthreads = pending = cputhreads = 0;
1618 for_each_td(td, i) {
1622 * ->io_ops is NULL for a thread that has closed its
1625 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1634 if (td->runstate == TD_REAPED)
1636 if (td->o.use_thread) {
1637 if (td->runstate == TD_EXITED) {
1638 td_set_runstate(td, TD_REAPED);
1645 if (td->runstate == TD_EXITED)
1649 * check if someone quit or got killed in an unusual way
1651 ret = waitpid(td->pid, &status, flags);
1653 if (errno == ECHILD) {
1654 log_err("fio: pid=%d disappeared %d\n",
1655 (int) td->pid, td->runstate);
1657 td_set_runstate(td, TD_REAPED);
1661 } else if (ret == td->pid) {
1662 if (WIFSIGNALED(status)) {
1663 int sig = WTERMSIG(status);
1665 if (sig != SIGTERM && sig != SIGUSR2)
1666 log_err("fio: pid=%d, got signal=%d\n",
1667 (int) td->pid, sig);
1669 td_set_runstate(td, TD_REAPED);
1672 if (WIFEXITED(status)) {
1673 if (WEXITSTATUS(status) && !td->error)
1674 td->error = WEXITSTATUS(status);
1676 td_set_runstate(td, TD_REAPED);
1682 * thread is not dead, continue
1688 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1689 (*t_rate) -= ddir_rw_sum(td->o.rate);
1696 done_secs += mtime_since_now(&td->epoch) / 1000;
1697 profile_td_exit(td);
1700 if (*nr_running == cputhreads && !pending && realthreads)
1701 fio_terminate_threads(TERMINATE_ALL);
1704 static void do_usleep(unsigned int usecs)
1706 check_for_running_stats();
1711 * Main function for kicking off and reaping jobs, as needed.
1713 static void run_threads(void)
1715 struct thread_data *td;
1716 unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
1719 if (fio_gtod_offload && fio_start_gtod_thread())
1722 fio_idle_prof_init();
1726 nr_thread = nr_process = 0;
1727 for_each_td(td, i) {
1728 if (td->o.use_thread)
1734 if (output_format == FIO_OUTPUT_NORMAL) {
1735 log_info("Starting ");
1737 log_info("%d thread%s", nr_thread,
1738 nr_thread > 1 ? "s" : "");
1742 log_info("%d process%s", nr_process,
1743 nr_process > 1 ? "es" : "");
1749 todo = thread_number;
1752 m_rate = t_rate = 0;
1754 for_each_td(td, i) {
1755 print_status_init(td->thread_number - 1);
1757 if (!td->o.create_serialize)
1761 * do file setup here so it happens sequentially,
1762 * we don't want X number of threads getting their
1763 * client data interspersed on disk
1765 if (setup_files(td)) {
1768 log_err("fio: pid=%d, err=%d/%s\n",
1769 (int) td->pid, td->error, td->verror);
1770 td_set_runstate(td, TD_REAPED);
1777 * for sharing to work, each job must always open
1778 * its own files. so close them, if we opened them
1781 for_each_file(td, f, j) {
1782 if (fio_file_open(f))
1783 td_io_close_file(td, f);
1788 /* start idle threads before io threads start to run */
1789 fio_idle_prof_start();
1794 struct thread_data *map[REAL_MAX_JOBS];
1795 struct timeval this_start;
1796 int this_jobs = 0, left;
1799 * create threads (TD_NOT_CREATED -> TD_CREATED)
1801 for_each_td(td, i) {
1802 if (td->runstate != TD_NOT_CREATED)
1806 * never got a chance to start, killed by other
1807 * thread for some reason
1809 if (td->terminate) {
1814 if (td->o.start_delay) {
1815 spent = utime_since_genesis();
1817 if (td->o.start_delay > spent)
1821 if (td->o.stonewall && (nr_started || nr_running)) {
1822 dprint(FD_PROCESS, "%s: stonewall wait\n",
1829 td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED);
1830 td->update_rusage = 0;
1833 * Set state to created. Thread will transition
1834 * to TD_INITIALIZED when it's done setting up.
1836 td_set_runstate(td, TD_CREATED);
1837 map[this_jobs++] = td;
1840 if (td->o.use_thread) {
1843 dprint(FD_PROCESS, "will pthread_create\n");
1844 ret = pthread_create(&td->thread, NULL,
1847 log_err("pthread_create: %s\n",
1852 ret = pthread_detach(td->thread);
1854 log_err("pthread_detach: %s",
1858 dprint(FD_PROCESS, "will fork\n");
1861 int ret = fork_main(shm_id, i);
1864 } else if (i == fio_debug_jobno)
1865 *fio_debug_jobp = pid;
1867 dprint(FD_MUTEX, "wait on startup_mutex\n");
1868 if (fio_mutex_down_timeout(startup_mutex, 10)) {
1869 log_err("fio: job startup hung? exiting.\n");
1870 fio_terminate_threads(TERMINATE_ALL);
1875 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1879 * Wait for the started threads to transition to
1882 fio_gettime(&this_start, NULL);
1884 while (left && !fio_abort) {
1885 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1890 for (i = 0; i < this_jobs; i++) {
1894 if (td->runstate == TD_INITIALIZED) {
1897 } else if (td->runstate >= TD_EXITED) {
1901 nr_running++; /* work-around... */
1907 log_err("fio: %d job%s failed to start\n", left,
1908 left > 1 ? "s" : "");
1909 for (i = 0; i < this_jobs; i++) {
1913 kill(td->pid, SIGTERM);
1919 * start created threads (TD_INITIALIZED -> TD_RUNNING).
1921 for_each_td(td, i) {
1922 if (td->runstate != TD_INITIALIZED)
1925 if (in_ramp_time(td))
1926 td_set_runstate(td, TD_RAMP);
1928 td_set_runstate(td, TD_RUNNING);
1931 m_rate += ddir_rw_sum(td->o.ratemin);
1932 t_rate += ddir_rw_sum(td->o.rate);
1934 fio_mutex_up(td->mutex);
1937 reap_threads(&nr_running, &t_rate, &m_rate);
1943 while (nr_running) {
1944 reap_threads(&nr_running, &t_rate, &m_rate);
1948 fio_idle_prof_stop();
1953 void wait_for_disk_thread_exit(void)
1955 fio_mutex_down(disk_thread_mutex);
1958 static void free_disk_util(void)
1960 disk_util_start_exit();
1961 wait_for_disk_thread_exit();
1962 disk_util_prune_entries();
1965 static void *disk_thread_main(void *data)
1969 fio_mutex_up(startup_mutex);
1971 while (threads && !ret) {
1972 usleep(DISK_UTIL_MSEC * 1000);
1975 ret = update_io_ticks();
1978 print_thread_status();
1981 fio_mutex_up(disk_thread_mutex);
1985 static int create_disk_util_thread(void)
1991 disk_thread_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
1993 ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
1995 fio_mutex_remove(disk_thread_mutex);
1996 log_err("Can't create disk util thread: %s\n", strerror(ret));
2000 ret = pthread_detach(disk_util_thread);
2002 fio_mutex_remove(disk_thread_mutex);
2003 log_err("Can't detatch disk util thread: %s\n", strerror(ret));
2007 dprint(FD_MUTEX, "wait on startup_mutex\n");
2008 fio_mutex_down(startup_mutex);
2009 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2013 int fio_backend(void)
2015 struct thread_data *td;
2019 if (load_profile(exec_profile))
2022 exec_profile = NULL;
2028 setup_log(&agg_io_log[DDIR_READ], 0, IO_LOG_TYPE_BW);
2029 setup_log(&agg_io_log[DDIR_WRITE], 0, IO_LOG_TYPE_BW);
2030 setup_log(&agg_io_log[DDIR_TRIM], 0, IO_LOG_TYPE_BW);
2033 startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
2034 if (startup_mutex == NULL)
2036 writeout_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
2037 if (writeout_mutex == NULL)
2042 create_disk_util_thread();
2044 cgroup_list = smalloc(sizeof(*cgroup_list));
2045 INIT_FLIST_HEAD(cgroup_list);
2052 __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
2053 __finish_log(agg_io_log[DDIR_WRITE],
2054 "agg-write_bw.log");
2055 __finish_log(agg_io_log[DDIR_TRIM],
2056 "agg-write_bw.log");
2061 fio_options_free(td);
2064 cgroup_kill(cgroup_list);
2068 fio_mutex_remove(startup_mutex);
2069 fio_mutex_remove(writeout_mutex);
2070 fio_mutex_remove(disk_thread_mutex);