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 (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
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 int io_bytes_exceeded(struct thread_data *td)
628 unsigned long long bytes;
631 bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE];
632 else if (td_write(td))
633 bytes = td->this_io_bytes[DDIR_WRITE];
634 else if (td_read(td))
635 bytes = td->this_io_bytes[DDIR_READ];
637 bytes = td->this_io_bytes[DDIR_TRIM];
639 return bytes >= td->o.size;
643 * Main IO worker function. It retrieves io_u's to process and queues
644 * and reaps them, checking for rate and errors along the way.
646 * Returns number of bytes written and trimmed.
648 static uint64_t do_io(struct thread_data *td)
650 uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
653 uint64_t total_bytes, bytes_issued = 0;
655 if (in_ramp_time(td))
656 td_set_runstate(td, TD_RAMP);
658 td_set_runstate(td, TD_RUNNING);
663 * If verify_backlog is enabled, we'll run the verify in this
664 * handler as well. For that case, we may need up to twice the
667 total_bytes = td->o.size;
668 if (td->o.verify != VERIFY_NONE &&
669 (td_write(td) && td->o.verify_backlog))
670 total_bytes += td->o.size;
672 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
673 (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td) ||
675 struct timeval comp_time;
681 check_update_rusage(td);
683 if (td->terminate || td->done)
688 if (runtime_exceeded(td, &td->tv_cache)) {
689 __update_tv_cache(td);
690 if (runtime_exceeded(td, &td->tv_cache)) {
696 if (flow_threshold_exceeded(td))
699 if (bytes_issued >= total_bytes)
703 if (IS_ERR_OR_NULL(io_u)) {
704 int err = PTR_ERR(io_u);
711 if (td->o.latency_target)
719 * Add verification end_io handler if:
720 * - Asked to verify (!td_rw(td))
721 * - Or the io_u is from our verify list (mixed write/ver)
723 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
724 ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
726 if (!td->o.verify_pattern_bytes) {
727 io_u->rand_seed = __rand(&td->__verify_state);
728 if (sizeof(int) != sizeof(long *))
729 io_u->rand_seed *= __rand(&td->__verify_state);
732 if (td->o.verify_async)
733 io_u->end_io = verify_io_u_async;
735 io_u->end_io = verify_io_u;
736 td_set_runstate(td, TD_VERIFYING);
737 } else if (in_ramp_time(td))
738 td_set_runstate(td, TD_RAMP);
740 td_set_runstate(td, TD_RUNNING);
743 * Always log IO before it's issued, so we know the specific
744 * order of it. The logged unit will track when the IO has
747 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
749 td->o.verify != VERIFY_NONE &&
750 !td->o.experimental_verify)
751 log_io_piece(td, io_u);
753 ret = td_io_queue(td, io_u);
755 case FIO_Q_COMPLETED:
758 clear_io_u(td, io_u);
759 } else if (io_u->resid) {
760 int bytes = io_u->xfer_buflen - io_u->resid;
761 struct fio_file *f = io_u->file;
763 bytes_issued += bytes;
768 td_verror(td, EIO, "full resid");
773 io_u->xfer_buflen = io_u->resid;
774 io_u->xfer_buf += bytes;
775 io_u->offset += bytes;
777 if (ddir_rw(io_u->ddir))
778 td->ts.short_io_u[io_u->ddir]++;
780 if (io_u->offset == f->real_file_size)
783 requeue_io_u(td, &io_u);
786 if (__should_check_rate(td, DDIR_READ) ||
787 __should_check_rate(td, DDIR_WRITE) ||
788 __should_check_rate(td, DDIR_TRIM))
789 fio_gettime(&comp_time, NULL);
791 ret = io_u_sync_complete(td, io_u, bytes_done);
794 bytes_issued += io_u->xfer_buflen;
799 * if the engine doesn't have a commit hook,
800 * the io_u is really queued. if it does have such
801 * a hook, it has to call io_u_queued() itself.
803 if (td->io_ops->commit == NULL)
804 io_u_queued(td, io_u);
805 bytes_issued += io_u->xfer_buflen;
808 requeue_io_u(td, &io_u);
809 ret2 = td_io_commit(td);
819 if (break_on_this_error(td, ddir, &ret))
823 * See if we need to complete some commands. Note that we
824 * can get BUSY even without IO queued, if the system is
828 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
829 if (full || !td->o.iodepth_batch_complete) {
830 min_evts = min(td->o.iodepth_batch_complete,
833 * if the queue is full, we MUST reap at least 1 event
835 if (full && !min_evts)
838 if (__should_check_rate(td, DDIR_READ) ||
839 __should_check_rate(td, DDIR_WRITE) ||
840 __should_check_rate(td, DDIR_TRIM))
841 fio_gettime(&comp_time, NULL);
844 ret = io_u_queued_complete(td, min_evts, bytes_done);
848 } while (full && (td->cur_depth > td->o.iodepth_low));
853 if (!ddir_rw_sum(bytes_done) && !(td->io_ops->flags & FIO_NOIO))
856 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
857 if (check_min_rate(td, &comp_time, bytes_done)) {
858 if (exitall_on_terminate)
859 fio_terminate_threads(td->groupid);
860 td_verror(td, EIO, "check_min_rate");
864 if (!in_ramp_time(td) && td->o.latency_target)
865 lat_target_check(td);
867 if (td->o.thinktime) {
868 unsigned long long b;
870 b = ddir_rw_sum(td->io_blocks);
871 if (!(b % td->o.thinktime_blocks)) {
876 if (td->o.thinktime_spin)
877 usec_spin(td->o.thinktime_spin);
879 left = td->o.thinktime - td->o.thinktime_spin;
881 usec_sleep(td, left);
886 check_update_rusage(td);
888 if (td->trim_entries)
889 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
891 if (td->o.fill_device && td->error == ENOSPC) {
900 ret = io_u_queued_complete(td, i, bytes_done);
901 if (td->o.fill_device && td->error == ENOSPC)
905 if (should_fsync(td) && td->o.end_fsync) {
906 td_set_runstate(td, TD_FSYNCING);
908 for_each_file(td, f, i) {
909 if (!fio_file_fsync(td, f))
912 log_err("fio: end_fsync failed for file %s\n",
917 cleanup_pending_aio(td);
920 * stop job if we failed doing any IO
922 if (!ddir_rw_sum(td->this_io_bytes))
925 return bytes_done[DDIR_WRITE] + bytes_done[DDIR_TRIM];
928 static void cleanup_io_u(struct thread_data *td)
932 while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
934 if (td->io_ops->io_u_free)
935 td->io_ops->io_u_free(td, io_u);
937 fio_memfree(io_u, sizeof(*io_u));
942 io_u_rexit(&td->io_u_requeues);
943 io_u_qexit(&td->io_u_freelist);
944 io_u_qexit(&td->io_u_all);
947 static int init_io_u(struct thread_data *td)
950 unsigned int max_bs, min_write;
951 int cl_align, i, max_units;
952 int data_xfer = 1, err;
955 max_units = td->o.iodepth;
956 max_bs = td_max_bs(td);
957 min_write = td->o.min_bs[DDIR_WRITE];
958 td->orig_buffer_size = (unsigned long long) max_bs
959 * (unsigned long long) max_units;
961 if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
965 err += io_u_rinit(&td->io_u_requeues, td->o.iodepth);
966 err += io_u_qinit(&td->io_u_freelist, td->o.iodepth);
967 err += io_u_qinit(&td->io_u_all, td->o.iodepth);
970 log_err("fio: failed setting up IO queues\n");
975 * if we may later need to do address alignment, then add any
976 * possible adjustment here so that we don't cause a buffer
977 * overflow later. this adjustment may be too much if we get
978 * lucky and the allocator gives us an aligned address.
980 if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
981 (td->io_ops->flags & FIO_RAWIO))
982 td->orig_buffer_size += page_mask + td->o.mem_align;
984 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
987 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
988 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
991 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
992 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
996 if (data_xfer && allocate_io_mem(td))
999 if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1000 (td->io_ops->flags & FIO_RAWIO))
1001 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
1003 p = td->orig_buffer;
1005 cl_align = os_cache_line_size();
1007 for (i = 0; i < max_units; i++) {
1013 ptr = fio_memalign(cl_align, sizeof(*io_u));
1015 log_err("fio: unable to allocate aligned memory\n");
1020 memset(io_u, 0, sizeof(*io_u));
1021 INIT_FLIST_HEAD(&io_u->verify_list);
1022 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1026 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1029 io_u_fill_buffer(td, io_u, min_write, max_bs);
1030 if (td_write(td) && td->o.verify_pattern_bytes) {
1032 * Fill the buffer with the pattern if we are
1033 * going to be doing writes.
1035 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1040 io_u->flags = IO_U_F_FREE;
1041 io_u_qpush(&td->io_u_freelist, io_u);
1044 * io_u never leaves this stack, used for iteration of all
1047 io_u_qpush(&td->io_u_all, io_u);
1049 if (td->io_ops->io_u_init) {
1050 int ret = td->io_ops->io_u_init(td, io_u);
1053 log_err("fio: failed to init engine data: %d\n", ret);
1064 static int switch_ioscheduler(struct thread_data *td)
1066 char tmp[256], tmp2[128];
1070 if (td->io_ops->flags & FIO_DISKLESSIO)
1073 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
1075 f = fopen(tmp, "r+");
1077 if (errno == ENOENT) {
1078 log_err("fio: os or kernel doesn't support IO scheduler"
1082 td_verror(td, errno, "fopen iosched");
1089 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1090 if (ferror(f) || ret != 1) {
1091 td_verror(td, errno, "fwrite");
1099 * Read back and check that the selected scheduler is now the default.
1101 ret = fread(tmp, 1, sizeof(tmp), f);
1102 if (ferror(f) || ret < 0) {
1103 td_verror(td, errno, "fread");
1108 sprintf(tmp2, "[%s]", td->o.ioscheduler);
1109 if (!strstr(tmp, tmp2)) {
1110 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1111 td_verror(td, EINVAL, "iosched_switch");
1120 static int keep_running(struct thread_data *td)
1124 if (td->o.time_based)
1131 if (td->o.size != -1ULL && ddir_rw_sum(td->io_bytes) < td->o.size) {
1135 * If the difference is less than the minimum IO size, we
1138 diff = td->o.size - ddir_rw_sum(td->io_bytes);
1139 if (diff < td_max_bs(td))
1142 if (fio_files_done(td))
1151 static int exec_string(struct thread_options *o, const char *string, const char *mode)
1153 int ret, newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
1156 str = malloc(newlen);
1157 sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
1159 log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1162 log_err("fio: exec of cmd <%s> failed\n", str);
1169 * Dry run to compute correct state of numberio for verification.
1171 static uint64_t do_dry_run(struct thread_data *td)
1173 uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
1175 td_set_runstate(td, TD_RUNNING);
1177 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1178 (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td)) {
1182 if (td->terminate || td->done)
1185 io_u = get_io_u(td);
1189 io_u->flags |= IO_U_F_FLIGHT;
1192 if (ddir_rw(acct_ddir(io_u)))
1193 td->io_issues[acct_ddir(io_u)]++;
1194 if (ddir_rw(io_u->ddir)) {
1195 io_u_mark_depth(td, 1);
1196 td->ts.total_io_u[io_u->ddir]++;
1199 ret = io_u_sync_complete(td, io_u, bytes_done);
1203 return bytes_done[DDIR_WRITE] + bytes_done[DDIR_TRIM];
1207 * Entry point for the thread based jobs. The process based jobs end up
1208 * here as well, after a little setup.
1210 static void *thread_main(void *data)
1212 unsigned long long elapsed;
1213 struct thread_data *td = data;
1214 struct thread_options *o = &td->o;
1215 pthread_condattr_t attr;
1219 if (!o->use_thread) {
1226 * fio_time_init() may not have been called yet if running as a server
1230 fio_local_clock_init(o->use_thread);
1232 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1235 fio_server_send_start(td);
1237 INIT_FLIST_HEAD(&td->io_log_list);
1238 INIT_FLIST_HEAD(&td->io_hist_list);
1239 INIT_FLIST_HEAD(&td->verify_list);
1240 INIT_FLIST_HEAD(&td->trim_list);
1241 INIT_FLIST_HEAD(&td->next_rand_list);
1242 pthread_mutex_init(&td->io_u_lock, NULL);
1243 td->io_hist_tree = RB_ROOT;
1245 pthread_condattr_init(&attr);
1246 pthread_cond_init(&td->verify_cond, &attr);
1247 pthread_cond_init(&td->free_cond, &attr);
1249 td_set_runstate(td, TD_INITIALIZED);
1250 dprint(FD_MUTEX, "up startup_mutex\n");
1251 fio_mutex_up(startup_mutex);
1252 dprint(FD_MUTEX, "wait on td->mutex\n");
1253 fio_mutex_down(td->mutex);
1254 dprint(FD_MUTEX, "done waiting on td->mutex\n");
1257 * A new gid requires privilege, so we need to do this before setting
1260 if (o->gid != -1U && setgid(o->gid)) {
1261 td_verror(td, errno, "setgid");
1264 if (o->uid != -1U && setuid(o->uid)) {
1265 td_verror(td, errno, "setuid");
1270 * If we have a gettimeofday() thread, make sure we exclude that
1271 * thread from this job
1274 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1277 * Set affinity first, in case it has an impact on the memory
1280 if (o->cpumask_set) {
1281 ret = fio_setaffinity(td->pid, o->cpumask);
1283 td_verror(td, errno, "cpu_set_affinity");
1288 #ifdef CONFIG_LIBNUMA
1289 /* numa node setup */
1290 if (o->numa_cpumask_set || o->numa_memmask_set) {
1293 if (numa_available() < 0) {
1294 td_verror(td, errno, "Does not support NUMA API\n");
1298 if (o->numa_cpumask_set) {
1299 ret = numa_run_on_node_mask(o->numa_cpunodesmask);
1301 td_verror(td, errno, \
1302 "numa_run_on_node_mask failed\n");
1307 if (o->numa_memmask_set) {
1309 switch (o->numa_mem_mode) {
1310 case MPOL_INTERLEAVE:
1311 numa_set_interleave_mask(o->numa_memnodesmask);
1314 numa_set_membind(o->numa_memnodesmask);
1317 numa_set_localalloc();
1319 case MPOL_PREFERRED:
1320 numa_set_preferred(o->numa_mem_prefer_node);
1331 if (fio_pin_memory(td))
1335 * May alter parameters that init_io_u() will use, so we need to
1344 if (o->verify_async && verify_async_init(td))
1348 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1350 td_verror(td, errno, "ioprio_set");
1355 if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1359 if (nice(o->nice) == -1 && errno != 0) {
1360 td_verror(td, errno, "nice");
1364 if (o->ioscheduler && switch_ioscheduler(td))
1367 if (!o->create_serialize && setup_files(td))
1373 if (init_random_map(td))
1376 if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
1380 if (pre_read_files(td) < 0)
1384 fio_verify_init(td);
1386 fio_gettime(&td->epoch, NULL);
1387 fio_getrusage(&td->ru_start);
1389 while (keep_running(td)) {
1390 uint64_t verify_bytes;
1392 fio_gettime(&td->start, NULL);
1393 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1394 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1395 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1397 if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1398 o->ratemin[DDIR_TRIM]) {
1399 memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1400 sizeof(td->bw_sample_time));
1401 memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1402 sizeof(td->bw_sample_time));
1403 memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1404 sizeof(td->bw_sample_time));
1410 prune_io_piece_log(td);
1412 if (td->o.verify_only && (td_write(td) || td_rw(td)))
1413 verify_bytes = do_dry_run(td);
1415 verify_bytes = do_io(td);
1419 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1420 elapsed = utime_since_now(&td->start);
1421 td->ts.runtime[DDIR_READ] += elapsed;
1423 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1424 elapsed = utime_since_now(&td->start);
1425 td->ts.runtime[DDIR_WRITE] += elapsed;
1427 if (td_trim(td) && td->io_bytes[DDIR_TRIM]) {
1428 elapsed = utime_since_now(&td->start);
1429 td->ts.runtime[DDIR_TRIM] += elapsed;
1432 if (td->error || td->terminate)
1435 if (!o->do_verify ||
1436 o->verify == VERIFY_NONE ||
1437 (td->io_ops->flags & FIO_UNIDIR))
1442 fio_gettime(&td->start, NULL);
1444 do_verify(td, verify_bytes);
1446 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1448 if (td->error || td->terminate)
1452 update_rusage_stat(td);
1453 td->ts.runtime[DDIR_READ] = (td->ts.runtime[DDIR_READ] + 999) / 1000;
1454 td->ts.runtime[DDIR_WRITE] = (td->ts.runtime[DDIR_WRITE] + 999) / 1000;
1455 td->ts.runtime[DDIR_TRIM] = (td->ts.runtime[DDIR_TRIM] + 999) / 1000;
1456 td->ts.total_run_time = mtime_since_now(&td->epoch);
1457 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1458 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1459 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1461 fio_unpin_memory(td);
1463 fio_mutex_down(writeout_mutex);
1465 if (o->bw_log_file) {
1466 finish_log_named(td, td->bw_log,
1467 o->bw_log_file, "bw");
1469 finish_log(td, td->bw_log, "bw");
1472 if (o->lat_log_file) {
1473 finish_log_named(td, td->lat_log,
1474 o->lat_log_file, "lat");
1476 finish_log(td, td->lat_log, "lat");
1479 if (o->lat_log_file) {
1480 finish_log_named(td, td->slat_log,
1481 o->lat_log_file, "slat");
1483 finish_log(td, td->slat_log, "slat");
1486 if (o->lat_log_file) {
1487 finish_log_named(td, td->clat_log,
1488 o->lat_log_file, "clat");
1490 finish_log(td, td->clat_log, "clat");
1493 if (o->iops_log_file) {
1494 finish_log_named(td, td->iops_log,
1495 o->iops_log_file, "iops");
1497 finish_log(td, td->iops_log, "iops");
1500 fio_mutex_up(writeout_mutex);
1501 if (o->exec_postrun)
1502 exec_string(o, o->exec_postrun, (const char *)"postrun");
1504 if (exitall_on_terminate)
1505 fio_terminate_threads(td->groupid);
1509 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1512 if (o->verify_async)
1513 verify_async_exit(td);
1515 close_and_free_files(td);
1518 cgroup_shutdown(td, &cgroup_mnt);
1520 if (o->cpumask_set) {
1521 int ret = fio_cpuset_exit(&o->cpumask);
1523 td_verror(td, ret, "fio_cpuset_exit");
1527 * do this very late, it will log file closing as well
1529 if (o->write_iolog_file)
1530 write_iolog_close(td);
1532 fio_mutex_remove(td->rusage_sem);
1533 td->rusage_sem = NULL;
1535 fio_mutex_remove(td->mutex);
1538 td_set_runstate(td, TD_EXITED);
1539 return (void *) (uintptr_t) td->error;
1544 * We cannot pass the td data into a forked process, so attach the td and
1545 * pass it to the thread worker.
1547 static int fork_main(int shmid, int offset)
1549 struct thread_data *td;
1553 data = shmat(shmid, NULL, 0);
1554 if (data == (void *) -1) {
1562 * HP-UX inherits shm mappings?
1567 td = data + offset * sizeof(struct thread_data);
1568 ret = thread_main(td);
1570 return (int) (uintptr_t) ret;
1574 * Run over the job map and reap the threads that have exited, if any.
1576 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1577 unsigned int *m_rate)
1579 struct thread_data *td;
1580 unsigned int cputhreads, realthreads, pending;
1584 * reap exited threads (TD_EXITED -> TD_REAPED)
1586 realthreads = pending = cputhreads = 0;
1587 for_each_td(td, i) {
1591 * ->io_ops is NULL for a thread that has closed its
1594 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1603 if (td->runstate == TD_REAPED)
1605 if (td->o.use_thread) {
1606 if (td->runstate == TD_EXITED) {
1607 td_set_runstate(td, TD_REAPED);
1614 if (td->runstate == TD_EXITED)
1618 * check if someone quit or got killed in an unusual way
1620 ret = waitpid(td->pid, &status, flags);
1622 if (errno == ECHILD) {
1623 log_err("fio: pid=%d disappeared %d\n",
1624 (int) td->pid, td->runstate);
1626 td_set_runstate(td, TD_REAPED);
1630 } else if (ret == td->pid) {
1631 if (WIFSIGNALED(status)) {
1632 int sig = WTERMSIG(status);
1634 if (sig != SIGTERM && sig != SIGUSR2)
1635 log_err("fio: pid=%d, got signal=%d\n",
1636 (int) td->pid, sig);
1638 td_set_runstate(td, TD_REAPED);
1641 if (WIFEXITED(status)) {
1642 if (WEXITSTATUS(status) && !td->error)
1643 td->error = WEXITSTATUS(status);
1645 td_set_runstate(td, TD_REAPED);
1651 * thread is not dead, continue
1657 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1658 (*t_rate) -= ddir_rw_sum(td->o.rate);
1665 done_secs += mtime_since_now(&td->epoch) / 1000;
1666 profile_td_exit(td);
1669 if (*nr_running == cputhreads && !pending && realthreads)
1670 fio_terminate_threads(TERMINATE_ALL);
1673 static void do_usleep(unsigned int usecs)
1675 check_for_running_stats();
1680 * Main function for kicking off and reaping jobs, as needed.
1682 static void run_threads(void)
1684 struct thread_data *td;
1685 unsigned long spent;
1686 unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
1688 if (fio_gtod_offload && fio_start_gtod_thread())
1691 fio_idle_prof_init();
1695 nr_thread = nr_process = 0;
1696 for_each_td(td, i) {
1697 if (td->o.use_thread)
1703 if (output_format == FIO_OUTPUT_NORMAL) {
1704 log_info("Starting ");
1706 log_info("%d thread%s", nr_thread,
1707 nr_thread > 1 ? "s" : "");
1711 log_info("%d process%s", nr_process,
1712 nr_process > 1 ? "es" : "");
1718 todo = thread_number;
1721 m_rate = t_rate = 0;
1723 for_each_td(td, i) {
1724 print_status_init(td->thread_number - 1);
1726 if (!td->o.create_serialize)
1730 * do file setup here so it happens sequentially,
1731 * we don't want X number of threads getting their
1732 * client data interspersed on disk
1734 if (setup_files(td)) {
1737 log_err("fio: pid=%d, err=%d/%s\n",
1738 (int) td->pid, td->error, td->verror);
1739 td_set_runstate(td, TD_REAPED);
1746 * for sharing to work, each job must always open
1747 * its own files. so close them, if we opened them
1750 for_each_file(td, f, j) {
1751 if (fio_file_open(f))
1752 td_io_close_file(td, f);
1757 /* start idle threads before io threads start to run */
1758 fio_idle_prof_start();
1763 struct thread_data *map[REAL_MAX_JOBS];
1764 struct timeval this_start;
1765 int this_jobs = 0, left;
1768 * create threads (TD_NOT_CREATED -> TD_CREATED)
1770 for_each_td(td, i) {
1771 if (td->runstate != TD_NOT_CREATED)
1775 * never got a chance to start, killed by other
1776 * thread for some reason
1778 if (td->terminate) {
1783 if (td->o.start_delay) {
1784 spent = mtime_since_genesis();
1786 if (td->o.start_delay * 1000 > spent)
1790 if (td->o.stonewall && (nr_started || nr_running)) {
1791 dprint(FD_PROCESS, "%s: stonewall wait\n",
1798 td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED);
1799 td->update_rusage = 0;
1802 * Set state to created. Thread will transition
1803 * to TD_INITIALIZED when it's done setting up.
1805 td_set_runstate(td, TD_CREATED);
1806 map[this_jobs++] = td;
1809 if (td->o.use_thread) {
1812 dprint(FD_PROCESS, "will pthread_create\n");
1813 ret = pthread_create(&td->thread, NULL,
1816 log_err("pthread_create: %s\n",
1821 ret = pthread_detach(td->thread);
1823 log_err("pthread_detach: %s",
1827 dprint(FD_PROCESS, "will fork\n");
1830 int ret = fork_main(shm_id, i);
1833 } else if (i == fio_debug_jobno)
1834 *fio_debug_jobp = pid;
1836 dprint(FD_MUTEX, "wait on startup_mutex\n");
1837 if (fio_mutex_down_timeout(startup_mutex, 10)) {
1838 log_err("fio: job startup hung? exiting.\n");
1839 fio_terminate_threads(TERMINATE_ALL);
1844 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1848 * Wait for the started threads to transition to
1851 fio_gettime(&this_start, NULL);
1853 while (left && !fio_abort) {
1854 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1859 for (i = 0; i < this_jobs; i++) {
1863 if (td->runstate == TD_INITIALIZED) {
1866 } else if (td->runstate >= TD_EXITED) {
1870 nr_running++; /* work-around... */
1876 log_err("fio: %d job%s failed to start\n", left,
1877 left > 1 ? "s" : "");
1878 for (i = 0; i < this_jobs; i++) {
1882 kill(td->pid, SIGTERM);
1888 * start created threads (TD_INITIALIZED -> TD_RUNNING).
1890 for_each_td(td, i) {
1891 if (td->runstate != TD_INITIALIZED)
1894 if (in_ramp_time(td))
1895 td_set_runstate(td, TD_RAMP);
1897 td_set_runstate(td, TD_RUNNING);
1900 m_rate += ddir_rw_sum(td->o.ratemin);
1901 t_rate += ddir_rw_sum(td->o.rate);
1903 fio_mutex_up(td->mutex);
1906 reap_threads(&nr_running, &t_rate, &m_rate);
1912 while (nr_running) {
1913 reap_threads(&nr_running, &t_rate, &m_rate);
1917 fio_idle_prof_stop();
1922 void wait_for_disk_thread_exit(void)
1924 fio_mutex_down(disk_thread_mutex);
1927 static void free_disk_util(void)
1929 disk_util_start_exit();
1930 wait_for_disk_thread_exit();
1931 disk_util_prune_entries();
1934 static void *disk_thread_main(void *data)
1938 fio_mutex_up(startup_mutex);
1940 while (threads && !ret) {
1941 usleep(DISK_UTIL_MSEC * 1000);
1944 ret = update_io_ticks();
1947 print_thread_status();
1950 fio_mutex_up(disk_thread_mutex);
1954 static int create_disk_util_thread(void)
1960 disk_thread_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
1962 ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
1964 fio_mutex_remove(disk_thread_mutex);
1965 log_err("Can't create disk util thread: %s\n", strerror(ret));
1969 ret = pthread_detach(disk_util_thread);
1971 fio_mutex_remove(disk_thread_mutex);
1972 log_err("Can't detatch disk util thread: %s\n", strerror(ret));
1976 dprint(FD_MUTEX, "wait on startup_mutex\n");
1977 fio_mutex_down(startup_mutex);
1978 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1982 int fio_backend(void)
1984 struct thread_data *td;
1988 if (load_profile(exec_profile))
1991 exec_profile = NULL;
1997 setup_log(&agg_io_log[DDIR_READ], 0, IO_LOG_TYPE_BW);
1998 setup_log(&agg_io_log[DDIR_WRITE], 0, IO_LOG_TYPE_BW);
1999 setup_log(&agg_io_log[DDIR_TRIM], 0, IO_LOG_TYPE_BW);
2002 startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
2003 if (startup_mutex == NULL)
2005 writeout_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
2006 if (writeout_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(writeout_mutex);
2039 fio_mutex_remove(disk_thread_mutex);