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 (td->o.verify_async)
530 io_u->end_io = verify_io_u_async;
532 io_u->end_io = verify_io_u;
536 ret = td_io_queue(td, io_u);
538 case FIO_Q_COMPLETED:
541 clear_io_u(td, io_u);
542 } else if (io_u->resid) {
543 int bytes = io_u->xfer_buflen - io_u->resid;
549 td_verror(td, EIO, "full resid");
554 io_u->xfer_buflen = io_u->resid;
555 io_u->xfer_buf += bytes;
556 io_u->offset += bytes;
558 if (ddir_rw(io_u->ddir))
559 td->ts.short_io_u[io_u->ddir]++;
562 if (io_u->offset == f->real_file_size)
565 requeue_io_u(td, &io_u);
568 ret = io_u_sync_complete(td, io_u, bytes_done);
576 requeue_io_u(td, &io_u);
577 ret2 = td_io_commit(td);
583 td_verror(td, -ret, "td_io_queue");
587 if (break_on_this_error(td, ddir, &ret))
591 * if we can queue more, do so. but check if there are
592 * completed io_u's first. Note that we can get BUSY even
593 * without IO queued, if the system is resource starved.
596 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
597 if (full || !td->o.iodepth_batch_complete) {
598 min_events = min(td->o.iodepth_batch_complete,
601 * if the queue is full, we MUST reap at least 1 event
603 if (full && !min_events)
608 * Reap required number of io units, if any,
609 * and do the verification on them through
610 * the callback handler
612 if (io_u_queued_complete(td, min_events, bytes_done) < 0) {
616 } while (full && (td->cur_depth > td->o.iodepth_low));
622 check_update_rusage(td);
625 min_events = td->cur_depth;
628 ret = io_u_queued_complete(td, min_events, NULL);
630 cleanup_pending_aio(td);
632 td_set_runstate(td, TD_RUNNING);
634 dprint(FD_VERIFY, "exiting loop\n");
637 static unsigned int exceeds_number_ios(struct thread_data *td)
639 unsigned long long number_ios;
641 if (!td->o.number_ios)
644 number_ios = ddir_rw_sum(td->this_io_blocks);
645 number_ios += td->io_u_queued + td->io_u_in_flight;
647 return number_ios >= td->o.number_ios;
650 static int io_bytes_exceeded(struct thread_data *td)
652 unsigned long long bytes, limit;
655 bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE];
656 else if (td_write(td))
657 bytes = td->this_io_bytes[DDIR_WRITE];
658 else if (td_read(td))
659 bytes = td->this_io_bytes[DDIR_READ];
661 bytes = td->this_io_bytes[DDIR_TRIM];
664 limit = td->o.io_limit;
668 return bytes >= limit || exceeds_number_ios(td);
672 * Main IO worker function. It retrieves io_u's to process and queues
673 * and reaps them, checking for rate and errors along the way.
675 * Returns number of bytes written and trimmed.
677 static uint64_t do_io(struct thread_data *td)
679 uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
682 uint64_t total_bytes, bytes_issued = 0;
684 if (in_ramp_time(td))
685 td_set_runstate(td, TD_RAMP);
687 td_set_runstate(td, TD_RUNNING);
692 * If verify_backlog is enabled, we'll run the verify in this
693 * handler as well. For that case, we may need up to twice the
696 total_bytes = td->o.size;
697 if (td->o.verify != VERIFY_NONE &&
698 (td_write(td) && td->o.verify_backlog))
699 total_bytes += td->o.size;
701 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
702 (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td) ||
704 struct timeval comp_time;
710 check_update_rusage(td);
712 if (td->terminate || td->done)
717 if (runtime_exceeded(td, &td->tv_cache)) {
718 __update_tv_cache(td);
719 if (runtime_exceeded(td, &td->tv_cache)) {
720 fio_mark_td_terminate(td);
725 if (flow_threshold_exceeded(td))
728 if (bytes_issued >= total_bytes)
732 if (IS_ERR_OR_NULL(io_u)) {
733 int err = PTR_ERR(io_u);
740 if (td->o.latency_target)
748 * Add verification end_io handler if:
749 * - Asked to verify (!td_rw(td))
750 * - Or the io_u is from our verify list (mixed write/ver)
752 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
753 ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
755 if (!td->o.verify_pattern_bytes) {
756 io_u->rand_seed = __rand(&td->verify_state);
757 if (sizeof(int) != sizeof(long *))
758 io_u->rand_seed *= __rand(&td->verify_state);
761 if (td->o.verify_async)
762 io_u->end_io = verify_io_u_async;
764 io_u->end_io = verify_io_u;
765 td_set_runstate(td, TD_VERIFYING);
766 } else if (in_ramp_time(td))
767 td_set_runstate(td, TD_RAMP);
769 td_set_runstate(td, TD_RUNNING);
772 * Always log IO before it's issued, so we know the specific
773 * order of it. The logged unit will track when the IO has
776 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
778 td->o.verify != VERIFY_NONE &&
779 !td->o.experimental_verify)
780 log_io_piece(td, io_u);
782 ret = td_io_queue(td, io_u);
784 case FIO_Q_COMPLETED:
787 unlog_io_piece(td, io_u);
788 clear_io_u(td, io_u);
789 } else if (io_u->resid) {
790 int bytes = io_u->xfer_buflen - io_u->resid;
791 struct fio_file *f = io_u->file;
793 bytes_issued += bytes;
795 trim_io_piece(td, io_u);
801 unlog_io_piece(td, io_u);
802 td_verror(td, EIO, "full resid");
807 io_u->xfer_buflen = io_u->resid;
808 io_u->xfer_buf += bytes;
809 io_u->offset += bytes;
811 if (ddir_rw(io_u->ddir))
812 td->ts.short_io_u[io_u->ddir]++;
814 if (io_u->offset == f->real_file_size)
817 requeue_io_u(td, &io_u);
820 if (__should_check_rate(td, DDIR_READ) ||
821 __should_check_rate(td, DDIR_WRITE) ||
822 __should_check_rate(td, DDIR_TRIM))
823 fio_gettime(&comp_time, NULL);
825 ret = io_u_sync_complete(td, io_u, bytes_done);
828 bytes_issued += io_u->xfer_buflen;
833 * if the engine doesn't have a commit hook,
834 * the io_u is really queued. if it does have such
835 * a hook, it has to call io_u_queued() itself.
837 if (td->io_ops->commit == NULL)
838 io_u_queued(td, io_u);
839 bytes_issued += io_u->xfer_buflen;
842 unlog_io_piece(td, io_u);
843 requeue_io_u(td, &io_u);
844 ret2 = td_io_commit(td);
854 if (break_on_this_error(td, ddir, &ret))
858 * See if we need to complete some commands. Note that we
859 * can get BUSY even without IO queued, if the system is
863 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
864 if (full || !td->o.iodepth_batch_complete) {
865 min_evts = min(td->o.iodepth_batch_complete,
868 * if the queue is full, we MUST reap at least 1 event
870 if (full && !min_evts)
873 if (__should_check_rate(td, DDIR_READ) ||
874 __should_check_rate(td, DDIR_WRITE) ||
875 __should_check_rate(td, DDIR_TRIM))
876 fio_gettime(&comp_time, NULL);
879 ret = io_u_queued_complete(td, min_evts, bytes_done);
883 } while (full && (td->cur_depth > td->o.iodepth_low));
888 if (!ddir_rw_sum(bytes_done) && !(td->io_ops->flags & FIO_NOIO))
891 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
892 if (check_min_rate(td, &comp_time, bytes_done)) {
893 if (exitall_on_terminate)
894 fio_terminate_threads(td->groupid);
895 td_verror(td, EIO, "check_min_rate");
899 if (!in_ramp_time(td) && td->o.latency_target)
900 lat_target_check(td);
902 if (td->o.thinktime) {
903 unsigned long long b;
905 b = ddir_rw_sum(td->io_blocks);
906 if (!(b % td->o.thinktime_blocks)) {
911 if (td->o.thinktime_spin)
912 usec_spin(td->o.thinktime_spin);
914 left = td->o.thinktime - td->o.thinktime_spin;
916 usec_sleep(td, left);
921 check_update_rusage(td);
923 if (td->trim_entries)
924 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
926 if (td->o.fill_device && td->error == ENOSPC) {
928 fio_mark_td_terminate(td);
935 ret = io_u_queued_complete(td, i, bytes_done);
936 if (td->o.fill_device && td->error == ENOSPC)
940 if (should_fsync(td) && td->o.end_fsync) {
941 td_set_runstate(td, TD_FSYNCING);
943 for_each_file(td, f, i) {
944 if (!fio_file_fsync(td, f))
947 log_err("fio: end_fsync failed for file %s\n",
952 cleanup_pending_aio(td);
955 * stop job if we failed doing any IO
957 if (!ddir_rw_sum(td->this_io_bytes))
960 return bytes_done[DDIR_WRITE] + bytes_done[DDIR_TRIM];
963 static void cleanup_io_u(struct thread_data *td)
967 while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
969 if (td->io_ops->io_u_free)
970 td->io_ops->io_u_free(td, io_u);
972 fio_memfree(io_u, sizeof(*io_u));
977 io_u_rexit(&td->io_u_requeues);
978 io_u_qexit(&td->io_u_freelist);
979 io_u_qexit(&td->io_u_all);
982 static int init_io_u(struct thread_data *td)
985 unsigned int max_bs, min_write;
986 int cl_align, i, max_units;
987 int data_xfer = 1, err;
990 max_units = td->o.iodepth;
991 max_bs = td_max_bs(td);
992 min_write = td->o.min_bs[DDIR_WRITE];
993 td->orig_buffer_size = (unsigned long long) max_bs
994 * (unsigned long long) max_units;
996 if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
1000 err += io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1001 err += io_u_qinit(&td->io_u_freelist, td->o.iodepth);
1002 err += io_u_qinit(&td->io_u_all, td->o.iodepth);
1005 log_err("fio: failed setting up IO queues\n");
1010 * if we may later need to do address alignment, then add any
1011 * possible adjustment here so that we don't cause a buffer
1012 * overflow later. this adjustment may be too much if we get
1013 * lucky and the allocator gives us an aligned address.
1015 if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1016 (td->io_ops->flags & FIO_RAWIO))
1017 td->orig_buffer_size += page_mask + td->o.mem_align;
1019 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1022 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1023 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1026 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1027 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1031 if (data_xfer && allocate_io_mem(td))
1034 if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1035 (td->io_ops->flags & FIO_RAWIO))
1036 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
1038 p = td->orig_buffer;
1040 cl_align = os_cache_line_size();
1042 for (i = 0; i < max_units; i++) {
1048 ptr = fio_memalign(cl_align, sizeof(*io_u));
1050 log_err("fio: unable to allocate aligned memory\n");
1055 memset(io_u, 0, sizeof(*io_u));
1056 INIT_FLIST_HEAD(&io_u->verify_list);
1057 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1061 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1064 io_u_fill_buffer(td, io_u, min_write, max_bs);
1065 if (td_write(td) && td->o.verify_pattern_bytes) {
1067 * Fill the buffer with the pattern if we are
1068 * going to be doing writes.
1070 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1075 io_u->flags = IO_U_F_FREE;
1076 io_u_qpush(&td->io_u_freelist, io_u);
1079 * io_u never leaves this stack, used for iteration of all
1082 io_u_qpush(&td->io_u_all, io_u);
1084 if (td->io_ops->io_u_init) {
1085 int ret = td->io_ops->io_u_init(td, io_u);
1088 log_err("fio: failed to init engine data: %d\n", ret);
1099 static int switch_ioscheduler(struct thread_data *td)
1101 char tmp[256], tmp2[128];
1105 if (td->io_ops->flags & FIO_DISKLESSIO)
1108 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
1110 f = fopen(tmp, "r+");
1112 if (errno == ENOENT) {
1113 log_err("fio: os or kernel doesn't support IO scheduler"
1117 td_verror(td, errno, "fopen iosched");
1124 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1125 if (ferror(f) || ret != 1) {
1126 td_verror(td, errno, "fwrite");
1134 * Read back and check that the selected scheduler is now the default.
1136 ret = fread(tmp, sizeof(tmp), 1, f);
1137 if (ferror(f) || ret < 0) {
1138 td_verror(td, errno, "fread");
1142 tmp[sizeof(tmp) - 1] = '\0';
1145 sprintf(tmp2, "[%s]", td->o.ioscheduler);
1146 if (!strstr(tmp, tmp2)) {
1147 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1148 td_verror(td, EINVAL, "iosched_switch");
1157 static int keep_running(struct thread_data *td)
1159 unsigned long long limit;
1163 if (td->o.time_based)
1169 if (exceeds_number_ios(td))
1173 limit = td->o.io_limit;
1177 if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1181 * If the difference is less than the minimum IO size, we
1184 diff = limit - ddir_rw_sum(td->io_bytes);
1185 if (diff < td_max_bs(td))
1188 if (fio_files_done(td))
1197 static int exec_string(struct thread_options *o, const char *string, const char *mode)
1199 int ret, newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
1202 str = malloc(newlen);
1203 sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
1205 log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1208 log_err("fio: exec of cmd <%s> failed\n", str);
1215 * Dry run to compute correct state of numberio for verification.
1217 static uint64_t do_dry_run(struct thread_data *td)
1219 uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
1221 td_set_runstate(td, TD_RUNNING);
1223 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1224 (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td)) {
1228 if (td->terminate || td->done)
1231 io_u = get_io_u(td);
1235 io_u->flags |= IO_U_F_FLIGHT;
1238 if (ddir_rw(acct_ddir(io_u)))
1239 td->io_issues[acct_ddir(io_u)]++;
1240 if (ddir_rw(io_u->ddir)) {
1241 io_u_mark_depth(td, 1);
1242 td->ts.total_io_u[io_u->ddir]++;
1245 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1247 td->o.verify != VERIFY_NONE &&
1248 !td->o.experimental_verify)
1249 log_io_piece(td, io_u);
1251 ret = io_u_sync_complete(td, io_u, bytes_done);
1255 return bytes_done[DDIR_WRITE] + bytes_done[DDIR_TRIM];
1259 * Entry point for the thread based jobs. The process based jobs end up
1260 * here as well, after a little setup.
1262 static void *thread_main(void *data)
1264 unsigned long long elapsed;
1265 struct thread_data *td = data;
1266 struct thread_options *o = &td->o;
1267 pthread_condattr_t attr;
1271 if (!o->use_thread) {
1277 fio_local_clock_init(o->use_thread);
1279 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1282 fio_server_send_start(td);
1284 INIT_FLIST_HEAD(&td->io_log_list);
1285 INIT_FLIST_HEAD(&td->io_hist_list);
1286 INIT_FLIST_HEAD(&td->verify_list);
1287 INIT_FLIST_HEAD(&td->trim_list);
1288 INIT_FLIST_HEAD(&td->next_rand_list);
1289 pthread_mutex_init(&td->io_u_lock, NULL);
1290 td->io_hist_tree = RB_ROOT;
1292 pthread_condattr_init(&attr);
1293 pthread_cond_init(&td->verify_cond, &attr);
1294 pthread_cond_init(&td->free_cond, &attr);
1296 td_set_runstate(td, TD_INITIALIZED);
1297 dprint(FD_MUTEX, "up startup_mutex\n");
1298 fio_mutex_up(startup_mutex);
1299 dprint(FD_MUTEX, "wait on td->mutex\n");
1300 fio_mutex_down(td->mutex);
1301 dprint(FD_MUTEX, "done waiting on td->mutex\n");
1304 * A new gid requires privilege, so we need to do this before setting
1307 if (o->gid != -1U && setgid(o->gid)) {
1308 td_verror(td, errno, "setgid");
1311 if (o->uid != -1U && setuid(o->uid)) {
1312 td_verror(td, errno, "setuid");
1317 * If we have a gettimeofday() thread, make sure we exclude that
1318 * thread from this job
1321 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1324 * Set affinity first, in case it has an impact on the memory
1327 if (o->cpumask_set) {
1328 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1329 ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1331 log_err("fio: no CPUs set\n");
1332 log_err("fio: Try increasing number of available CPUs\n");
1333 td_verror(td, EINVAL, "cpus_split");
1337 ret = fio_setaffinity(td->pid, o->cpumask);
1339 td_verror(td, errno, "cpu_set_affinity");
1344 #ifdef CONFIG_LIBNUMA
1345 /* numa node setup */
1346 if (o->numa_cpumask_set || o->numa_memmask_set) {
1347 struct bitmask *mask;
1349 if (numa_available() < 0) {
1350 td_verror(td, errno, "Does not support NUMA API\n");
1354 if (o->numa_cpumask_set) {
1355 mask = numa_parse_nodestring(o->numa_cpunodes);
1356 ret = numa_run_on_node_mask(mask);
1357 numa_free_nodemask(mask);
1359 td_verror(td, errno, \
1360 "numa_run_on_node_mask failed\n");
1365 if (o->numa_memmask_set) {
1368 if (o->numa_memnodes)
1369 mask = numa_parse_nodestring(o->numa_memnodes);
1371 switch (o->numa_mem_mode) {
1372 case MPOL_INTERLEAVE:
1373 numa_set_interleave_mask(mask);
1376 numa_set_membind(mask);
1379 numa_set_localalloc();
1381 case MPOL_PREFERRED:
1382 numa_set_preferred(o->numa_mem_prefer_node);
1390 numa_free_nodemask(mask);
1396 if (fio_pin_memory(td))
1400 * May alter parameters that init_io_u() will use, so we need to
1409 if (o->verify_async && verify_async_init(td))
1413 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1415 td_verror(td, errno, "ioprio_set");
1420 if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1424 if (nice(o->nice) == -1 && errno != 0) {
1425 td_verror(td, errno, "nice");
1429 if (o->ioscheduler && switch_ioscheduler(td))
1432 if (!o->create_serialize && setup_files(td))
1438 if (init_random_map(td))
1441 if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
1445 if (pre_read_files(td) < 0)
1449 if (td->flags & TD_F_COMPRESS_LOG)
1450 tp_init(&td->tp_data);
1452 fio_verify_init(td);
1454 fio_gettime(&td->epoch, NULL);
1455 fio_getrusage(&td->ru_start);
1457 while (keep_running(td)) {
1458 uint64_t verify_bytes;
1460 fio_gettime(&td->start, NULL);
1461 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1462 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1463 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1465 if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1466 o->ratemin[DDIR_TRIM]) {
1467 memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1468 sizeof(td->bw_sample_time));
1469 memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1470 sizeof(td->bw_sample_time));
1471 memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1472 sizeof(td->bw_sample_time));
1478 prune_io_piece_log(td);
1480 if (td->o.verify_only && (td_write(td) || td_rw(td)))
1481 verify_bytes = do_dry_run(td);
1483 verify_bytes = do_io(td);
1487 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1488 elapsed = utime_since_now(&td->start);
1489 td->ts.runtime[DDIR_READ] += elapsed;
1491 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1492 elapsed = utime_since_now(&td->start);
1493 td->ts.runtime[DDIR_WRITE] += elapsed;
1495 if (td_trim(td) && td->io_bytes[DDIR_TRIM]) {
1496 elapsed = utime_since_now(&td->start);
1497 td->ts.runtime[DDIR_TRIM] += elapsed;
1500 if (td->error || td->terminate)
1503 if (!o->do_verify ||
1504 o->verify == VERIFY_NONE ||
1505 (td->io_ops->flags & FIO_UNIDIR))
1510 fio_gettime(&td->start, NULL);
1512 do_verify(td, verify_bytes);
1514 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1516 if (td->error || td->terminate)
1520 update_rusage_stat(td);
1521 td->ts.runtime[DDIR_READ] = (td->ts.runtime[DDIR_READ] + 999) / 1000;
1522 td->ts.runtime[DDIR_WRITE] = (td->ts.runtime[DDIR_WRITE] + 999) / 1000;
1523 td->ts.runtime[DDIR_TRIM] = (td->ts.runtime[DDIR_TRIM] + 999) / 1000;
1524 td->ts.total_run_time = mtime_since_now(&td->epoch);
1525 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1526 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1527 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1529 fio_unpin_memory(td);
1531 fio_writeout_logs(td);
1533 if (td->flags & TD_F_COMPRESS_LOG)
1534 tp_exit(&td->tp_data);
1536 if (o->exec_postrun)
1537 exec_string(o, o->exec_postrun, (const char *)"postrun");
1539 if (exitall_on_terminate)
1540 fio_terminate_threads(td->groupid);
1544 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1547 if (o->verify_async)
1548 verify_async_exit(td);
1550 close_and_free_files(td);
1553 cgroup_shutdown(td, &cgroup_mnt);
1555 if (o->cpumask_set) {
1556 ret = fio_cpuset_exit(&o->cpumask);
1558 td_verror(td, ret, "fio_cpuset_exit");
1562 * do this very late, it will log file closing as well
1564 if (o->write_iolog_file)
1565 write_iolog_close(td);
1567 fio_mutex_remove(td->mutex);
1570 td_set_runstate(td, TD_EXITED);
1573 * Do this last after setting our runstate to exited, so we
1574 * know that the stat thread is signaled.
1576 check_update_rusage(td);
1578 return (void *) (uintptr_t) td->error;
1583 * We cannot pass the td data into a forked process, so attach the td and
1584 * pass it to the thread worker.
1586 static int fork_main(int shmid, int offset)
1588 struct thread_data *td;
1591 #if !defined(__hpux) && !defined(CONFIG_NO_SHM)
1592 data = shmat(shmid, NULL, 0);
1593 if (data == (void *) -1) {
1601 * HP-UX inherits shm mappings?
1606 td = data + offset * sizeof(struct thread_data);
1607 ret = thread_main(td);
1609 return (int) (uintptr_t) ret;
1612 static void dump_td_info(struct thread_data *td)
1614 log_err("fio: job '%s' hasn't exited in %lu seconds, it appears to "
1615 "be stuck. Doing forceful exit of this job.\n", td->o.name,
1616 (unsigned long) time_since_now(&td->terminate_time));
1620 * Run over the job map and reap the threads that have exited, if any.
1622 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1623 unsigned int *m_rate)
1625 struct thread_data *td;
1626 unsigned int cputhreads, realthreads, pending;
1630 * reap exited threads (TD_EXITED -> TD_REAPED)
1632 realthreads = pending = cputhreads = 0;
1633 for_each_td(td, i) {
1637 * ->io_ops is NULL for a thread that has closed its
1640 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1649 if (td->runstate == TD_REAPED)
1651 if (td->o.use_thread) {
1652 if (td->runstate == TD_EXITED) {
1653 td_set_runstate(td, TD_REAPED);
1660 if (td->runstate == TD_EXITED)
1664 * check if someone quit or got killed in an unusual way
1666 ret = waitpid(td->pid, &status, flags);
1668 if (errno == ECHILD) {
1669 log_err("fio: pid=%d disappeared %d\n",
1670 (int) td->pid, td->runstate);
1672 td_set_runstate(td, TD_REAPED);
1676 } else if (ret == td->pid) {
1677 if (WIFSIGNALED(status)) {
1678 int sig = WTERMSIG(status);
1680 if (sig != SIGTERM && sig != SIGUSR2)
1681 log_err("fio: pid=%d, got signal=%d\n",
1682 (int) td->pid, sig);
1684 td_set_runstate(td, TD_REAPED);
1687 if (WIFEXITED(status)) {
1688 if (WEXITSTATUS(status) && !td->error)
1689 td->error = WEXITSTATUS(status);
1691 td_set_runstate(td, TD_REAPED);
1697 * If the job is stuck, do a forceful timeout of it and
1700 if (td->terminate &&
1701 time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
1703 td_set_runstate(td, TD_REAPED);
1708 * thread is not dead, continue
1714 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1715 (*t_rate) -= ddir_rw_sum(td->o.rate);
1722 done_secs += mtime_since_now(&td->epoch) / 1000;
1723 profile_td_exit(td);
1726 if (*nr_running == cputhreads && !pending && realthreads)
1727 fio_terminate_threads(TERMINATE_ALL);
1730 static void do_usleep(unsigned int usecs)
1732 check_for_running_stats();
1737 * Main function for kicking off and reaping jobs, as needed.
1739 static void run_threads(void)
1741 struct thread_data *td;
1742 unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
1745 if (fio_gtod_offload && fio_start_gtod_thread())
1748 fio_idle_prof_init();
1752 nr_thread = nr_process = 0;
1753 for_each_td(td, i) {
1754 if (td->o.use_thread)
1760 if (output_format == FIO_OUTPUT_NORMAL) {
1761 log_info("Starting ");
1763 log_info("%d thread%s", nr_thread,
1764 nr_thread > 1 ? "s" : "");
1768 log_info("%d process%s", nr_process,
1769 nr_process > 1 ? "es" : "");
1775 todo = thread_number;
1778 m_rate = t_rate = 0;
1780 for_each_td(td, i) {
1781 print_status_init(td->thread_number - 1);
1783 if (!td->o.create_serialize)
1787 * do file setup here so it happens sequentially,
1788 * we don't want X number of threads getting their
1789 * client data interspersed on disk
1791 if (setup_files(td)) {
1794 log_err("fio: pid=%d, err=%d/%s\n",
1795 (int) td->pid, td->error, td->verror);
1796 td_set_runstate(td, TD_REAPED);
1803 * for sharing to work, each job must always open
1804 * its own files. so close them, if we opened them
1807 for_each_file(td, f, j) {
1808 if (fio_file_open(f))
1809 td_io_close_file(td, f);
1814 /* start idle threads before io threads start to run */
1815 fio_idle_prof_start();
1820 struct thread_data *map[REAL_MAX_JOBS];
1821 struct timeval this_start;
1822 int this_jobs = 0, left;
1825 * create threads (TD_NOT_CREATED -> TD_CREATED)
1827 for_each_td(td, i) {
1828 if (td->runstate != TD_NOT_CREATED)
1832 * never got a chance to start, killed by other
1833 * thread for some reason
1835 if (td->terminate) {
1840 if (td->o.start_delay) {
1841 spent = utime_since_genesis();
1843 if (td->o.start_delay > spent)
1847 if (td->o.stonewall && (nr_started || nr_running)) {
1848 dprint(FD_PROCESS, "%s: stonewall wait\n",
1855 td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED);
1856 td->update_rusage = 0;
1859 * Set state to created. Thread will transition
1860 * to TD_INITIALIZED when it's done setting up.
1862 td_set_runstate(td, TD_CREATED);
1863 map[this_jobs++] = td;
1866 if (td->o.use_thread) {
1869 dprint(FD_PROCESS, "will pthread_create\n");
1870 ret = pthread_create(&td->thread, NULL,
1873 log_err("pthread_create: %s\n",
1878 ret = pthread_detach(td->thread);
1880 log_err("pthread_detach: %s",
1884 dprint(FD_PROCESS, "will fork\n");
1887 int ret = fork_main(shm_id, i);
1890 } else if (i == fio_debug_jobno)
1891 *fio_debug_jobp = pid;
1893 dprint(FD_MUTEX, "wait on startup_mutex\n");
1894 if (fio_mutex_down_timeout(startup_mutex, 10)) {
1895 log_err("fio: job startup hung? exiting.\n");
1896 fio_terminate_threads(TERMINATE_ALL);
1901 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1905 * Wait for the started threads to transition to
1908 fio_gettime(&this_start, NULL);
1910 while (left && !fio_abort) {
1911 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1916 for (i = 0; i < this_jobs; i++) {
1920 if (td->runstate == TD_INITIALIZED) {
1923 } else if (td->runstate >= TD_EXITED) {
1927 nr_running++; /* work-around... */
1933 log_err("fio: %d job%s failed to start\n", left,
1934 left > 1 ? "s" : "");
1935 for (i = 0; i < this_jobs; i++) {
1939 kill(td->pid, SIGTERM);
1945 * start created threads (TD_INITIALIZED -> TD_RUNNING).
1947 for_each_td(td, i) {
1948 if (td->runstate != TD_INITIALIZED)
1951 if (in_ramp_time(td))
1952 td_set_runstate(td, TD_RAMP);
1954 td_set_runstate(td, TD_RUNNING);
1957 m_rate += ddir_rw_sum(td->o.ratemin);
1958 t_rate += ddir_rw_sum(td->o.rate);
1960 fio_mutex_up(td->mutex);
1963 reap_threads(&nr_running, &t_rate, &m_rate);
1969 while (nr_running) {
1970 reap_threads(&nr_running, &t_rate, &m_rate);
1974 fio_idle_prof_stop();
1979 static void wait_for_helper_thread_exit(void)
1984 pthread_cond_signal(&helper_cond);
1985 pthread_join(helper_thread, &ret);
1988 static void free_disk_util(void)
1990 disk_util_prune_entries();
1992 pthread_cond_destroy(&helper_cond);
1995 static void *helper_thread_main(void *data)
1999 fio_mutex_up(startup_mutex);
2002 uint64_t sec = DISK_UTIL_MSEC / 1000;
2003 uint64_t nsec = (DISK_UTIL_MSEC % 1000) * 1000000;
2007 gettimeofday(&tv, NULL);
2008 ts.tv_sec = tv.tv_sec + sec;
2009 ts.tv_nsec = (tv.tv_usec * 1000) + nsec;
2010 if (ts.tv_nsec > 1000000000ULL) {
2011 ts.tv_nsec -= 1000000000ULL;
2015 pthread_cond_timedwait(&helper_cond, &helper_lock, &ts);
2017 ret = update_io_ticks();
2019 if (helper_do_stat) {
2021 __show_running_run_stats();
2025 print_thread_status();
2031 static int create_helper_thread(void)
2037 pthread_cond_init(&helper_cond, NULL);
2038 pthread_mutex_init(&helper_lock, NULL);
2040 ret = pthread_create(&helper_thread, NULL, helper_thread_main, NULL);
2042 log_err("Can't create helper thread: %s\n", strerror(ret));
2046 dprint(FD_MUTEX, "wait on startup_mutex\n");
2047 fio_mutex_down(startup_mutex);
2048 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2052 int fio_backend(void)
2054 struct thread_data *td;
2058 if (load_profile(exec_profile))
2061 exec_profile = NULL;
2067 struct log_params p = {
2068 .log_type = IO_LOG_TYPE_BW,
2071 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2072 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2073 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2076 startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
2077 if (startup_mutex == NULL)
2082 create_helper_thread();
2084 cgroup_list = smalloc(sizeof(*cgroup_list));
2085 INIT_FLIST_HEAD(cgroup_list);
2089 wait_for_helper_thread_exit();
2094 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2095 struct io_log *log = agg_io_log[i];
2103 for_each_td(td, i) {
2104 fio_options_free(td);
2105 fio_mutex_remove(td->rusage_sem);
2106 td->rusage_sem = NULL;
2110 cgroup_kill(cgroup_list);
2114 fio_mutex_remove(startup_mutex);