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,
233 uint64_t *bytes_done)
237 if (bytes_done[DDIR_READ])
238 ret |= __check_min_rate(td, now, DDIR_READ);
239 if (bytes_done[DDIR_WRITE])
240 ret |= __check_min_rate(td, now, DDIR_WRITE);
241 if (bytes_done[DDIR_TRIM])
242 ret |= __check_min_rate(td, now, DDIR_TRIM);
248 * When job exits, we can cancel the in-flight IO if we are using async
249 * io. Attempt to do so.
251 static void cleanup_pending_aio(struct thread_data *td)
256 * get immediately available events, if any
258 r = io_u_queued_complete(td, 0, NULL);
263 * now cancel remaining active events
265 if (td->io_ops->cancel) {
269 io_u_qiter(&td->io_u_all, io_u, i) {
270 if (io_u->flags & IO_U_F_FLIGHT) {
271 r = td->io_ops->cancel(td, io_u);
279 r = io_u_queued_complete(td, td->cur_depth, NULL);
283 * Helper to handle the final sync of a file. Works just like the normal
284 * io path, just does everything sync.
286 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
288 struct io_u *io_u = __get_io_u(td);
294 io_u->ddir = DDIR_SYNC;
297 if (td_io_prep(td, io_u)) {
303 ret = td_io_queue(td, io_u);
305 td_verror(td, io_u->error, "td_io_queue");
308 } else if (ret == FIO_Q_QUEUED) {
309 if (io_u_queued_complete(td, 1, NULL) < 0)
311 } else if (ret == FIO_Q_COMPLETED) {
313 td_verror(td, io_u->error, "td_io_queue");
317 if (io_u_sync_complete(td, io_u, NULL) < 0)
319 } else if (ret == FIO_Q_BUSY) {
320 if (td_io_commit(td))
328 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
332 if (fio_file_open(f))
333 return fio_io_sync(td, f);
335 if (td_io_open_file(td, f))
338 ret = fio_io_sync(td, f);
339 td_io_close_file(td, f);
343 static inline void __update_tv_cache(struct thread_data *td)
345 fio_gettime(&td->tv_cache, NULL);
348 static inline void update_tv_cache(struct thread_data *td)
350 if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
351 __update_tv_cache(td);
354 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
356 if (in_ramp_time(td))
360 if (utime_since(&td->epoch, t) >= td->o.timeout)
366 static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
371 if (ret < 0 || td->error) {
373 enum error_type_bit eb;
378 eb = td_error_type(ddir, err);
379 if (!(td->o.continue_on_error & (1 << eb)))
382 if (td_non_fatal_error(td, eb, err)) {
384 * Continue with the I/Os in case of
387 update_error_count(td, err);
391 } else if (td->o.fill_device && err == ENOSPC) {
393 * We expect to hit this error if
394 * fill_device option is set.
397 fio_mark_td_terminate(td);
401 * Stop the I/O in case of a fatal
404 update_error_count(td, err);
412 static void check_update_rusage(struct thread_data *td)
414 if (td->update_rusage) {
415 td->update_rusage = 0;
416 update_rusage_stat(td);
417 fio_mutex_up(td->rusage_sem);
422 * The main verify engine. Runs over the writes we previously submitted,
423 * reads the blocks back in, and checks the crc/md5 of the data.
425 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
427 uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
433 dprint(FD_VERIFY, "starting loop\n");
436 * sync io first and invalidate cache, to make sure we really
439 for_each_file(td, f, i) {
440 if (!fio_file_open(f))
442 if (fio_io_sync(td, f))
444 if (file_invalidate_cache(td, f))
448 check_update_rusage(td);
453 td_set_runstate(td, TD_VERIFYING);
456 while (!td->terminate) {
461 check_update_rusage(td);
463 if (runtime_exceeded(td, &td->tv_cache)) {
464 __update_tv_cache(td);
465 if (runtime_exceeded(td, &td->tv_cache)) {
466 fio_mark_td_terminate(td);
471 if (flow_threshold_exceeded(td))
474 if (!td->o.experimental_verify) {
475 io_u = __get_io_u(td);
479 if (get_next_verify(td, io_u)) {
484 if (td_io_prep(td, io_u)) {
489 if (ddir_rw_sum(bytes_done) + td->o.rw_min_bs > verify_bytes)
492 while ((io_u = get_io_u(td)) != NULL) {
500 * We are only interested in the places where
501 * we wrote or trimmed IOs. Turn those into
502 * reads for verification purposes.
504 if (io_u->ddir == DDIR_READ) {
506 * Pretend we issued it for rwmix
509 td->io_issues[DDIR_READ]++;
512 } else if (io_u->ddir == DDIR_TRIM) {
513 io_u->ddir = DDIR_READ;
514 io_u->flags |= IO_U_F_TRIMMED;
516 } else if (io_u->ddir == DDIR_WRITE) {
517 io_u->ddir = DDIR_READ;
529 if (verify_state_should_stop(td, io_u)) {
534 if (td->o.verify_async)
535 io_u->end_io = verify_io_u_async;
537 io_u->end_io = verify_io_u;
541 ret = td_io_queue(td, io_u);
543 case FIO_Q_COMPLETED:
546 clear_io_u(td, io_u);
547 } else if (io_u->resid) {
548 int bytes = io_u->xfer_buflen - io_u->resid;
554 td_verror(td, EIO, "full resid");
559 io_u->xfer_buflen = io_u->resid;
560 io_u->xfer_buf += bytes;
561 io_u->offset += bytes;
563 if (ddir_rw(io_u->ddir))
564 td->ts.short_io_u[io_u->ddir]++;
567 if (io_u->offset == f->real_file_size)
570 requeue_io_u(td, &io_u);
573 ret = io_u_sync_complete(td, io_u, bytes_done);
581 requeue_io_u(td, &io_u);
582 ret2 = td_io_commit(td);
588 td_verror(td, -ret, "td_io_queue");
592 if (break_on_this_error(td, ddir, &ret))
596 * if we can queue more, do so. but check if there are
597 * completed io_u's first. Note that we can get BUSY even
598 * without IO queued, if the system is resource starved.
601 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
602 if (full || !td->o.iodepth_batch_complete) {
603 min_events = min(td->o.iodepth_batch_complete,
606 * if the queue is full, we MUST reap at least 1 event
608 if (full && !min_events)
613 * Reap required number of io units, if any,
614 * and do the verification on them through
615 * the callback handler
617 if (io_u_queued_complete(td, min_events, bytes_done) < 0) {
621 } while (full && (td->cur_depth > td->o.iodepth_low));
627 check_update_rusage(td);
630 min_events = td->cur_depth;
633 ret = io_u_queued_complete(td, min_events, NULL);
635 cleanup_pending_aio(td);
637 td_set_runstate(td, TD_RUNNING);
639 dprint(FD_VERIFY, "exiting loop\n");
642 static unsigned int exceeds_number_ios(struct thread_data *td)
644 unsigned long long number_ios;
646 if (!td->o.number_ios)
649 number_ios = ddir_rw_sum(td->this_io_blocks);
650 number_ios += td->io_u_queued + td->io_u_in_flight;
652 return number_ios >= td->o.number_ios;
655 static int io_bytes_exceeded(struct thread_data *td)
657 unsigned long long bytes, limit;
660 bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE];
661 else if (td_write(td))
662 bytes = td->this_io_bytes[DDIR_WRITE];
663 else if (td_read(td))
664 bytes = td->this_io_bytes[DDIR_READ];
666 bytes = td->this_io_bytes[DDIR_TRIM];
669 limit = td->o.io_limit;
673 return bytes >= limit || exceeds_number_ios(td);
677 * Main IO worker function. It retrieves io_u's to process and queues
678 * and reaps them, checking for rate and errors along the way.
680 * Returns number of bytes written and trimmed.
682 static uint64_t do_io(struct thread_data *td)
684 uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
687 uint64_t total_bytes, bytes_issued = 0;
689 if (in_ramp_time(td))
690 td_set_runstate(td, TD_RAMP);
692 td_set_runstate(td, TD_RUNNING);
697 * If verify_backlog is enabled, we'll run the verify in this
698 * handler as well. For that case, we may need up to twice the
701 total_bytes = td->o.size;
702 if (td->o.verify != VERIFY_NONE &&
703 (td_write(td) && td->o.verify_backlog))
704 total_bytes += td->o.size;
706 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
707 (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td) ||
709 struct timeval comp_time;
715 check_update_rusage(td);
717 if (td->terminate || td->done)
722 if (runtime_exceeded(td, &td->tv_cache)) {
723 __update_tv_cache(td);
724 if (runtime_exceeded(td, &td->tv_cache)) {
725 fio_mark_td_terminate(td);
730 if (flow_threshold_exceeded(td))
733 if (bytes_issued >= total_bytes)
737 if (IS_ERR_OR_NULL(io_u)) {
738 int err = PTR_ERR(io_u);
745 if (td->o.latency_target)
753 * Add verification end_io handler if:
754 * - Asked to verify (!td_rw(td))
755 * - Or the io_u is from our verify list (mixed write/ver)
757 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
758 ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
760 if (!td->o.verify_pattern_bytes) {
761 io_u->rand_seed = __rand(&td->verify_state);
762 if (sizeof(int) != sizeof(long *))
763 io_u->rand_seed *= __rand(&td->verify_state);
766 if (verify_state_should_stop(td, io_u)) {
771 if (td->o.verify_async)
772 io_u->end_io = verify_io_u_async;
774 io_u->end_io = verify_io_u;
775 td_set_runstate(td, TD_VERIFYING);
776 } else if (in_ramp_time(td))
777 td_set_runstate(td, TD_RAMP);
779 td_set_runstate(td, TD_RUNNING);
782 * Always log IO before it's issued, so we know the specific
783 * order of it. The logged unit will track when the IO has
786 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
788 td->o.verify != VERIFY_NONE &&
789 !td->o.experimental_verify)
790 log_io_piece(td, io_u);
792 ret = td_io_queue(td, io_u);
794 case FIO_Q_COMPLETED:
797 unlog_io_piece(td, io_u);
798 clear_io_u(td, io_u);
799 } else if (io_u->resid) {
800 int bytes = io_u->xfer_buflen - io_u->resid;
801 struct fio_file *f = io_u->file;
803 bytes_issued += bytes;
805 trim_io_piece(td, io_u);
811 unlog_io_piece(td, io_u);
812 td_verror(td, EIO, "full resid");
817 io_u->xfer_buflen = io_u->resid;
818 io_u->xfer_buf += bytes;
819 io_u->offset += bytes;
821 if (ddir_rw(io_u->ddir))
822 td->ts.short_io_u[io_u->ddir]++;
824 if (io_u->offset == f->real_file_size)
827 requeue_io_u(td, &io_u);
830 if (__should_check_rate(td, DDIR_READ) ||
831 __should_check_rate(td, DDIR_WRITE) ||
832 __should_check_rate(td, DDIR_TRIM))
833 fio_gettime(&comp_time, NULL);
835 ret = io_u_sync_complete(td, io_u, bytes_done);
838 bytes_issued += io_u->xfer_buflen;
843 * if the engine doesn't have a commit hook,
844 * the io_u is really queued. if it does have such
845 * a hook, it has to call io_u_queued() itself.
847 if (td->io_ops->commit == NULL)
848 io_u_queued(td, io_u);
849 bytes_issued += io_u->xfer_buflen;
852 unlog_io_piece(td, io_u);
853 requeue_io_u(td, &io_u);
854 ret2 = td_io_commit(td);
864 if (break_on_this_error(td, ddir, &ret))
868 * See if we need to complete some commands. Note that we
869 * can get BUSY even without IO queued, if the system is
873 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
874 if (full || !td->o.iodepth_batch_complete) {
875 min_evts = min(td->o.iodepth_batch_complete,
878 * if the queue is full, we MUST reap at least 1 event
880 if (full && !min_evts)
883 if (__should_check_rate(td, DDIR_READ) ||
884 __should_check_rate(td, DDIR_WRITE) ||
885 __should_check_rate(td, DDIR_TRIM))
886 fio_gettime(&comp_time, NULL);
889 ret = io_u_queued_complete(td, min_evts, bytes_done);
893 } while (full && (td->cur_depth > td->o.iodepth_low));
898 if (!ddir_rw_sum(bytes_done) && !(td->io_ops->flags & FIO_NOIO))
901 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
902 if (check_min_rate(td, &comp_time, bytes_done)) {
903 if (exitall_on_terminate)
904 fio_terminate_threads(td->groupid);
905 td_verror(td, EIO, "check_min_rate");
909 if (!in_ramp_time(td) && td->o.latency_target)
910 lat_target_check(td);
912 if (td->o.thinktime) {
913 unsigned long long b;
915 b = ddir_rw_sum(td->io_blocks);
916 if (!(b % td->o.thinktime_blocks)) {
921 if (td->o.thinktime_spin)
922 usec_spin(td->o.thinktime_spin);
924 left = td->o.thinktime - td->o.thinktime_spin;
926 usec_sleep(td, left);
931 check_update_rusage(td);
933 if (td->trim_entries)
934 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
936 if (td->o.fill_device && td->error == ENOSPC) {
938 fio_mark_td_terminate(td);
945 ret = io_u_queued_complete(td, i, bytes_done);
946 if (td->o.fill_device && td->error == ENOSPC)
950 if (should_fsync(td) && td->o.end_fsync) {
951 td_set_runstate(td, TD_FSYNCING);
953 for_each_file(td, f, i) {
954 if (!fio_file_fsync(td, f))
957 log_err("fio: end_fsync failed for file %s\n",
962 cleanup_pending_aio(td);
965 * stop job if we failed doing any IO
967 if (!ddir_rw_sum(td->this_io_bytes))
970 return bytes_done[DDIR_WRITE] + bytes_done[DDIR_TRIM];
973 static void cleanup_io_u(struct thread_data *td)
977 while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
979 if (td->io_ops->io_u_free)
980 td->io_ops->io_u_free(td, io_u);
982 fio_memfree(io_u, sizeof(*io_u));
987 io_u_rexit(&td->io_u_requeues);
988 io_u_qexit(&td->io_u_freelist);
989 io_u_qexit(&td->io_u_all);
991 if (td->last_write_comp)
992 sfree(td->last_write_comp);
995 static int init_io_u(struct thread_data *td)
998 unsigned int max_bs, min_write;
999 int cl_align, i, max_units;
1000 int data_xfer = 1, err;
1003 max_units = td->o.iodepth;
1004 max_bs = td_max_bs(td);
1005 min_write = td->o.min_bs[DDIR_WRITE];
1006 td->orig_buffer_size = (unsigned long long) max_bs
1007 * (unsigned long long) max_units;
1009 if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
1013 err += io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1014 err += io_u_qinit(&td->io_u_freelist, td->o.iodepth);
1015 err += io_u_qinit(&td->io_u_all, td->o.iodepth);
1018 log_err("fio: failed setting up IO queues\n");
1023 * if we may later need to do address alignment, then add any
1024 * possible adjustment here so that we don't cause a buffer
1025 * overflow later. this adjustment may be too much if we get
1026 * lucky and the allocator gives us an aligned address.
1028 if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1029 (td->io_ops->flags & FIO_RAWIO))
1030 td->orig_buffer_size += page_mask + td->o.mem_align;
1032 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1035 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1036 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1039 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1040 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1044 if (data_xfer && allocate_io_mem(td))
1047 if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1048 (td->io_ops->flags & FIO_RAWIO))
1049 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
1051 p = td->orig_buffer;
1053 cl_align = os_cache_line_size();
1055 for (i = 0; i < max_units; i++) {
1061 ptr = fio_memalign(cl_align, sizeof(*io_u));
1063 log_err("fio: unable to allocate aligned memory\n");
1068 memset(io_u, 0, sizeof(*io_u));
1069 INIT_FLIST_HEAD(&io_u->verify_list);
1070 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1074 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1077 io_u_fill_buffer(td, io_u, min_write, max_bs);
1078 if (td_write(td) && td->o.verify_pattern_bytes) {
1080 * Fill the buffer with the pattern if we are
1081 * going to be doing writes.
1083 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1088 io_u->flags = IO_U_F_FREE;
1089 io_u_qpush(&td->io_u_freelist, io_u);
1092 * io_u never leaves this stack, used for iteration of all
1095 io_u_qpush(&td->io_u_all, io_u);
1097 if (td->io_ops->io_u_init) {
1098 int ret = td->io_ops->io_u_init(td, io_u);
1101 log_err("fio: failed to init engine data: %d\n", ret);
1109 if (td->o.verify != VERIFY_NONE) {
1110 td->last_write_comp = scalloc(max_units, sizeof(uint64_t));
1111 if (!td->last_write_comp) {
1112 log_err("fio: failed to alloc write comp data\n");
1120 static int switch_ioscheduler(struct thread_data *td)
1122 char tmp[256], tmp2[128];
1126 if (td->io_ops->flags & FIO_DISKLESSIO)
1129 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
1131 f = fopen(tmp, "r+");
1133 if (errno == ENOENT) {
1134 log_err("fio: os or kernel doesn't support IO scheduler"
1138 td_verror(td, errno, "fopen iosched");
1145 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1146 if (ferror(f) || ret != 1) {
1147 td_verror(td, errno, "fwrite");
1155 * Read back and check that the selected scheduler is now the default.
1157 ret = fread(tmp, sizeof(tmp), 1, f);
1158 if (ferror(f) || ret < 0) {
1159 td_verror(td, errno, "fread");
1163 tmp[sizeof(tmp) - 1] = '\0';
1166 sprintf(tmp2, "[%s]", td->o.ioscheduler);
1167 if (!strstr(tmp, tmp2)) {
1168 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1169 td_verror(td, EINVAL, "iosched_switch");
1178 static int keep_running(struct thread_data *td)
1180 unsigned long long limit;
1184 if (td->o.time_based)
1190 if (exceeds_number_ios(td))
1194 limit = td->o.io_limit;
1198 if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1202 * If the difference is less than the minimum IO size, we
1205 diff = limit - ddir_rw_sum(td->io_bytes);
1206 if (diff < td_max_bs(td))
1209 if (fio_files_done(td))
1218 static int exec_string(struct thread_options *o, const char *string, const char *mode)
1220 int ret, newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
1223 str = malloc(newlen);
1224 sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
1226 log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1229 log_err("fio: exec of cmd <%s> failed\n", str);
1236 * Dry run to compute correct state of numberio for verification.
1238 static uint64_t do_dry_run(struct thread_data *td)
1240 uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
1242 td_set_runstate(td, TD_RUNNING);
1244 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1245 (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td)) {
1249 if (td->terminate || td->done)
1252 io_u = get_io_u(td);
1256 io_u->flags |= IO_U_F_FLIGHT;
1259 if (ddir_rw(acct_ddir(io_u)))
1260 td->io_issues[acct_ddir(io_u)]++;
1261 if (ddir_rw(io_u->ddir)) {
1262 io_u_mark_depth(td, 1);
1263 td->ts.total_io_u[io_u->ddir]++;
1266 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1268 td->o.verify != VERIFY_NONE &&
1269 !td->o.experimental_verify)
1270 log_io_piece(td, io_u);
1272 ret = io_u_sync_complete(td, io_u, bytes_done);
1276 return bytes_done[DDIR_WRITE] + bytes_done[DDIR_TRIM];
1280 * Entry point for the thread based jobs. The process based jobs end up
1281 * here as well, after a little setup.
1283 static void *thread_main(void *data)
1285 unsigned long long elapsed;
1286 struct thread_data *td = data;
1287 struct thread_options *o = &td->o;
1288 pthread_condattr_t attr;
1292 if (!o->use_thread) {
1298 fio_local_clock_init(o->use_thread);
1300 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1303 fio_server_send_start(td);
1305 INIT_FLIST_HEAD(&td->io_log_list);
1306 INIT_FLIST_HEAD(&td->io_hist_list);
1307 INIT_FLIST_HEAD(&td->verify_list);
1308 INIT_FLIST_HEAD(&td->trim_list);
1309 INIT_FLIST_HEAD(&td->next_rand_list);
1310 pthread_mutex_init(&td->io_u_lock, NULL);
1311 td->io_hist_tree = RB_ROOT;
1313 pthread_condattr_init(&attr);
1314 pthread_cond_init(&td->verify_cond, &attr);
1315 pthread_cond_init(&td->free_cond, &attr);
1317 td_set_runstate(td, TD_INITIALIZED);
1318 dprint(FD_MUTEX, "up startup_mutex\n");
1319 fio_mutex_up(startup_mutex);
1320 dprint(FD_MUTEX, "wait on td->mutex\n");
1321 fio_mutex_down(td->mutex);
1322 dprint(FD_MUTEX, "done waiting on td->mutex\n");
1325 * A new gid requires privilege, so we need to do this before setting
1328 if (o->gid != -1U && setgid(o->gid)) {
1329 td_verror(td, errno, "setgid");
1332 if (o->uid != -1U && setuid(o->uid)) {
1333 td_verror(td, errno, "setuid");
1338 * If we have a gettimeofday() thread, make sure we exclude that
1339 * thread from this job
1342 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1345 * Set affinity first, in case it has an impact on the memory
1348 if (o->cpumask_set) {
1349 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1350 ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1352 log_err("fio: no CPUs set\n");
1353 log_err("fio: Try increasing number of available CPUs\n");
1354 td_verror(td, EINVAL, "cpus_split");
1358 ret = fio_setaffinity(td->pid, o->cpumask);
1360 td_verror(td, errno, "cpu_set_affinity");
1365 #ifdef CONFIG_LIBNUMA
1366 /* numa node setup */
1367 if (o->numa_cpumask_set || o->numa_memmask_set) {
1368 struct bitmask *mask;
1370 if (numa_available() < 0) {
1371 td_verror(td, errno, "Does not support NUMA API\n");
1375 if (o->numa_cpumask_set) {
1376 mask = numa_parse_nodestring(o->numa_cpunodes);
1377 ret = numa_run_on_node_mask(mask);
1378 numa_free_nodemask(mask);
1380 td_verror(td, errno, \
1381 "numa_run_on_node_mask failed\n");
1386 if (o->numa_memmask_set) {
1389 if (o->numa_memnodes)
1390 mask = numa_parse_nodestring(o->numa_memnodes);
1392 switch (o->numa_mem_mode) {
1393 case MPOL_INTERLEAVE:
1394 numa_set_interleave_mask(mask);
1397 numa_set_membind(mask);
1400 numa_set_localalloc();
1402 case MPOL_PREFERRED:
1403 numa_set_preferred(o->numa_mem_prefer_node);
1411 numa_free_nodemask(mask);
1417 if (fio_pin_memory(td))
1421 * May alter parameters that init_io_u() will use, so we need to
1430 if (o->verify_async && verify_async_init(td))
1434 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1436 td_verror(td, errno, "ioprio_set");
1441 if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1445 if (nice(o->nice) == -1 && errno != 0) {
1446 td_verror(td, errno, "nice");
1450 if (o->ioscheduler && switch_ioscheduler(td))
1453 if (!o->create_serialize && setup_files(td))
1459 if (init_random_map(td))
1462 if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
1466 if (pre_read_files(td) < 0)
1470 if (td->flags & TD_F_COMPRESS_LOG)
1471 tp_init(&td->tp_data);
1473 fio_verify_init(td);
1475 fio_gettime(&td->epoch, NULL);
1476 fio_getrusage(&td->ru_start);
1478 while (keep_running(td)) {
1479 uint64_t verify_bytes;
1481 fio_gettime(&td->start, NULL);
1482 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1483 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1484 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1486 if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1487 o->ratemin[DDIR_TRIM]) {
1488 memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1489 sizeof(td->bw_sample_time));
1490 memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1491 sizeof(td->bw_sample_time));
1492 memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1493 sizeof(td->bw_sample_time));
1499 prune_io_piece_log(td);
1501 if (td->o.verify_only && (td_write(td) || td_rw(td)))
1502 verify_bytes = do_dry_run(td);
1504 verify_bytes = do_io(td);
1508 fio_mutex_down(stat_mutex);
1509 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1510 elapsed = mtime_since_now(&td->start);
1511 td->ts.runtime[DDIR_READ] += elapsed;
1513 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1514 elapsed = mtime_since_now(&td->start);
1515 td->ts.runtime[DDIR_WRITE] += elapsed;
1517 if (td_trim(td) && td->io_bytes[DDIR_TRIM]) {
1518 elapsed = mtime_since_now(&td->start);
1519 td->ts.runtime[DDIR_TRIM] += elapsed;
1521 fio_gettime(&td->start, NULL);
1522 fio_mutex_up(stat_mutex);
1524 if (td->error || td->terminate)
1527 if (!o->do_verify ||
1528 o->verify == VERIFY_NONE ||
1529 (td->io_ops->flags & FIO_UNIDIR))
1534 fio_gettime(&td->start, NULL);
1536 do_verify(td, verify_bytes);
1538 fio_mutex_down(stat_mutex);
1539 td->ts.runtime[DDIR_READ] += mtime_since_now(&td->start);
1540 fio_gettime(&td->start, NULL);
1541 fio_mutex_up(stat_mutex);
1543 if (td->error || td->terminate)
1547 update_rusage_stat(td);
1548 td->ts.total_run_time = mtime_since_now(&td->epoch);
1549 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1550 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1551 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1553 if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
1554 (td->o.verify != VERIFY_NONE && td_write(td))) {
1555 struct all_io_list *state;
1558 state = get_all_io_list(td->thread_number, &sz);
1560 __verify_save_state(state, "local");
1565 fio_unpin_memory(td);
1567 fio_writeout_logs(td);
1569 if (td->flags & TD_F_COMPRESS_LOG)
1570 tp_exit(&td->tp_data);
1572 if (o->exec_postrun)
1573 exec_string(o, o->exec_postrun, (const char *)"postrun");
1575 if (exitall_on_terminate)
1576 fio_terminate_threads(td->groupid);
1580 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1583 if (o->verify_async)
1584 verify_async_exit(td);
1586 close_and_free_files(td);
1589 cgroup_shutdown(td, &cgroup_mnt);
1590 verify_free_state(td);
1592 if (o->cpumask_set) {
1593 ret = fio_cpuset_exit(&o->cpumask);
1595 td_verror(td, ret, "fio_cpuset_exit");
1599 * do this very late, it will log file closing as well
1601 if (o->write_iolog_file)
1602 write_iolog_close(td);
1604 fio_mutex_remove(td->mutex);
1607 td_set_runstate(td, TD_EXITED);
1610 * Do this last after setting our runstate to exited, so we
1611 * know that the stat thread is signaled.
1613 check_update_rusage(td);
1615 return (void *) (uintptr_t) td->error;
1620 * We cannot pass the td data into a forked process, so attach the td and
1621 * pass it to the thread worker.
1623 static int fork_main(int shmid, int offset)
1625 struct thread_data *td;
1628 #if !defined(__hpux) && !defined(CONFIG_NO_SHM)
1629 data = shmat(shmid, NULL, 0);
1630 if (data == (void *) -1) {
1638 * HP-UX inherits shm mappings?
1643 td = data + offset * sizeof(struct thread_data);
1644 ret = thread_main(td);
1646 return (int) (uintptr_t) ret;
1649 static void dump_td_info(struct thread_data *td)
1651 log_err("fio: job '%s' hasn't exited in %lu seconds, it appears to "
1652 "be stuck. Doing forceful exit of this job.\n", td->o.name,
1653 (unsigned long) time_since_now(&td->terminate_time));
1657 * Run over the job map and reap the threads that have exited, if any.
1659 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1660 unsigned int *m_rate)
1662 struct thread_data *td;
1663 unsigned int cputhreads, realthreads, pending;
1667 * reap exited threads (TD_EXITED -> TD_REAPED)
1669 realthreads = pending = cputhreads = 0;
1670 for_each_td(td, i) {
1674 * ->io_ops is NULL for a thread that has closed its
1677 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1686 if (td->runstate == TD_REAPED)
1688 if (td->o.use_thread) {
1689 if (td->runstate == TD_EXITED) {
1690 td_set_runstate(td, TD_REAPED);
1697 if (td->runstate == TD_EXITED)
1701 * check if someone quit or got killed in an unusual way
1703 ret = waitpid(td->pid, &status, flags);
1705 if (errno == ECHILD) {
1706 log_err("fio: pid=%d disappeared %d\n",
1707 (int) td->pid, td->runstate);
1709 td_set_runstate(td, TD_REAPED);
1713 } else if (ret == td->pid) {
1714 if (WIFSIGNALED(status)) {
1715 int sig = WTERMSIG(status);
1717 if (sig != SIGTERM && sig != SIGUSR2)
1718 log_err("fio: pid=%d, got signal=%d\n",
1719 (int) td->pid, sig);
1721 td_set_runstate(td, TD_REAPED);
1724 if (WIFEXITED(status)) {
1725 if (WEXITSTATUS(status) && !td->error)
1726 td->error = WEXITSTATUS(status);
1728 td_set_runstate(td, TD_REAPED);
1734 * If the job is stuck, do a forceful timeout of it and
1737 if (td->terminate &&
1738 time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
1740 td_set_runstate(td, TD_REAPED);
1745 * thread is not dead, continue
1751 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1752 (*t_rate) -= ddir_rw_sum(td->o.rate);
1759 done_secs += mtime_since_now(&td->epoch) / 1000;
1760 profile_td_exit(td);
1763 if (*nr_running == cputhreads && !pending && realthreads)
1764 fio_terminate_threads(TERMINATE_ALL);
1767 static int __check_trigger_file(void)
1774 if (stat(trigger_file, &sb))
1777 if (unlink(trigger_file) < 0)
1778 log_err("fio: failed to unlink %s: %s\n", trigger_file,
1784 static int trigger_timedout(void)
1786 if (trigger_timeout)
1787 return time_since_genesis() >= trigger_timeout;
1792 void exec_trigger(const char *cmd)
1801 log_err("fio: failed executing %s trigger\n", cmd);
1804 void check_trigger_file(void)
1806 if (__check_trigger_file() || trigger_timedout()) {
1808 fio_clients_send_trigger(trigger_remote_cmd);
1810 verify_save_state();
1811 fio_terminate_threads(TERMINATE_ALL);
1812 exec_trigger(trigger_cmd);
1817 static int fio_verify_load_state(struct thread_data *td)
1821 if (!td->o.verify_state)
1827 ret = fio_server_get_verify_state(td->o.name,
1828 td->thread_number - 1, &data);
1830 verify_convert_assign_state(td, data);
1832 ret = verify_load_state(td, "local");
1837 static void do_usleep(unsigned int usecs)
1839 check_for_running_stats();
1840 check_trigger_file();
1845 * Main function for kicking off and reaping jobs, as needed.
1847 static void run_threads(void)
1849 struct thread_data *td;
1850 unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
1853 if (fio_gtod_offload && fio_start_gtod_thread())
1856 fio_idle_prof_init();
1860 nr_thread = nr_process = 0;
1861 for_each_td(td, i) {
1862 if (td->o.use_thread)
1868 if (output_format == FIO_OUTPUT_NORMAL) {
1869 log_info("Starting ");
1871 log_info("%d thread%s", nr_thread,
1872 nr_thread > 1 ? "s" : "");
1876 log_info("%d process%s", nr_process,
1877 nr_process > 1 ? "es" : "");
1883 todo = thread_number;
1886 m_rate = t_rate = 0;
1888 for_each_td(td, i) {
1889 print_status_init(td->thread_number - 1);
1891 if (!td->o.create_serialize)
1894 if (fio_verify_load_state(td))
1898 * do file setup here so it happens sequentially,
1899 * we don't want X number of threads getting their
1900 * client data interspersed on disk
1902 if (setup_files(td)) {
1906 log_err("fio: pid=%d, err=%d/%s\n",
1907 (int) td->pid, td->error, td->verror);
1908 td_set_runstate(td, TD_REAPED);
1915 * for sharing to work, each job must always open
1916 * its own files. so close them, if we opened them
1919 for_each_file(td, f, j) {
1920 if (fio_file_open(f))
1921 td_io_close_file(td, f);
1926 /* start idle threads before io threads start to run */
1927 fio_idle_prof_start();
1932 struct thread_data *map[REAL_MAX_JOBS];
1933 struct timeval this_start;
1934 int this_jobs = 0, left;
1937 * create threads (TD_NOT_CREATED -> TD_CREATED)
1939 for_each_td(td, i) {
1940 if (td->runstate != TD_NOT_CREATED)
1944 * never got a chance to start, killed by other
1945 * thread for some reason
1947 if (td->terminate) {
1952 if (td->o.start_delay) {
1953 spent = utime_since_genesis();
1955 if (td->o.start_delay > spent)
1959 if (td->o.stonewall && (nr_started || nr_running)) {
1960 dprint(FD_PROCESS, "%s: stonewall wait\n",
1967 td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED);
1968 td->update_rusage = 0;
1971 * Set state to created. Thread will transition
1972 * to TD_INITIALIZED when it's done setting up.
1974 td_set_runstate(td, TD_CREATED);
1975 map[this_jobs++] = td;
1978 if (td->o.use_thread) {
1981 dprint(FD_PROCESS, "will pthread_create\n");
1982 ret = pthread_create(&td->thread, NULL,
1985 log_err("pthread_create: %s\n",
1990 ret = pthread_detach(td->thread);
1992 log_err("pthread_detach: %s",
1996 dprint(FD_PROCESS, "will fork\n");
1999 int ret = fork_main(shm_id, i);
2002 } else if (i == fio_debug_jobno)
2003 *fio_debug_jobp = pid;
2005 dprint(FD_MUTEX, "wait on startup_mutex\n");
2006 if (fio_mutex_down_timeout(startup_mutex, 10)) {
2007 log_err("fio: job startup hung? exiting.\n");
2008 fio_terminate_threads(TERMINATE_ALL);
2013 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2017 * Wait for the started threads to transition to
2020 fio_gettime(&this_start, NULL);
2022 while (left && !fio_abort) {
2023 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2028 for (i = 0; i < this_jobs; i++) {
2032 if (td->runstate == TD_INITIALIZED) {
2035 } else if (td->runstate >= TD_EXITED) {
2039 nr_running++; /* work-around... */
2045 log_err("fio: %d job%s failed to start\n", left,
2046 left > 1 ? "s" : "");
2047 for (i = 0; i < this_jobs; i++) {
2051 kill(td->pid, SIGTERM);
2057 * start created threads (TD_INITIALIZED -> TD_RUNNING).
2059 for_each_td(td, i) {
2060 if (td->runstate != TD_INITIALIZED)
2063 if (in_ramp_time(td))
2064 td_set_runstate(td, TD_RAMP);
2066 td_set_runstate(td, TD_RUNNING);
2069 m_rate += ddir_rw_sum(td->o.ratemin);
2070 t_rate += ddir_rw_sum(td->o.rate);
2072 fio_mutex_up(td->mutex);
2075 reap_threads(&nr_running, &t_rate, &m_rate);
2081 while (nr_running) {
2082 reap_threads(&nr_running, &t_rate, &m_rate);
2086 fio_idle_prof_stop();
2091 static void wait_for_helper_thread_exit(void)
2096 pthread_cond_signal(&helper_cond);
2097 pthread_join(helper_thread, &ret);
2100 static void free_disk_util(void)
2102 disk_util_prune_entries();
2104 pthread_cond_destroy(&helper_cond);
2107 static void *helper_thread_main(void *data)
2111 fio_mutex_up(startup_mutex);
2114 uint64_t sec = DISK_UTIL_MSEC / 1000;
2115 uint64_t nsec = (DISK_UTIL_MSEC % 1000) * 1000000;
2119 gettimeofday(&tv, NULL);
2120 ts.tv_sec = tv.tv_sec + sec;
2121 ts.tv_nsec = (tv.tv_usec * 1000) + nsec;
2122 if (ts.tv_nsec > 1000000000ULL) {
2123 ts.tv_nsec -= 1000000000ULL;
2127 pthread_cond_timedwait(&helper_cond, &helper_lock, &ts);
2129 ret = update_io_ticks();
2131 if (helper_do_stat) {
2133 __show_running_run_stats();
2137 print_thread_status();
2143 static int create_helper_thread(void)
2149 pthread_cond_init(&helper_cond, NULL);
2150 pthread_mutex_init(&helper_lock, NULL);
2152 ret = pthread_create(&helper_thread, NULL, helper_thread_main, NULL);
2154 log_err("Can't create helper thread: %s\n", strerror(ret));
2158 dprint(FD_MUTEX, "wait on startup_mutex\n");
2159 fio_mutex_down(startup_mutex);
2160 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2164 int fio_backend(void)
2166 struct thread_data *td;
2170 if (load_profile(exec_profile))
2173 exec_profile = NULL;
2179 struct log_params p = {
2180 .log_type = IO_LOG_TYPE_BW,
2183 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2184 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2185 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2188 startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
2189 if (startup_mutex == NULL)
2194 create_helper_thread();
2196 cgroup_list = smalloc(sizeof(*cgroup_list));
2197 INIT_FLIST_HEAD(cgroup_list);
2201 wait_for_helper_thread_exit();
2206 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2207 struct io_log *log = agg_io_log[i];
2215 for_each_td(td, i) {
2216 fio_options_free(td);
2217 if (td->rusage_sem) {
2218 fio_mutex_remove(td->rusage_sem);
2219 td->rusage_sem = NULL;
2224 cgroup_kill(cgroup_list);
2228 fio_mutex_remove(startup_mutex);