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"
58 static pthread_t helper_thread;
59 static pthread_mutex_t helper_lock;
60 pthread_cond_t helper_cond;
61 int helper_do_stat = 0;
63 static struct fio_mutex *startup_mutex;
64 static struct flist_head *cgroup_list;
65 static char *cgroup_mnt;
66 static int exit_value;
67 static volatile int fio_abort;
68 static unsigned int nr_process = 0;
69 static unsigned int nr_thread = 0;
71 struct io_log *agg_io_log[DDIR_RWDIR_CNT];
74 unsigned int thread_number = 0;
75 unsigned int stat_number = 0;
78 unsigned long done_secs = 0;
79 volatile int helper_exit = 0;
81 #define PAGE_ALIGN(buf) \
82 (char *) (((uintptr_t) (buf) + page_mask) & ~page_mask)
84 #define JOB_START_TIMEOUT (5 * 1000)
86 static void sig_int(int sig)
90 fio_server_got_signal(sig);
92 log_info("\nfio: terminating on signal %d\n", sig);
97 fio_terminate_threads(TERMINATE_ALL);
101 static void sig_show_status(int sig)
103 show_running_run_stats();
106 static void set_sig_handlers(void)
108 struct sigaction act;
110 memset(&act, 0, sizeof(act));
111 act.sa_handler = sig_int;
112 act.sa_flags = SA_RESTART;
113 sigaction(SIGINT, &act, NULL);
115 memset(&act, 0, sizeof(act));
116 act.sa_handler = sig_int;
117 act.sa_flags = SA_RESTART;
118 sigaction(SIGTERM, &act, NULL);
120 /* Windows uses SIGBREAK as a quit signal from other applications */
122 memset(&act, 0, sizeof(act));
123 act.sa_handler = sig_int;
124 act.sa_flags = SA_RESTART;
125 sigaction(SIGBREAK, &act, NULL);
128 memset(&act, 0, sizeof(act));
129 act.sa_handler = sig_show_status;
130 act.sa_flags = SA_RESTART;
131 sigaction(SIGUSR1, &act, NULL);
134 memset(&act, 0, sizeof(act));
135 act.sa_handler = sig_int;
136 act.sa_flags = SA_RESTART;
137 sigaction(SIGPIPE, &act, NULL);
142 * Check if we are above the minimum rate given.
144 static int __check_min_rate(struct thread_data *td, struct timeval *now,
147 unsigned long long bytes = 0;
148 unsigned long iops = 0;
151 unsigned int ratemin = 0;
152 unsigned int rate_iops = 0;
153 unsigned int rate_iops_min = 0;
155 assert(ddir_rw(ddir));
157 if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
161 * allow a 2 second settle period in the beginning
163 if (mtime_since(&td->start, now) < 2000)
166 iops += td->this_io_blocks[ddir];
167 bytes += td->this_io_bytes[ddir];
168 ratemin += td->o.ratemin[ddir];
169 rate_iops += td->o.rate_iops[ddir];
170 rate_iops_min += td->o.rate_iops_min[ddir];
173 * if rate blocks is set, sample is running
175 if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
176 spent = mtime_since(&td->lastrate[ddir], now);
177 if (spent < td->o.ratecycle)
180 if (td->o.rate[ddir]) {
182 * check bandwidth specified rate
184 if (bytes < td->rate_bytes[ddir]) {
185 log_err("%s: min rate %u not met\n", td->o.name,
190 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
194 if (rate < ratemin ||
195 bytes < td->rate_bytes[ddir]) {
196 log_err("%s: min rate %u not met, got"
197 " %luKB/sec\n", td->o.name,
204 * checks iops specified rate
206 if (iops < rate_iops) {
207 log_err("%s: min iops rate %u not met\n",
208 td->o.name, rate_iops);
212 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
216 if (rate < rate_iops_min ||
217 iops < td->rate_blocks[ddir]) {
218 log_err("%s: min iops rate %u not met,"
219 " got %lu\n", td->o.name,
220 rate_iops_min, rate);
226 td->rate_bytes[ddir] = bytes;
227 td->rate_blocks[ddir] = iops;
228 memcpy(&td->lastrate[ddir], now, sizeof(*now));
232 static int check_min_rate(struct thread_data *td, struct timeval *now)
236 if (td->bytes_done[DDIR_READ])
237 ret |= __check_min_rate(td, now, DDIR_READ);
238 if (td->bytes_done[DDIR_WRITE])
239 ret |= __check_min_rate(td, now, DDIR_WRITE);
240 if (td->bytes_done[DDIR_TRIM])
241 ret |= __check_min_rate(td, now, DDIR_TRIM);
247 * When job exits, we can cancel the in-flight IO if we are using async
248 * io. Attempt to do so.
250 static void cleanup_pending_aio(struct thread_data *td)
255 * get immediately available events, if any
257 r = io_u_queued_complete(td, 0);
262 * now cancel remaining active events
264 if (td->io_ops->cancel) {
268 io_u_qiter(&td->io_u_all, io_u, i) {
269 if (io_u->flags & IO_U_F_FLIGHT) {
270 r = td->io_ops->cancel(td, io_u);
278 r = io_u_queued_complete(td, td->cur_depth);
282 * Helper to handle the final sync of a file. Works just like the normal
283 * io path, just does everything sync.
285 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
287 struct io_u *io_u = __get_io_u(td);
293 io_u->ddir = DDIR_SYNC;
296 if (td_io_prep(td, io_u)) {
302 ret = td_io_queue(td, io_u);
304 td_verror(td, io_u->error, "td_io_queue");
307 } else if (ret == FIO_Q_QUEUED) {
308 if (io_u_queued_complete(td, 1) < 0)
310 } else if (ret == FIO_Q_COMPLETED) {
312 td_verror(td, io_u->error, "td_io_queue");
316 if (io_u_sync_complete(td, io_u) < 0)
318 } else if (ret == FIO_Q_BUSY) {
319 if (td_io_commit(td))
327 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
331 if (fio_file_open(f))
332 return fio_io_sync(td, f);
334 if (td_io_open_file(td, f))
337 ret = fio_io_sync(td, f);
338 td_io_close_file(td, f);
342 static inline void __update_tv_cache(struct thread_data *td)
344 fio_gettime(&td->tv_cache, NULL);
347 static inline void update_tv_cache(struct thread_data *td)
349 if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
350 __update_tv_cache(td);
353 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
355 if (in_ramp_time(td))
359 if (utime_since(&td->epoch, t) >= td->o.timeout)
365 static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
370 if (ret < 0 || td->error) {
372 enum error_type_bit eb;
377 eb = td_error_type(ddir, err);
378 if (!(td->o.continue_on_error & (1 << eb)))
381 if (td_non_fatal_error(td, eb, err)) {
383 * Continue with the I/Os in case of
386 update_error_count(td, err);
390 } else if (td->o.fill_device && err == ENOSPC) {
392 * We expect to hit this error if
393 * fill_device option is set.
396 fio_mark_td_terminate(td);
400 * Stop the I/O in case of a fatal
403 update_error_count(td, err);
411 static void check_update_rusage(struct thread_data *td)
413 if (td->update_rusage) {
414 td->update_rusage = 0;
415 update_rusage_stat(td);
416 fio_mutex_up(td->rusage_sem);
420 static int wait_for_completions(struct thread_data *td, struct timeval *time)
422 const int full = queue_full(td);
427 * if the queue is full, we MUST reap at least 1 event
429 min_evts = min(td->o.iodepth_batch_complete, td->cur_depth);
430 if (full && !min_evts)
433 if (time && (__should_check_rate(td, DDIR_READ) ||
434 __should_check_rate(td, DDIR_WRITE) ||
435 __should_check_rate(td, DDIR_TRIM)))
436 fio_gettime(time, NULL);
439 ret = io_u_queued_complete(td, min_evts);
442 } while (full && (td->cur_depth > td->o.iodepth_low));
447 int io_queue_event(struct thread_data *td, struct io_u *io_u, int *ret,
448 enum fio_ddir ddir, uint64_t *bytes_issued, int from_verify,
449 struct timeval *comp_time)
454 case FIO_Q_COMPLETED:
457 clear_io_u(td, io_u);
458 } else if (io_u->resid) {
459 int bytes = io_u->xfer_buflen - io_u->resid;
460 struct fio_file *f = io_u->file;
463 *bytes_issued += bytes;
466 trim_io_piece(td, io_u);
473 unlog_io_piece(td, io_u);
474 td_verror(td, EIO, "full resid");
479 io_u->xfer_buflen = io_u->resid;
480 io_u->xfer_buf += bytes;
481 io_u->offset += bytes;
483 if (ddir_rw(io_u->ddir))
484 td->ts.short_io_u[io_u->ddir]++;
487 if (io_u->offset == f->real_file_size)
490 requeue_io_u(td, &io_u);
493 if (comp_time && (__should_check_rate(td, DDIR_READ) ||
494 __should_check_rate(td, DDIR_WRITE) ||
495 __should_check_rate(td, DDIR_TRIM)))
496 fio_gettime(comp_time, NULL);
498 *ret = io_u_sync_complete(td, io_u);
505 * if the engine doesn't have a commit hook,
506 * the io_u is really queued. if it does have such
507 * a hook, it has to call io_u_queued() itself.
509 if (td->io_ops->commit == NULL)
510 io_u_queued(td, io_u);
512 *bytes_issued += io_u->xfer_buflen;
516 unlog_io_piece(td, io_u);
517 requeue_io_u(td, &io_u);
518 ret2 = td_io_commit(td);
524 td_verror(td, -(*ret), "td_io_queue");
528 if (break_on_this_error(td, ddir, ret))
535 * The main verify engine. Runs over the writes we previously submitted,
536 * reads the blocks back in, and checks the crc/md5 of the data.
538 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
545 dprint(FD_VERIFY, "starting loop\n");
548 * sync io first and invalidate cache, to make sure we really
551 for_each_file(td, f, i) {
552 if (!fio_file_open(f))
554 if (fio_io_sync(td, f))
556 if (file_invalidate_cache(td, f))
560 check_update_rusage(td);
565 td_set_runstate(td, TD_VERIFYING);
568 while (!td->terminate) {
573 check_update_rusage(td);
575 if (runtime_exceeded(td, &td->tv_cache)) {
576 __update_tv_cache(td);
577 if (runtime_exceeded(td, &td->tv_cache)) {
578 fio_mark_td_terminate(td);
583 if (flow_threshold_exceeded(td))
586 if (!td->o.experimental_verify) {
587 io_u = __get_io_u(td);
591 if (get_next_verify(td, io_u)) {
596 if (td_io_prep(td, io_u)) {
601 if (ddir_rw_sum(td->bytes_done) + td->o.rw_min_bs > verify_bytes)
604 while ((io_u = get_io_u(td)) != NULL) {
612 * We are only interested in the places where
613 * we wrote or trimmed IOs. Turn those into
614 * reads for verification purposes.
616 if (io_u->ddir == DDIR_READ) {
618 * Pretend we issued it for rwmix
621 td->io_issues[DDIR_READ]++;
624 } else if (io_u->ddir == DDIR_TRIM) {
625 io_u->ddir = DDIR_READ;
626 io_u->flags |= IO_U_F_TRIMMED;
628 } else if (io_u->ddir == DDIR_WRITE) {
629 io_u->ddir = DDIR_READ;
641 if (verify_state_should_stop(td, io_u)) {
646 if (td->o.verify_async)
647 io_u->end_io = verify_io_u_async;
649 io_u->end_io = verify_io_u;
652 if (!td->o.disable_slat)
653 fio_gettime(&io_u->start_time, NULL);
655 ret = td_io_queue(td, io_u);
657 if (io_queue_event(td, io_u, &ret, ddir, NULL, 1, NULL))
661 * if we can queue more, do so. but check if there are
662 * completed io_u's first. Note that we can get BUSY even
663 * without IO queued, if the system is resource starved.
666 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
667 if (full || !td->o.iodepth_batch_complete)
668 ret = wait_for_completions(td, NULL);
674 check_update_rusage(td);
677 min_events = td->cur_depth;
680 ret = io_u_queued_complete(td, min_events);
682 cleanup_pending_aio(td);
684 td_set_runstate(td, TD_RUNNING);
686 dprint(FD_VERIFY, "exiting loop\n");
689 static unsigned int exceeds_number_ios(struct thread_data *td)
691 unsigned long long number_ios;
693 if (!td->o.number_ios)
696 number_ios = ddir_rw_sum(td->io_blocks);
697 number_ios += td->io_u_queued + td->io_u_in_flight;
699 return number_ios >= (td->o.number_ios * td->loops);
702 static int io_issue_bytes_exceeded(struct thread_data *td)
704 unsigned long long bytes, limit;
707 bytes = td->io_issue_bytes[DDIR_READ] + td->io_issue_bytes[DDIR_WRITE];
708 else if (td_write(td))
709 bytes = td->io_issue_bytes[DDIR_WRITE];
710 else if (td_read(td))
711 bytes = td->io_issue_bytes[DDIR_READ];
713 bytes = td->io_issue_bytes[DDIR_TRIM];
716 limit = td->o.io_limit;
721 return bytes >= limit || exceeds_number_ios(td);
724 static int io_complete_bytes_exceeded(struct thread_data *td)
726 unsigned long long bytes, limit;
729 bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE];
730 else if (td_write(td))
731 bytes = td->this_io_bytes[DDIR_WRITE];
732 else if (td_read(td))
733 bytes = td->this_io_bytes[DDIR_READ];
735 bytes = td->this_io_bytes[DDIR_TRIM];
738 limit = td->o.io_limit;
743 return bytes >= limit || exceeds_number_ios(td);
747 * Main IO worker function. It retrieves io_u's to process and queues
748 * and reaps them, checking for rate and errors along the way.
750 * Returns number of bytes written and trimmed.
752 static uint64_t do_io(struct thread_data *td)
756 uint64_t total_bytes, bytes_issued = 0;
758 if (in_ramp_time(td))
759 td_set_runstate(td, TD_RAMP);
761 td_set_runstate(td, TD_RUNNING);
765 total_bytes = td->o.size;
767 * Allow random overwrite workloads to write up to io_limit
768 * before starting verification phase as 'size' doesn't apply.
770 if (td_write(td) && td_random(td) && td->o.norandommap)
771 total_bytes = max(total_bytes, (uint64_t) td->o.io_limit);
773 * If verify_backlog is enabled, we'll run the verify in this
774 * handler as well. For that case, we may need up to twice the
777 if (td->o.verify != VERIFY_NONE &&
778 (td_write(td) && td->o.verify_backlog))
779 total_bytes += td->o.size;
781 /* In trimwrite mode, each byte is trimmed and then written, so
782 * allow total_bytes to be twice as big */
783 if (td_trimwrite(td))
784 total_bytes += td->total_io_size;
786 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
787 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
789 struct timeval comp_time;
794 check_update_rusage(td);
796 if (td->terminate || td->done)
801 if (runtime_exceeded(td, &td->tv_cache)) {
802 __update_tv_cache(td);
803 if (runtime_exceeded(td, &td->tv_cache)) {
804 fio_mark_td_terminate(td);
809 if (flow_threshold_exceeded(td))
812 if (bytes_issued >= total_bytes)
816 if (IS_ERR_OR_NULL(io_u)) {
817 int err = PTR_ERR(io_u);
824 if (td->o.latency_target)
832 * Add verification end_io handler if:
833 * - Asked to verify (!td_rw(td))
834 * - Or the io_u is from our verify list (mixed write/ver)
836 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
837 ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
839 if (!td->o.verify_pattern_bytes) {
840 io_u->rand_seed = __rand(&td->verify_state);
841 if (sizeof(int) != sizeof(long *))
842 io_u->rand_seed *= __rand(&td->verify_state);
845 if (verify_state_should_stop(td, io_u)) {
850 if (td->o.verify_async)
851 io_u->end_io = verify_io_u_async;
853 io_u->end_io = verify_io_u;
854 td_set_runstate(td, TD_VERIFYING);
855 } else if (in_ramp_time(td))
856 td_set_runstate(td, TD_RAMP);
858 td_set_runstate(td, TD_RUNNING);
861 * Always log IO before it's issued, so we know the specific
862 * order of it. The logged unit will track when the IO has
865 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
867 td->o.verify != VERIFY_NONE &&
868 !td->o.experimental_verify)
869 log_io_piece(td, io_u);
871 ret = td_io_queue(td, io_u);
873 if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 1, &comp_time))
877 * See if we need to complete some commands. Note that we
878 * can get BUSY even without IO queued, if the system is
882 full = queue_full(td) ||
883 (ret == FIO_Q_BUSY && td->cur_depth);
884 if (full || !td->o.iodepth_batch_complete)
885 ret = wait_for_completions(td, &comp_time);
888 if (!ddir_rw_sum(td->bytes_done) &&
889 !(td->io_ops->flags & FIO_NOIO))
892 if (!in_ramp_time(td) && should_check_rate(td)) {
893 if (check_min_rate(td, &comp_time)) {
894 if (exitall_on_terminate)
895 fio_terminate_threads(td->groupid);
896 td_verror(td, EIO, "check_min_rate");
900 if (!in_ramp_time(td) && td->o.latency_target)
901 lat_target_check(td);
903 if (td->o.thinktime) {
904 unsigned long long b;
906 b = ddir_rw_sum(td->io_blocks);
907 if (!(b % td->o.thinktime_blocks)) {
912 if (td->o.thinktime_spin)
913 usec_spin(td->o.thinktime_spin);
915 left = td->o.thinktime - td->o.thinktime_spin;
917 usec_sleep(td, left);
922 check_update_rusage(td);
924 if (td->trim_entries)
925 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
927 if (td->o.fill_device && td->error == ENOSPC) {
929 fio_mark_td_terminate(td);
936 ret = io_u_queued_complete(td, i);
937 if (td->o.fill_device && td->error == ENOSPC)
941 if (should_fsync(td) && td->o.end_fsync) {
942 td_set_runstate(td, TD_FSYNCING);
944 for_each_file(td, f, i) {
945 if (!fio_file_fsync(td, f))
948 log_err("fio: end_fsync failed for file %s\n",
953 cleanup_pending_aio(td);
956 * stop job if we failed doing any IO
958 if (!ddir_rw_sum(td->this_io_bytes))
961 return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
964 static void cleanup_io_u(struct thread_data *td)
968 while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
970 if (td->io_ops->io_u_free)
971 td->io_ops->io_u_free(td, io_u);
973 fio_memfree(io_u, sizeof(*io_u));
978 io_u_rexit(&td->io_u_requeues);
979 io_u_qexit(&td->io_u_freelist);
980 io_u_qexit(&td->io_u_all);
982 if (td->last_write_comp)
983 sfree(td->last_write_comp);
986 static int init_io_u(struct thread_data *td)
989 unsigned int max_bs, min_write;
990 int cl_align, i, max_units;
991 int data_xfer = 1, err;
994 max_units = td->o.iodepth;
995 max_bs = td_max_bs(td);
996 min_write = td->o.min_bs[DDIR_WRITE];
997 td->orig_buffer_size = (unsigned long long) max_bs
998 * (unsigned long long) max_units;
1000 if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
1004 err += io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1005 err += io_u_qinit(&td->io_u_freelist, td->o.iodepth);
1006 err += io_u_qinit(&td->io_u_all, td->o.iodepth);
1009 log_err("fio: failed setting up IO queues\n");
1014 * if we may later need to do address alignment, then add any
1015 * possible adjustment here so that we don't cause a buffer
1016 * overflow later. this adjustment may be too much if we get
1017 * lucky and the allocator gives us an aligned address.
1019 if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1020 (td->io_ops->flags & FIO_RAWIO))
1021 td->orig_buffer_size += page_mask + td->o.mem_align;
1023 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1026 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1027 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1030 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1031 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1035 if (data_xfer && allocate_io_mem(td))
1038 if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1039 (td->io_ops->flags & FIO_RAWIO))
1040 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
1042 p = td->orig_buffer;
1044 cl_align = os_cache_line_size();
1046 for (i = 0; i < max_units; i++) {
1052 ptr = fio_memalign(cl_align, sizeof(*io_u));
1054 log_err("fio: unable to allocate aligned memory\n");
1059 memset(io_u, 0, sizeof(*io_u));
1060 INIT_FLIST_HEAD(&io_u->verify_list);
1061 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1065 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1068 io_u_fill_buffer(td, io_u, min_write, max_bs);
1069 if (td_write(td) && td->o.verify_pattern_bytes) {
1071 * Fill the buffer with the pattern if we are
1072 * going to be doing writes.
1074 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1079 io_u->flags = IO_U_F_FREE;
1080 io_u_qpush(&td->io_u_freelist, io_u);
1083 * io_u never leaves this stack, used for iteration of all
1086 io_u_qpush(&td->io_u_all, io_u);
1088 if (td->io_ops->io_u_init) {
1089 int ret = td->io_ops->io_u_init(td, io_u);
1092 log_err("fio: failed to init engine data: %d\n", ret);
1100 if (td->o.verify != VERIFY_NONE) {
1101 td->last_write_comp = scalloc(max_units, sizeof(uint64_t));
1102 if (!td->last_write_comp) {
1103 log_err("fio: failed to alloc write comp data\n");
1111 static int switch_ioscheduler(struct thread_data *td)
1113 char tmp[256], tmp2[128];
1117 if (td->io_ops->flags & FIO_DISKLESSIO)
1120 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
1122 f = fopen(tmp, "r+");
1124 if (errno == ENOENT) {
1125 log_err("fio: os or kernel doesn't support IO scheduler"
1129 td_verror(td, errno, "fopen iosched");
1136 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1137 if (ferror(f) || ret != 1) {
1138 td_verror(td, errno, "fwrite");
1146 * Read back and check that the selected scheduler is now the default.
1148 ret = fread(tmp, sizeof(tmp), 1, f);
1149 if (ferror(f) || ret < 0) {
1150 td_verror(td, errno, "fread");
1154 tmp[sizeof(tmp) - 1] = '\0';
1157 sprintf(tmp2, "[%s]", td->o.ioscheduler);
1158 if (!strstr(tmp, tmp2)) {
1159 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1160 td_verror(td, EINVAL, "iosched_switch");
1169 static int keep_running(struct thread_data *td)
1171 unsigned long long limit;
1175 if (td->o.time_based)
1181 if (exceeds_number_ios(td))
1185 limit = td->o.io_limit;
1189 if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1193 * If the difference is less than the minimum IO size, we
1196 diff = limit - ddir_rw_sum(td->io_bytes);
1197 if (diff < td_max_bs(td))
1200 if (fio_files_done(td))
1209 static int exec_string(struct thread_options *o, const char *string, const char *mode)
1211 int ret, newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
1214 str = malloc(newlen);
1215 sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
1217 log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1220 log_err("fio: exec of cmd <%s> failed\n", str);
1227 * Dry run to compute correct state of numberio for verification.
1229 static uint64_t do_dry_run(struct thread_data *td)
1231 td_set_runstate(td, TD_RUNNING);
1233 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1234 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
1238 if (td->terminate || td->done)
1241 io_u = get_io_u(td);
1245 io_u->flags |= IO_U_F_FLIGHT;
1248 if (ddir_rw(acct_ddir(io_u)))
1249 td->io_issues[acct_ddir(io_u)]++;
1250 if (ddir_rw(io_u->ddir)) {
1251 io_u_mark_depth(td, 1);
1252 td->ts.total_io_u[io_u->ddir]++;
1255 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1257 td->o.verify != VERIFY_NONE &&
1258 !td->o.experimental_verify)
1259 log_io_piece(td, io_u);
1261 ret = io_u_sync_complete(td, io_u);
1265 return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1269 * Entry point for the thread based jobs. The process based jobs end up
1270 * here as well, after a little setup.
1272 static void *thread_main(void *data)
1274 unsigned long long elapsed;
1275 struct thread_data *td = data;
1276 struct thread_options *o = &td->o;
1277 pthread_condattr_t attr;
1281 if (!o->use_thread) {
1287 fio_local_clock_init(o->use_thread);
1289 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1292 fio_server_send_start(td);
1294 INIT_FLIST_HEAD(&td->io_log_list);
1295 INIT_FLIST_HEAD(&td->io_hist_list);
1296 INIT_FLIST_HEAD(&td->verify_list);
1297 INIT_FLIST_HEAD(&td->trim_list);
1298 INIT_FLIST_HEAD(&td->next_rand_list);
1299 pthread_mutex_init(&td->io_u_lock, NULL);
1300 td->io_hist_tree = RB_ROOT;
1302 pthread_condattr_init(&attr);
1303 pthread_cond_init(&td->verify_cond, &attr);
1304 pthread_cond_init(&td->free_cond, &attr);
1306 td_set_runstate(td, TD_INITIALIZED);
1307 dprint(FD_MUTEX, "up startup_mutex\n");
1308 fio_mutex_up(startup_mutex);
1309 dprint(FD_MUTEX, "wait on td->mutex\n");
1310 fio_mutex_down(td->mutex);
1311 dprint(FD_MUTEX, "done waiting on td->mutex\n");
1314 * A new gid requires privilege, so we need to do this before setting
1317 if (o->gid != -1U && setgid(o->gid)) {
1318 td_verror(td, errno, "setgid");
1321 if (o->uid != -1U && setuid(o->uid)) {
1322 td_verror(td, errno, "setuid");
1327 * If we have a gettimeofday() thread, make sure we exclude that
1328 * thread from this job
1331 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1334 * Set affinity first, in case it has an impact on the memory
1337 if (fio_option_is_set(o, cpumask)) {
1338 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1339 ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1341 log_err("fio: no CPUs set\n");
1342 log_err("fio: Try increasing number of available CPUs\n");
1343 td_verror(td, EINVAL, "cpus_split");
1347 ret = fio_setaffinity(td->pid, o->cpumask);
1349 td_verror(td, errno, "cpu_set_affinity");
1354 #ifdef CONFIG_LIBNUMA
1355 /* numa node setup */
1356 if (fio_option_is_set(o, numa_cpunodes) ||
1357 fio_option_is_set(o, numa_memnodes)) {
1358 struct bitmask *mask;
1360 if (numa_available() < 0) {
1361 td_verror(td, errno, "Does not support NUMA API\n");
1365 if (fio_option_is_set(o, numa_cpunodes)) {
1366 mask = numa_parse_nodestring(o->numa_cpunodes);
1367 ret = numa_run_on_node_mask(mask);
1368 numa_free_nodemask(mask);
1370 td_verror(td, errno, \
1371 "numa_run_on_node_mask failed\n");
1376 if (fio_option_is_set(o, numa_memnodes)) {
1378 if (o->numa_memnodes)
1379 mask = numa_parse_nodestring(o->numa_memnodes);
1381 switch (o->numa_mem_mode) {
1382 case MPOL_INTERLEAVE:
1383 numa_set_interleave_mask(mask);
1386 numa_set_membind(mask);
1389 numa_set_localalloc();
1391 case MPOL_PREFERRED:
1392 numa_set_preferred(o->numa_mem_prefer_node);
1400 numa_free_nodemask(mask);
1406 if (fio_pin_memory(td))
1410 * May alter parameters that init_io_u() will use, so we need to
1419 if (o->verify_async && verify_async_init(td))
1422 if (fio_option_is_set(o, ioprio) ||
1423 fio_option_is_set(o, ioprio_class)) {
1424 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1426 td_verror(td, errno, "ioprio_set");
1431 if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1435 if (nice(o->nice) == -1 && errno != 0) {
1436 td_verror(td, errno, "nice");
1440 if (o->ioscheduler && switch_ioscheduler(td))
1443 if (!o->create_serialize && setup_files(td))
1449 if (init_random_map(td))
1452 if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
1456 if (pre_read_files(td) < 0)
1460 if (td->flags & TD_F_COMPRESS_LOG)
1461 tp_init(&td->tp_data);
1463 fio_verify_init(td);
1465 fio_gettime(&td->epoch, NULL);
1466 fio_getrusage(&td->ru_start);
1468 while (keep_running(td)) {
1469 uint64_t verify_bytes;
1471 fio_gettime(&td->start, NULL);
1472 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1473 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1474 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1476 if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1477 o->ratemin[DDIR_TRIM]) {
1478 memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1479 sizeof(td->bw_sample_time));
1480 memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1481 sizeof(td->bw_sample_time));
1482 memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1483 sizeof(td->bw_sample_time));
1489 prune_io_piece_log(td);
1491 if (td->o.verify_only && (td_write(td) || td_rw(td)))
1492 verify_bytes = do_dry_run(td);
1494 verify_bytes = do_io(td);
1499 * Make sure we've successfully updated the rusage stats
1500 * before waiting on the stat mutex. Otherwise we could have
1501 * the stat thread holding stat mutex and waiting for
1502 * the rusage_sem, which would never get upped because
1503 * this thread is waiting for the stat mutex.
1505 check_update_rusage(td);
1507 fio_mutex_down(stat_mutex);
1508 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1509 elapsed = mtime_since_now(&td->start);
1510 td->ts.runtime[DDIR_READ] += elapsed;
1512 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1513 elapsed = mtime_since_now(&td->start);
1514 td->ts.runtime[DDIR_WRITE] += elapsed;
1516 if (td_trim(td) && td->io_bytes[DDIR_TRIM]) {
1517 elapsed = mtime_since_now(&td->start);
1518 td->ts.runtime[DDIR_TRIM] += elapsed;
1520 fio_gettime(&td->start, NULL);
1521 fio_mutex_up(stat_mutex);
1523 if (td->error || td->terminate)
1526 if (!o->do_verify ||
1527 o->verify == VERIFY_NONE ||
1528 (td->io_ops->flags & FIO_UNIDIR))
1533 fio_gettime(&td->start, NULL);
1535 do_verify(td, verify_bytes);
1538 * See comment further up for why this is done here.
1540 check_update_rusage(td);
1542 fio_mutex_down(stat_mutex);
1543 td->ts.runtime[DDIR_READ] += mtime_since_now(&td->start);
1544 fio_gettime(&td->start, NULL);
1545 fio_mutex_up(stat_mutex);
1547 if (td->error || td->terminate)
1551 update_rusage_stat(td);
1552 td->ts.total_run_time = mtime_since_now(&td->epoch);
1553 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1554 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1555 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1557 if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
1558 (td->o.verify != VERIFY_NONE && td_write(td))) {
1559 struct all_io_list *state;
1562 state = get_all_io_list(td->thread_number, &sz);
1564 __verify_save_state(state, "local");
1569 fio_unpin_memory(td);
1571 fio_writeout_logs(td);
1573 if (td->flags & TD_F_COMPRESS_LOG)
1574 tp_exit(&td->tp_data);
1576 if (o->exec_postrun)
1577 exec_string(o, o->exec_postrun, (const char *)"postrun");
1579 if (exitall_on_terminate)
1580 fio_terminate_threads(td->groupid);
1584 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1587 if (o->verify_async)
1588 verify_async_exit(td);
1590 close_and_free_files(td);
1593 cgroup_shutdown(td, &cgroup_mnt);
1594 verify_free_state(td);
1596 if (fio_option_is_set(o, cpumask)) {
1597 ret = fio_cpuset_exit(&o->cpumask);
1599 td_verror(td, ret, "fio_cpuset_exit");
1603 * do this very late, it will log file closing as well
1605 if (o->write_iolog_file)
1606 write_iolog_close(td);
1608 fio_mutex_remove(td->mutex);
1611 td_set_runstate(td, TD_EXITED);
1614 * Do this last after setting our runstate to exited, so we
1615 * know that the stat thread is signaled.
1617 check_update_rusage(td);
1619 return (void *) (uintptr_t) td->error;
1624 * We cannot pass the td data into a forked process, so attach the td and
1625 * pass it to the thread worker.
1627 static int fork_main(int shmid, int offset)
1629 struct thread_data *td;
1632 #if !defined(__hpux) && !defined(CONFIG_NO_SHM)
1633 data = shmat(shmid, NULL, 0);
1634 if (data == (void *) -1) {
1642 * HP-UX inherits shm mappings?
1647 td = data + offset * sizeof(struct thread_data);
1648 ret = thread_main(td);
1650 return (int) (uintptr_t) ret;
1653 static void dump_td_info(struct thread_data *td)
1655 log_err("fio: job '%s' hasn't exited in %lu seconds, it appears to "
1656 "be stuck. Doing forceful exit of this job.\n", td->o.name,
1657 (unsigned long) time_since_now(&td->terminate_time));
1661 * Run over the job map and reap the threads that have exited, if any.
1663 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1664 unsigned int *m_rate)
1666 struct thread_data *td;
1667 unsigned int cputhreads, realthreads, pending;
1671 * reap exited threads (TD_EXITED -> TD_REAPED)
1673 realthreads = pending = cputhreads = 0;
1674 for_each_td(td, i) {
1678 * ->io_ops is NULL for a thread that has closed its
1681 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1690 if (td->runstate == TD_REAPED)
1692 if (td->o.use_thread) {
1693 if (td->runstate == TD_EXITED) {
1694 td_set_runstate(td, TD_REAPED);
1701 if (td->runstate == TD_EXITED)
1705 * check if someone quit or got killed in an unusual way
1707 ret = waitpid(td->pid, &status, flags);
1709 if (errno == ECHILD) {
1710 log_err("fio: pid=%d disappeared %d\n",
1711 (int) td->pid, td->runstate);
1713 td_set_runstate(td, TD_REAPED);
1717 } else if (ret == td->pid) {
1718 if (WIFSIGNALED(status)) {
1719 int sig = WTERMSIG(status);
1721 if (sig != SIGTERM && sig != SIGUSR2)
1722 log_err("fio: pid=%d, got signal=%d\n",
1723 (int) td->pid, sig);
1725 td_set_runstate(td, TD_REAPED);
1728 if (WIFEXITED(status)) {
1729 if (WEXITSTATUS(status) && !td->error)
1730 td->error = WEXITSTATUS(status);
1732 td_set_runstate(td, TD_REAPED);
1738 * If the job is stuck, do a forceful timeout of it and
1741 if (td->terminate &&
1742 time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
1744 td_set_runstate(td, TD_REAPED);
1749 * thread is not dead, continue
1755 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1756 (*t_rate) -= ddir_rw_sum(td->o.rate);
1763 done_secs += mtime_since_now(&td->epoch) / 1000;
1764 profile_td_exit(td);
1767 if (*nr_running == cputhreads && !pending && realthreads)
1768 fio_terminate_threads(TERMINATE_ALL);
1771 static int __check_trigger_file(void)
1778 if (stat(trigger_file, &sb))
1781 if (unlink(trigger_file) < 0)
1782 log_err("fio: failed to unlink %s: %s\n", trigger_file,
1788 static int trigger_timedout(void)
1790 if (trigger_timeout)
1791 return time_since_genesis() >= trigger_timeout;
1796 void exec_trigger(const char *cmd)
1805 log_err("fio: failed executing %s trigger\n", cmd);
1808 void check_trigger_file(void)
1810 if (__check_trigger_file() || trigger_timedout()) {
1812 fio_clients_send_trigger(trigger_remote_cmd);
1814 verify_save_state();
1815 fio_terminate_threads(TERMINATE_ALL);
1816 exec_trigger(trigger_cmd);
1821 static int fio_verify_load_state(struct thread_data *td)
1825 if (!td->o.verify_state)
1831 ret = fio_server_get_verify_state(td->o.name,
1832 td->thread_number - 1, &data);
1834 verify_convert_assign_state(td, data);
1836 ret = verify_load_state(td, "local");
1841 static void do_usleep(unsigned int usecs)
1843 check_for_running_stats();
1844 check_trigger_file();
1849 * Main function for kicking off and reaping jobs, as needed.
1851 static void run_threads(void)
1853 struct thread_data *td;
1854 unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
1857 if (fio_gtod_offload && fio_start_gtod_thread())
1860 fio_idle_prof_init();
1864 nr_thread = nr_process = 0;
1865 for_each_td(td, i) {
1866 if (td->o.use_thread)
1872 if (output_format == FIO_OUTPUT_NORMAL) {
1873 log_info("Starting ");
1875 log_info("%d thread%s", nr_thread,
1876 nr_thread > 1 ? "s" : "");
1880 log_info("%d process%s", nr_process,
1881 nr_process > 1 ? "es" : "");
1887 todo = thread_number;
1890 m_rate = t_rate = 0;
1892 for_each_td(td, i) {
1893 print_status_init(td->thread_number - 1);
1895 if (!td->o.create_serialize)
1898 if (fio_verify_load_state(td))
1902 * do file setup here so it happens sequentially,
1903 * we don't want X number of threads getting their
1904 * client data interspersed on disk
1906 if (setup_files(td)) {
1910 log_err("fio: pid=%d, err=%d/%s\n",
1911 (int) td->pid, td->error, td->verror);
1912 td_set_runstate(td, TD_REAPED);
1919 * for sharing to work, each job must always open
1920 * its own files. so close them, if we opened them
1923 for_each_file(td, f, j) {
1924 if (fio_file_open(f))
1925 td_io_close_file(td, f);
1930 /* start idle threads before io threads start to run */
1931 fio_idle_prof_start();
1936 struct thread_data *map[REAL_MAX_JOBS];
1937 struct timeval this_start;
1938 int this_jobs = 0, left;
1941 * create threads (TD_NOT_CREATED -> TD_CREATED)
1943 for_each_td(td, i) {
1944 if (td->runstate != TD_NOT_CREATED)
1948 * never got a chance to start, killed by other
1949 * thread for some reason
1951 if (td->terminate) {
1956 if (td->o.start_delay) {
1957 spent = utime_since_genesis();
1959 if (td->o.start_delay > spent)
1963 if (td->o.stonewall && (nr_started || nr_running)) {
1964 dprint(FD_PROCESS, "%s: stonewall wait\n",
1971 td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED);
1972 td->update_rusage = 0;
1975 * Set state to created. Thread will transition
1976 * to TD_INITIALIZED when it's done setting up.
1978 td_set_runstate(td, TD_CREATED);
1979 map[this_jobs++] = td;
1982 if (td->o.use_thread) {
1985 dprint(FD_PROCESS, "will pthread_create\n");
1986 ret = pthread_create(&td->thread, NULL,
1989 log_err("pthread_create: %s\n",
1994 ret = pthread_detach(td->thread);
1996 log_err("pthread_detach: %s",
2000 dprint(FD_PROCESS, "will fork\n");
2003 int ret = fork_main(shm_id, i);
2006 } else if (i == fio_debug_jobno)
2007 *fio_debug_jobp = pid;
2009 dprint(FD_MUTEX, "wait on startup_mutex\n");
2010 if (fio_mutex_down_timeout(startup_mutex, 10)) {
2011 log_err("fio: job startup hung? exiting.\n");
2012 fio_terminate_threads(TERMINATE_ALL);
2017 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2021 * Wait for the started threads to transition to
2024 fio_gettime(&this_start, NULL);
2026 while (left && !fio_abort) {
2027 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2032 for (i = 0; i < this_jobs; i++) {
2036 if (td->runstate == TD_INITIALIZED) {
2039 } else if (td->runstate >= TD_EXITED) {
2043 nr_running++; /* work-around... */
2049 log_err("fio: %d job%s failed to start\n", left,
2050 left > 1 ? "s" : "");
2051 for (i = 0; i < this_jobs; i++) {
2055 kill(td->pid, SIGTERM);
2061 * start created threads (TD_INITIALIZED -> TD_RUNNING).
2063 for_each_td(td, i) {
2064 if (td->runstate != TD_INITIALIZED)
2067 if (in_ramp_time(td))
2068 td_set_runstate(td, TD_RAMP);
2070 td_set_runstate(td, TD_RUNNING);
2073 m_rate += ddir_rw_sum(td->o.ratemin);
2074 t_rate += ddir_rw_sum(td->o.rate);
2076 fio_mutex_up(td->mutex);
2079 reap_threads(&nr_running, &t_rate, &m_rate);
2085 while (nr_running) {
2086 reap_threads(&nr_running, &t_rate, &m_rate);
2090 fio_idle_prof_stop();
2095 static void wait_for_helper_thread_exit(void)
2100 pthread_cond_signal(&helper_cond);
2101 pthread_join(helper_thread, &ret);
2104 static void free_disk_util(void)
2106 disk_util_prune_entries();
2108 pthread_cond_destroy(&helper_cond);
2111 static void *helper_thread_main(void *data)
2115 fio_mutex_up(startup_mutex);
2118 uint64_t sec = DISK_UTIL_MSEC / 1000;
2119 uint64_t nsec = (DISK_UTIL_MSEC % 1000) * 1000000;
2123 gettimeofday(&tv, NULL);
2124 ts.tv_sec = tv.tv_sec + sec;
2125 ts.tv_nsec = (tv.tv_usec * 1000) + nsec;
2127 if (ts.tv_nsec >= 1000000000ULL) {
2128 ts.tv_nsec -= 1000000000ULL;
2132 pthread_cond_timedwait(&helper_cond, &helper_lock, &ts);
2134 ret = update_io_ticks();
2136 if (helper_do_stat) {
2138 __show_running_run_stats();
2142 print_thread_status();
2148 static int create_helper_thread(void)
2154 pthread_cond_init(&helper_cond, NULL);
2155 pthread_mutex_init(&helper_lock, NULL);
2157 ret = pthread_create(&helper_thread, NULL, helper_thread_main, NULL);
2159 log_err("Can't create helper thread: %s\n", strerror(ret));
2163 dprint(FD_MUTEX, "wait on startup_mutex\n");
2164 fio_mutex_down(startup_mutex);
2165 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2169 int fio_backend(void)
2171 struct thread_data *td;
2175 if (load_profile(exec_profile))
2178 exec_profile = NULL;
2184 struct log_params p = {
2185 .log_type = IO_LOG_TYPE_BW,
2188 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2189 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2190 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2193 startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
2194 if (startup_mutex == NULL)
2199 create_helper_thread();
2201 cgroup_list = smalloc(sizeof(*cgroup_list));
2202 INIT_FLIST_HEAD(cgroup_list);
2206 wait_for_helper_thread_exit();
2211 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2212 struct io_log *log = agg_io_log[i];
2220 for_each_td(td, i) {
2221 fio_options_free(td);
2222 if (td->rusage_sem) {
2223 fio_mutex_remove(td->rusage_sem);
2224 td->rusage_sem = NULL;
2229 cgroup_kill(cgroup_list);
2233 fio_mutex_remove(startup_mutex);