2 * fio - the flexible io tester
4 * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
5 * Copyright (C) 2006-2012 Jens Axboe <axboe@kernel.dk>
7 * The license below covers all files distributed with fio unless otherwise
8 * noted in the file itself.
10 * This program is free software; you can redistribute it and/or modify
11 * it under the terms of the GNU General Public License version 2 as
12 * published by the Free Software Foundation.
14 * This program is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
17 * GNU General Public License for more details.
19 * You should have received a copy of the GNU General Public License
20 * along with this program; if not, write to the Free Software
21 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
40 #ifndef FIO_NO_HAVE_SHM_H
53 #include "lib/getrusage.h"
57 static pthread_t disk_util_thread;
58 static struct fio_mutex *disk_thread_mutex;
59 static struct fio_mutex *startup_mutex;
60 static struct flist_head *cgroup_list;
61 static char *cgroup_mnt;
62 static int exit_value;
63 static volatile int fio_abort;
64 static unsigned int nr_process = 0;
65 static unsigned int nr_thread = 0;
67 struct io_log *agg_io_log[DDIR_RWDIR_CNT];
70 unsigned int thread_number = 0;
71 unsigned int stat_number = 0;
74 unsigned long done_secs = 0;
75 volatile int disk_util_exit = 0;
77 #define PAGE_ALIGN(buf) \
78 (char *) (((uintptr_t) (buf) + page_mask) & ~page_mask)
80 #define JOB_START_TIMEOUT (5 * 1000)
82 static void sig_int(int sig)
86 fio_server_got_signal(sig);
88 log_info("\nfio: terminating on signal %d\n", sig);
93 fio_terminate_threads(TERMINATE_ALL);
97 static void sig_show_status(int sig)
99 show_running_run_stats();
102 static void set_sig_handlers(void)
104 struct sigaction act;
106 memset(&act, 0, sizeof(act));
107 act.sa_handler = sig_int;
108 act.sa_flags = SA_RESTART;
109 sigaction(SIGINT, &act, NULL);
111 memset(&act, 0, sizeof(act));
112 act.sa_handler = sig_int;
113 act.sa_flags = SA_RESTART;
114 sigaction(SIGTERM, &act, NULL);
116 /* Windows uses SIGBREAK as a quit signal from other applications */
118 memset(&act, 0, sizeof(act));
119 act.sa_handler = sig_int;
120 act.sa_flags = SA_RESTART;
121 sigaction(SIGBREAK, &act, NULL);
124 memset(&act, 0, sizeof(act));
125 act.sa_handler = sig_show_status;
126 act.sa_flags = SA_RESTART;
127 sigaction(SIGUSR1, &act, NULL);
130 memset(&act, 0, sizeof(act));
131 act.sa_handler = sig_int;
132 act.sa_flags = SA_RESTART;
133 sigaction(SIGPIPE, &act, NULL);
138 * Check if we are above the minimum rate given.
140 static int __check_min_rate(struct thread_data *td, struct timeval *now,
143 unsigned long long bytes = 0;
144 unsigned long iops = 0;
147 unsigned int ratemin = 0;
148 unsigned int rate_iops = 0;
149 unsigned int rate_iops_min = 0;
151 assert(ddir_rw(ddir));
153 if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
157 * allow a 2 second settle period in the beginning
159 if (mtime_since(&td->start, now) < 2000)
162 iops += td->this_io_blocks[ddir];
163 bytes += td->this_io_bytes[ddir];
164 ratemin += td->o.ratemin[ddir];
165 rate_iops += td->o.rate_iops[ddir];
166 rate_iops_min += td->o.rate_iops_min[ddir];
169 * if rate blocks is set, sample is running
171 if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
172 spent = mtime_since(&td->lastrate[ddir], now);
173 if (spent < td->o.ratecycle)
176 if (td->o.rate[ddir]) {
178 * check bandwidth specified rate
180 if (bytes < td->rate_bytes[ddir]) {
181 log_err("%s: min rate %u not met\n", td->o.name,
186 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
190 if (rate < ratemin ||
191 bytes < td->rate_bytes[ddir]) {
192 log_err("%s: min rate %u not met, got"
193 " %luKB/sec\n", td->o.name,
200 * checks iops specified rate
202 if (iops < rate_iops) {
203 log_err("%s: min iops rate %u not met\n",
204 td->o.name, rate_iops);
208 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
212 if (rate < rate_iops_min ||
213 iops < td->rate_blocks[ddir]) {
214 log_err("%s: min iops rate %u not met,"
215 " got %lu\n", td->o.name,
216 rate_iops_min, rate);
222 td->rate_bytes[ddir] = bytes;
223 td->rate_blocks[ddir] = iops;
224 memcpy(&td->lastrate[ddir], now, sizeof(*now));
228 static int check_min_rate(struct thread_data *td, struct timeval *now,
229 uint64_t *bytes_done)
233 if (bytes_done[DDIR_READ])
234 ret |= __check_min_rate(td, now, DDIR_READ);
235 if (bytes_done[DDIR_WRITE])
236 ret |= __check_min_rate(td, now, DDIR_WRITE);
237 if (bytes_done[DDIR_TRIM])
238 ret |= __check_min_rate(td, now, DDIR_TRIM);
244 * When job exits, we can cancel the in-flight IO if we are using async
245 * io. Attempt to do so.
247 static void cleanup_pending_aio(struct thread_data *td)
252 * get immediately available events, if any
254 r = io_u_queued_complete(td, 0, NULL);
259 * now cancel remaining active events
261 if (td->io_ops->cancel) {
265 io_u_qiter(&td->io_u_all, io_u, i) {
266 if (io_u->flags & IO_U_F_FLIGHT) {
267 r = td->io_ops->cancel(td, io_u);
275 r = io_u_queued_complete(td, td->cur_depth, NULL);
279 * Helper to handle the final sync of a file. Works just like the normal
280 * io path, just does everything sync.
282 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
284 struct io_u *io_u = __get_io_u(td);
290 io_u->ddir = DDIR_SYNC;
293 if (td_io_prep(td, io_u)) {
299 ret = td_io_queue(td, io_u);
301 td_verror(td, io_u->error, "td_io_queue");
304 } else if (ret == FIO_Q_QUEUED) {
305 if (io_u_queued_complete(td, 1, NULL) < 0)
307 } else if (ret == FIO_Q_COMPLETED) {
309 td_verror(td, io_u->error, "td_io_queue");
313 if (io_u_sync_complete(td, io_u, NULL) < 0)
315 } else if (ret == FIO_Q_BUSY) {
316 if (td_io_commit(td))
324 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
328 if (fio_file_open(f))
329 return fio_io_sync(td, f);
331 if (td_io_open_file(td, f))
334 ret = fio_io_sync(td, f);
335 td_io_close_file(td, f);
339 static inline void __update_tv_cache(struct thread_data *td)
341 fio_gettime(&td->tv_cache, NULL);
344 static inline void update_tv_cache(struct thread_data *td)
346 if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
347 __update_tv_cache(td);
350 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
352 if (in_ramp_time(td))
356 if (utime_since(&td->epoch, t) >= td->o.timeout)
362 static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
367 if (ret < 0 || td->error) {
369 enum error_type_bit eb;
374 eb = td_error_type(ddir, err);
375 if (!(td->o.continue_on_error & (1 << eb)))
378 if (td_non_fatal_error(td, eb, err)) {
380 * Continue with the I/Os in case of
383 update_error_count(td, err);
387 } else if (td->o.fill_device && err == ENOSPC) {
389 * We expect to hit this error if
390 * fill_device option is set.
397 * Stop the I/O in case of a fatal
400 update_error_count(td, err);
408 static void check_update_rusage(struct thread_data *td)
410 if (td->update_rusage) {
411 td->update_rusage = 0;
412 update_rusage_stat(td);
413 fio_mutex_up(td->rusage_sem);
418 * The main verify engine. Runs over the writes we previously submitted,
419 * reads the blocks back in, and checks the crc/md5 of the data.
421 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
423 uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
429 dprint(FD_VERIFY, "starting loop\n");
432 * sync io first and invalidate cache, to make sure we really
435 for_each_file(td, f, i) {
436 if (!fio_file_open(f))
438 if (fio_io_sync(td, f))
440 if (file_invalidate_cache(td, f))
444 check_update_rusage(td);
449 td_set_runstate(td, TD_VERIFYING);
452 while (!td->terminate) {
457 check_update_rusage(td);
459 if (runtime_exceeded(td, &td->tv_cache)) {
460 __update_tv_cache(td);
461 if (runtime_exceeded(td, &td->tv_cache)) {
467 if (flow_threshold_exceeded(td))
470 if (!td->o.experimental_verify) {
471 io_u = __get_io_u(td);
475 if (get_next_verify(td, io_u)) {
480 if (td_io_prep(td, io_u)) {
485 if (ddir_rw_sum(bytes_done) + td->o.rw_min_bs > verify_bytes)
488 while ((io_u = get_io_u(td)) != NULL) {
496 * We are only interested in the places where
497 * we wrote or trimmed IOs. Turn those into
498 * reads for verification purposes.
500 if (io_u->ddir == DDIR_READ) {
502 * Pretend we issued it for rwmix
505 td->io_issues[DDIR_READ]++;
508 } else if (io_u->ddir == DDIR_TRIM) {
509 io_u->ddir = DDIR_READ;
510 io_u->flags |= IO_U_F_TRIMMED;
512 } else if (io_u->ddir == DDIR_WRITE) {
513 io_u->ddir = DDIR_READ;
525 if (td->o.verify_async)
526 io_u->end_io = verify_io_u_async;
528 io_u->end_io = verify_io_u;
532 ret = td_io_queue(td, io_u);
534 case FIO_Q_COMPLETED:
537 clear_io_u(td, io_u);
538 } else if (io_u->resid) {
539 int bytes = io_u->xfer_buflen - io_u->resid;
545 td_verror(td, EIO, "full resid");
550 io_u->xfer_buflen = io_u->resid;
551 io_u->xfer_buf += bytes;
552 io_u->offset += bytes;
554 if (ddir_rw(io_u->ddir))
555 td->ts.short_io_u[io_u->ddir]++;
558 if (io_u->offset == f->real_file_size)
561 requeue_io_u(td, &io_u);
564 ret = io_u_sync_complete(td, io_u, bytes_done);
572 requeue_io_u(td, &io_u);
573 ret2 = td_io_commit(td);
579 td_verror(td, -ret, "td_io_queue");
583 if (break_on_this_error(td, ddir, &ret))
587 * if we can queue more, do so. but check if there are
588 * completed io_u's first. Note that we can get BUSY even
589 * without IO queued, if the system is resource starved.
592 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
593 if (full || !td->o.iodepth_batch_complete) {
594 min_events = min(td->o.iodepth_batch_complete,
597 * if the queue is full, we MUST reap at least 1 event
599 if (full && !min_events)
604 * Reap required number of io units, if any,
605 * and do the verification on them through
606 * the callback handler
608 if (io_u_queued_complete(td, min_events, bytes_done) < 0) {
612 } while (full && (td->cur_depth > td->o.iodepth_low));
618 check_update_rusage(td);
621 min_events = td->cur_depth;
624 ret = io_u_queued_complete(td, min_events, NULL);
626 cleanup_pending_aio(td);
628 td_set_runstate(td, TD_RUNNING);
630 dprint(FD_VERIFY, "exiting loop\n");
633 static unsigned int exceeds_number_ios(struct thread_data *td)
635 unsigned long long number_ios;
637 if (!td->o.number_ios)
640 number_ios = ddir_rw_sum(td->this_io_blocks);
641 number_ios += td->io_u_queued + td->io_u_in_flight;
643 return number_ios >= td->o.number_ios;
646 static int io_bytes_exceeded(struct thread_data *td)
648 unsigned long long bytes, limit;
651 bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE];
652 else if (td_write(td))
653 bytes = td->this_io_bytes[DDIR_WRITE];
654 else if (td_read(td))
655 bytes = td->this_io_bytes[DDIR_READ];
657 bytes = td->this_io_bytes[DDIR_TRIM];
660 limit = td->o.io_limit;
664 return bytes >= limit || exceeds_number_ios(td);
668 * Main IO worker function. It retrieves io_u's to process and queues
669 * and reaps them, checking for rate and errors along the way.
671 * Returns number of bytes written and trimmed.
673 static uint64_t do_io(struct thread_data *td)
675 uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
678 uint64_t total_bytes, bytes_issued = 0;
680 if (in_ramp_time(td))
681 td_set_runstate(td, TD_RAMP);
683 td_set_runstate(td, TD_RUNNING);
688 * If verify_backlog is enabled, we'll run the verify in this
689 * handler as well. For that case, we may need up to twice the
692 total_bytes = td->o.size;
693 if (td->o.verify != VERIFY_NONE &&
694 (td_write(td) && td->o.verify_backlog))
695 total_bytes += td->o.size;
697 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
698 (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td) ||
700 struct timeval comp_time;
706 check_update_rusage(td);
708 if (td->terminate || td->done)
713 if (runtime_exceeded(td, &td->tv_cache)) {
714 __update_tv_cache(td);
715 if (runtime_exceeded(td, &td->tv_cache)) {
721 if (flow_threshold_exceeded(td))
724 if (bytes_issued >= total_bytes)
728 if (IS_ERR_OR_NULL(io_u)) {
729 int err = PTR_ERR(io_u);
736 if (td->o.latency_target)
744 * Add verification end_io handler if:
745 * - Asked to verify (!td_rw(td))
746 * - Or the io_u is from our verify list (mixed write/ver)
748 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
749 ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
751 if (!td->o.verify_pattern_bytes) {
752 io_u->rand_seed = __rand(&td->__verify_state);
753 if (sizeof(int) != sizeof(long *))
754 io_u->rand_seed *= __rand(&td->__verify_state);
757 if (td->o.verify_async)
758 io_u->end_io = verify_io_u_async;
760 io_u->end_io = verify_io_u;
761 td_set_runstate(td, TD_VERIFYING);
762 } else if (in_ramp_time(td))
763 td_set_runstate(td, TD_RAMP);
765 td_set_runstate(td, TD_RUNNING);
768 * Always log IO before it's issued, so we know the specific
769 * order of it. The logged unit will track when the IO has
772 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
774 td->o.verify != VERIFY_NONE &&
775 !td->o.experimental_verify)
776 log_io_piece(td, io_u);
778 ret = td_io_queue(td, io_u);
780 case FIO_Q_COMPLETED:
783 clear_io_u(td, io_u);
784 } else if (io_u->resid) {
785 int bytes = io_u->xfer_buflen - io_u->resid;
786 struct fio_file *f = io_u->file;
788 bytes_issued += bytes;
793 td_verror(td, EIO, "full resid");
798 io_u->xfer_buflen = io_u->resid;
799 io_u->xfer_buf += bytes;
800 io_u->offset += bytes;
802 if (ddir_rw(io_u->ddir))
803 td->ts.short_io_u[io_u->ddir]++;
805 if (io_u->offset == f->real_file_size)
808 requeue_io_u(td, &io_u);
811 if (__should_check_rate(td, DDIR_READ) ||
812 __should_check_rate(td, DDIR_WRITE) ||
813 __should_check_rate(td, DDIR_TRIM))
814 fio_gettime(&comp_time, NULL);
816 ret = io_u_sync_complete(td, io_u, bytes_done);
819 bytes_issued += io_u->xfer_buflen;
824 * if the engine doesn't have a commit hook,
825 * the io_u is really queued. if it does have such
826 * a hook, it has to call io_u_queued() itself.
828 if (td->io_ops->commit == NULL)
829 io_u_queued(td, io_u);
830 bytes_issued += io_u->xfer_buflen;
833 requeue_io_u(td, &io_u);
834 ret2 = td_io_commit(td);
844 if (break_on_this_error(td, ddir, &ret))
848 * See if we need to complete some commands. Note that we
849 * can get BUSY even without IO queued, if the system is
853 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
854 if (full || !td->o.iodepth_batch_complete) {
855 min_evts = min(td->o.iodepth_batch_complete,
858 * if the queue is full, we MUST reap at least 1 event
860 if (full && !min_evts)
863 if (__should_check_rate(td, DDIR_READ) ||
864 __should_check_rate(td, DDIR_WRITE) ||
865 __should_check_rate(td, DDIR_TRIM))
866 fio_gettime(&comp_time, NULL);
869 ret = io_u_queued_complete(td, min_evts, bytes_done);
873 } while (full && (td->cur_depth > td->o.iodepth_low));
878 if (!ddir_rw_sum(bytes_done) && !(td->io_ops->flags & FIO_NOIO))
881 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
882 if (check_min_rate(td, &comp_time, bytes_done)) {
883 if (exitall_on_terminate)
884 fio_terminate_threads(td->groupid);
885 td_verror(td, EIO, "check_min_rate");
889 if (!in_ramp_time(td) && td->o.latency_target)
890 lat_target_check(td);
892 if (td->o.thinktime) {
893 unsigned long long b;
895 b = ddir_rw_sum(td->io_blocks);
896 if (!(b % td->o.thinktime_blocks)) {
901 if (td->o.thinktime_spin)
902 usec_spin(td->o.thinktime_spin);
904 left = td->o.thinktime - td->o.thinktime_spin;
906 usec_sleep(td, left);
911 check_update_rusage(td);
913 if (td->trim_entries)
914 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
916 if (td->o.fill_device && td->error == ENOSPC) {
925 ret = io_u_queued_complete(td, i, bytes_done);
926 if (td->o.fill_device && td->error == ENOSPC)
930 if (should_fsync(td) && td->o.end_fsync) {
931 td_set_runstate(td, TD_FSYNCING);
933 for_each_file(td, f, i) {
934 if (!fio_file_fsync(td, f))
937 log_err("fio: end_fsync failed for file %s\n",
942 cleanup_pending_aio(td);
945 * stop job if we failed doing any IO
947 if (!ddir_rw_sum(td->this_io_bytes))
950 return bytes_done[DDIR_WRITE] + bytes_done[DDIR_TRIM];
953 static void cleanup_io_u(struct thread_data *td)
957 while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
959 if (td->io_ops->io_u_free)
960 td->io_ops->io_u_free(td, io_u);
962 fio_memfree(io_u, sizeof(*io_u));
967 io_u_rexit(&td->io_u_requeues);
968 io_u_qexit(&td->io_u_freelist);
969 io_u_qexit(&td->io_u_all);
972 static int init_io_u(struct thread_data *td)
975 unsigned int max_bs, min_write;
976 int cl_align, i, max_units;
977 int data_xfer = 1, err;
980 max_units = td->o.iodepth;
981 max_bs = td_max_bs(td);
982 min_write = td->o.min_bs[DDIR_WRITE];
983 td->orig_buffer_size = (unsigned long long) max_bs
984 * (unsigned long long) max_units;
986 if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
990 err += io_u_rinit(&td->io_u_requeues, td->o.iodepth);
991 err += io_u_qinit(&td->io_u_freelist, td->o.iodepth);
992 err += io_u_qinit(&td->io_u_all, td->o.iodepth);
995 log_err("fio: failed setting up IO queues\n");
1000 * if we may later need to do address alignment, then add any
1001 * possible adjustment here so that we don't cause a buffer
1002 * overflow later. this adjustment may be too much if we get
1003 * lucky and the allocator gives us an aligned address.
1005 if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1006 (td->io_ops->flags & FIO_RAWIO))
1007 td->orig_buffer_size += page_mask + td->o.mem_align;
1009 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1012 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1013 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1016 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1017 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1021 if (data_xfer && allocate_io_mem(td))
1024 if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1025 (td->io_ops->flags & FIO_RAWIO))
1026 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
1028 p = td->orig_buffer;
1030 cl_align = os_cache_line_size();
1032 for (i = 0; i < max_units; i++) {
1038 ptr = fio_memalign(cl_align, sizeof(*io_u));
1040 log_err("fio: unable to allocate aligned memory\n");
1045 memset(io_u, 0, sizeof(*io_u));
1046 INIT_FLIST_HEAD(&io_u->verify_list);
1047 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1051 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1054 io_u_fill_buffer(td, io_u, min_write, max_bs);
1055 if (td_write(td) && td->o.verify_pattern_bytes) {
1057 * Fill the buffer with the pattern if we are
1058 * going to be doing writes.
1060 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1065 io_u->flags = IO_U_F_FREE;
1066 io_u_qpush(&td->io_u_freelist, io_u);
1069 * io_u never leaves this stack, used for iteration of all
1072 io_u_qpush(&td->io_u_all, io_u);
1074 if (td->io_ops->io_u_init) {
1075 int ret = td->io_ops->io_u_init(td, io_u);
1078 log_err("fio: failed to init engine data: %d\n", ret);
1089 static int switch_ioscheduler(struct thread_data *td)
1091 char tmp[256], tmp2[128];
1095 if (td->io_ops->flags & FIO_DISKLESSIO)
1098 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
1100 f = fopen(tmp, "r+");
1102 if (errno == ENOENT) {
1103 log_err("fio: os or kernel doesn't support IO scheduler"
1107 td_verror(td, errno, "fopen iosched");
1114 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1115 if (ferror(f) || ret != 1) {
1116 td_verror(td, errno, "fwrite");
1124 * Read back and check that the selected scheduler is now the default.
1126 ret = fread(tmp, sizeof(tmp), 1, f);
1127 if (ferror(f) || ret < 0) {
1128 td_verror(td, errno, "fread");
1132 tmp[sizeof(tmp) - 1] = '\0';
1135 sprintf(tmp2, "[%s]", td->o.ioscheduler);
1136 if (!strstr(tmp, tmp2)) {
1137 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1138 td_verror(td, EINVAL, "iosched_switch");
1147 static int keep_running(struct thread_data *td)
1149 unsigned long long limit;
1153 if (td->o.time_based)
1159 if (exceeds_number_ios(td))
1163 limit = td->o.io_limit;
1167 if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1171 * If the difference is less than the minimum IO size, we
1174 diff = limit - ddir_rw_sum(td->io_bytes);
1175 if (diff < td_max_bs(td))
1178 if (fio_files_done(td))
1187 static int exec_string(struct thread_options *o, const char *string, const char *mode)
1189 int ret, newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
1192 str = malloc(newlen);
1193 sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
1195 log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1198 log_err("fio: exec of cmd <%s> failed\n", str);
1205 * Dry run to compute correct state of numberio for verification.
1207 static uint64_t do_dry_run(struct thread_data *td)
1209 uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
1211 td_set_runstate(td, TD_RUNNING);
1213 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1214 (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td)) {
1218 if (td->terminate || td->done)
1221 io_u = get_io_u(td);
1225 io_u->flags |= IO_U_F_FLIGHT;
1228 if (ddir_rw(acct_ddir(io_u)))
1229 td->io_issues[acct_ddir(io_u)]++;
1230 if (ddir_rw(io_u->ddir)) {
1231 io_u_mark_depth(td, 1);
1232 td->ts.total_io_u[io_u->ddir]++;
1235 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1237 td->o.verify != VERIFY_NONE &&
1238 !td->o.experimental_verify)
1239 log_io_piece(td, io_u);
1241 ret = io_u_sync_complete(td, io_u, bytes_done);
1245 return bytes_done[DDIR_WRITE] + bytes_done[DDIR_TRIM];
1249 * Entry point for the thread based jobs. The process based jobs end up
1250 * here as well, after a little setup.
1252 static void *thread_main(void *data)
1254 unsigned long long elapsed;
1255 struct thread_data *td = data;
1256 struct thread_options *o = &td->o;
1257 pthread_condattr_t attr;
1261 if (!o->use_thread) {
1268 * fio_time_init() may not have been called yet if running as a server
1272 fio_local_clock_init(o->use_thread);
1274 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1277 fio_server_send_start(td);
1279 INIT_FLIST_HEAD(&td->io_log_list);
1280 INIT_FLIST_HEAD(&td->io_hist_list);
1281 INIT_FLIST_HEAD(&td->verify_list);
1282 INIT_FLIST_HEAD(&td->trim_list);
1283 INIT_FLIST_HEAD(&td->next_rand_list);
1284 pthread_mutex_init(&td->io_u_lock, NULL);
1285 td->io_hist_tree = RB_ROOT;
1287 pthread_condattr_init(&attr);
1288 pthread_cond_init(&td->verify_cond, &attr);
1289 pthread_cond_init(&td->free_cond, &attr);
1291 td_set_runstate(td, TD_INITIALIZED);
1292 dprint(FD_MUTEX, "up startup_mutex\n");
1293 fio_mutex_up(startup_mutex);
1294 dprint(FD_MUTEX, "wait on td->mutex\n");
1295 fio_mutex_down(td->mutex);
1296 dprint(FD_MUTEX, "done waiting on td->mutex\n");
1299 * A new gid requires privilege, so we need to do this before setting
1302 if (o->gid != -1U && setgid(o->gid)) {
1303 td_verror(td, errno, "setgid");
1306 if (o->uid != -1U && setuid(o->uid)) {
1307 td_verror(td, errno, "setuid");
1312 * If we have a gettimeofday() thread, make sure we exclude that
1313 * thread from this job
1316 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1319 * Set affinity first, in case it has an impact on the memory
1322 if (o->cpumask_set) {
1323 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1324 ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1326 log_err("fio: no CPUs set\n");
1327 log_err("fio: Try increasing number of available CPUs\n");
1328 td_verror(td, EINVAL, "cpus_split");
1332 ret = fio_setaffinity(td->pid, o->cpumask);
1334 td_verror(td, errno, "cpu_set_affinity");
1339 #ifdef CONFIG_LIBNUMA
1340 /* numa node setup */
1341 if (o->numa_cpumask_set || o->numa_memmask_set) {
1344 if (numa_available() < 0) {
1345 td_verror(td, errno, "Does not support NUMA API\n");
1349 if (o->numa_cpumask_set) {
1350 ret = numa_run_on_node_mask(o->numa_cpunodesmask);
1352 td_verror(td, errno, \
1353 "numa_run_on_node_mask failed\n");
1358 if (o->numa_memmask_set) {
1360 switch (o->numa_mem_mode) {
1361 case MPOL_INTERLEAVE:
1362 numa_set_interleave_mask(o->numa_memnodesmask);
1365 numa_set_membind(o->numa_memnodesmask);
1368 numa_set_localalloc();
1370 case MPOL_PREFERRED:
1371 numa_set_preferred(o->numa_mem_prefer_node);
1382 if (fio_pin_memory(td))
1386 * May alter parameters that init_io_u() will use, so we need to
1395 if (o->verify_async && verify_async_init(td))
1399 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1401 td_verror(td, errno, "ioprio_set");
1406 if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1410 if (nice(o->nice) == -1 && errno != 0) {
1411 td_verror(td, errno, "nice");
1415 if (o->ioscheduler && switch_ioscheduler(td))
1418 if (!o->create_serialize && setup_files(td))
1424 if (init_random_map(td))
1427 if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
1431 if (pre_read_files(td) < 0)
1435 fio_verify_init(td);
1437 fio_gettime(&td->epoch, NULL);
1438 fio_getrusage(&td->ru_start);
1440 while (keep_running(td)) {
1441 uint64_t verify_bytes;
1443 fio_gettime(&td->start, NULL);
1444 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1445 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1446 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1448 if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1449 o->ratemin[DDIR_TRIM]) {
1450 memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1451 sizeof(td->bw_sample_time));
1452 memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1453 sizeof(td->bw_sample_time));
1454 memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1455 sizeof(td->bw_sample_time));
1461 prune_io_piece_log(td);
1463 if (td->o.verify_only && (td_write(td) || td_rw(td)))
1464 verify_bytes = do_dry_run(td);
1466 verify_bytes = do_io(td);
1470 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1471 elapsed = utime_since_now(&td->start);
1472 td->ts.runtime[DDIR_READ] += elapsed;
1474 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1475 elapsed = utime_since_now(&td->start);
1476 td->ts.runtime[DDIR_WRITE] += elapsed;
1478 if (td_trim(td) && td->io_bytes[DDIR_TRIM]) {
1479 elapsed = utime_since_now(&td->start);
1480 td->ts.runtime[DDIR_TRIM] += elapsed;
1483 if (td->error || td->terminate)
1486 if (!o->do_verify ||
1487 o->verify == VERIFY_NONE ||
1488 (td->io_ops->flags & FIO_UNIDIR))
1493 fio_gettime(&td->start, NULL);
1495 do_verify(td, verify_bytes);
1497 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1499 if (td->error || td->terminate)
1503 update_rusage_stat(td);
1504 td->ts.runtime[DDIR_READ] = (td->ts.runtime[DDIR_READ] + 999) / 1000;
1505 td->ts.runtime[DDIR_WRITE] = (td->ts.runtime[DDIR_WRITE] + 999) / 1000;
1506 td->ts.runtime[DDIR_TRIM] = (td->ts.runtime[DDIR_TRIM] + 999) / 1000;
1507 td->ts.total_run_time = mtime_since_now(&td->epoch);
1508 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1509 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1510 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1512 fio_unpin_memory(td);
1514 fio_writeout_logs(td);
1516 if (o->exec_postrun)
1517 exec_string(o, o->exec_postrun, (const char *)"postrun");
1519 if (exitall_on_terminate)
1520 fio_terminate_threads(td->groupid);
1524 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1527 if (o->verify_async)
1528 verify_async_exit(td);
1530 close_and_free_files(td);
1533 cgroup_shutdown(td, &cgroup_mnt);
1535 if (o->cpumask_set) {
1536 int ret = fio_cpuset_exit(&o->cpumask);
1538 td_verror(td, ret, "fio_cpuset_exit");
1542 * do this very late, it will log file closing as well
1544 if (o->write_iolog_file)
1545 write_iolog_close(td);
1547 fio_mutex_remove(td->rusage_sem);
1548 td->rusage_sem = NULL;
1550 fio_mutex_remove(td->mutex);
1553 td_set_runstate(td, TD_EXITED);
1554 return (void *) (uintptr_t) td->error;
1559 * We cannot pass the td data into a forked process, so attach the td and
1560 * pass it to the thread worker.
1562 static int fork_main(int shmid, int offset)
1564 struct thread_data *td;
1568 data = shmat(shmid, NULL, 0);
1569 if (data == (void *) -1) {
1577 * HP-UX inherits shm mappings?
1582 td = data + offset * sizeof(struct thread_data);
1583 ret = thread_main(td);
1585 return (int) (uintptr_t) ret;
1589 * Run over the job map and reap the threads that have exited, if any.
1591 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1592 unsigned int *m_rate)
1594 struct thread_data *td;
1595 unsigned int cputhreads, realthreads, pending;
1599 * reap exited threads (TD_EXITED -> TD_REAPED)
1601 realthreads = pending = cputhreads = 0;
1602 for_each_td(td, i) {
1606 * ->io_ops is NULL for a thread that has closed its
1609 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1618 if (td->runstate == TD_REAPED)
1620 if (td->o.use_thread) {
1621 if (td->runstate == TD_EXITED) {
1622 td_set_runstate(td, TD_REAPED);
1629 if (td->runstate == TD_EXITED)
1633 * check if someone quit or got killed in an unusual way
1635 ret = waitpid(td->pid, &status, flags);
1637 if (errno == ECHILD) {
1638 log_err("fio: pid=%d disappeared %d\n",
1639 (int) td->pid, td->runstate);
1641 td_set_runstate(td, TD_REAPED);
1645 } else if (ret == td->pid) {
1646 if (WIFSIGNALED(status)) {
1647 int sig = WTERMSIG(status);
1649 if (sig != SIGTERM && sig != SIGUSR2)
1650 log_err("fio: pid=%d, got signal=%d\n",
1651 (int) td->pid, sig);
1653 td_set_runstate(td, TD_REAPED);
1656 if (WIFEXITED(status)) {
1657 if (WEXITSTATUS(status) && !td->error)
1658 td->error = WEXITSTATUS(status);
1660 td_set_runstate(td, TD_REAPED);
1666 * thread is not dead, continue
1672 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1673 (*t_rate) -= ddir_rw_sum(td->o.rate);
1680 done_secs += mtime_since_now(&td->epoch) / 1000;
1681 profile_td_exit(td);
1684 if (*nr_running == cputhreads && !pending && realthreads)
1685 fio_terminate_threads(TERMINATE_ALL);
1688 static void do_usleep(unsigned int usecs)
1690 check_for_running_stats();
1695 * Main function for kicking off and reaping jobs, as needed.
1697 static void run_threads(void)
1699 struct thread_data *td;
1700 unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
1703 if (fio_gtod_offload && fio_start_gtod_thread())
1706 fio_idle_prof_init();
1710 nr_thread = nr_process = 0;
1711 for_each_td(td, i) {
1712 if (td->o.use_thread)
1718 if (output_format == FIO_OUTPUT_NORMAL) {
1719 log_info("Starting ");
1721 log_info("%d thread%s", nr_thread,
1722 nr_thread > 1 ? "s" : "");
1726 log_info("%d process%s", nr_process,
1727 nr_process > 1 ? "es" : "");
1733 todo = thread_number;
1736 m_rate = t_rate = 0;
1738 for_each_td(td, i) {
1739 print_status_init(td->thread_number - 1);
1741 if (!td->o.create_serialize)
1745 * do file setup here so it happens sequentially,
1746 * we don't want X number of threads getting their
1747 * client data interspersed on disk
1749 if (setup_files(td)) {
1752 log_err("fio: pid=%d, err=%d/%s\n",
1753 (int) td->pid, td->error, td->verror);
1754 td_set_runstate(td, TD_REAPED);
1761 * for sharing to work, each job must always open
1762 * its own files. so close them, if we opened them
1765 for_each_file(td, f, j) {
1766 if (fio_file_open(f))
1767 td_io_close_file(td, f);
1772 /* start idle threads before io threads start to run */
1773 fio_idle_prof_start();
1778 struct thread_data *map[REAL_MAX_JOBS];
1779 struct timeval this_start;
1780 int this_jobs = 0, left;
1783 * create threads (TD_NOT_CREATED -> TD_CREATED)
1785 for_each_td(td, i) {
1786 if (td->runstate != TD_NOT_CREATED)
1790 * never got a chance to start, killed by other
1791 * thread for some reason
1793 if (td->terminate) {
1798 if (td->o.start_delay) {
1799 spent = utime_since_genesis();
1801 if (td->o.start_delay > spent)
1805 if (td->o.stonewall && (nr_started || nr_running)) {
1806 dprint(FD_PROCESS, "%s: stonewall wait\n",
1813 td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED);
1814 td->update_rusage = 0;
1817 * Set state to created. Thread will transition
1818 * to TD_INITIALIZED when it's done setting up.
1820 td_set_runstate(td, TD_CREATED);
1821 map[this_jobs++] = td;
1824 if (td->o.use_thread) {
1827 dprint(FD_PROCESS, "will pthread_create\n");
1828 ret = pthread_create(&td->thread, NULL,
1831 log_err("pthread_create: %s\n",
1836 ret = pthread_detach(td->thread);
1838 log_err("pthread_detach: %s",
1842 dprint(FD_PROCESS, "will fork\n");
1845 int ret = fork_main(shm_id, i);
1848 } else if (i == fio_debug_jobno)
1849 *fio_debug_jobp = pid;
1851 dprint(FD_MUTEX, "wait on startup_mutex\n");
1852 if (fio_mutex_down_timeout(startup_mutex, 10)) {
1853 log_err("fio: job startup hung? exiting.\n");
1854 fio_terminate_threads(TERMINATE_ALL);
1859 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1863 * Wait for the started threads to transition to
1866 fio_gettime(&this_start, NULL);
1868 while (left && !fio_abort) {
1869 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1874 for (i = 0; i < this_jobs; i++) {
1878 if (td->runstate == TD_INITIALIZED) {
1881 } else if (td->runstate >= TD_EXITED) {
1885 nr_running++; /* work-around... */
1891 log_err("fio: %d job%s failed to start\n", left,
1892 left > 1 ? "s" : "");
1893 for (i = 0; i < this_jobs; i++) {
1897 kill(td->pid, SIGTERM);
1903 * start created threads (TD_INITIALIZED -> TD_RUNNING).
1905 for_each_td(td, i) {
1906 if (td->runstate != TD_INITIALIZED)
1909 if (in_ramp_time(td))
1910 td_set_runstate(td, TD_RAMP);
1912 td_set_runstate(td, TD_RUNNING);
1915 m_rate += ddir_rw_sum(td->o.ratemin);
1916 t_rate += ddir_rw_sum(td->o.rate);
1918 fio_mutex_up(td->mutex);
1921 reap_threads(&nr_running, &t_rate, &m_rate);
1927 while (nr_running) {
1928 reap_threads(&nr_running, &t_rate, &m_rate);
1932 fio_idle_prof_stop();
1937 void wait_for_disk_thread_exit(void)
1939 fio_mutex_down(disk_thread_mutex);
1942 static void free_disk_util(void)
1944 disk_util_start_exit();
1945 wait_for_disk_thread_exit();
1946 disk_util_prune_entries();
1949 static void *disk_thread_main(void *data)
1953 fio_mutex_up(startup_mutex);
1955 while (threads && !ret) {
1956 usleep(DISK_UTIL_MSEC * 1000);
1959 ret = update_io_ticks();
1962 print_thread_status();
1965 fio_mutex_up(disk_thread_mutex);
1969 static int create_disk_util_thread(void)
1975 disk_thread_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
1977 ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
1979 fio_mutex_remove(disk_thread_mutex);
1980 log_err("Can't create disk util thread: %s\n", strerror(ret));
1984 ret = pthread_detach(disk_util_thread);
1986 fio_mutex_remove(disk_thread_mutex);
1987 log_err("Can't detatch disk util thread: %s\n", strerror(ret));
1991 dprint(FD_MUTEX, "wait on startup_mutex\n");
1992 fio_mutex_down(startup_mutex);
1993 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1997 int fio_backend(void)
1999 struct thread_data *td;
2003 if (load_profile(exec_profile))
2006 exec_profile = NULL;
2012 setup_log(&agg_io_log[DDIR_READ], 0, IO_LOG_TYPE_BW);
2013 setup_log(&agg_io_log[DDIR_WRITE], 0, IO_LOG_TYPE_BW);
2014 setup_log(&agg_io_log[DDIR_TRIM], 0, IO_LOG_TYPE_BW);
2017 startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
2018 if (startup_mutex == NULL)
2023 create_disk_util_thread();
2025 cgroup_list = smalloc(sizeof(*cgroup_list));
2026 INIT_FLIST_HEAD(cgroup_list);
2033 __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
2034 __finish_log(agg_io_log[DDIR_WRITE],
2035 "agg-write_bw.log");
2036 __finish_log(agg_io_log[DDIR_TRIM],
2037 "agg-write_bw.log");
2042 fio_options_free(td);
2045 cgroup_kill(cgroup_list);
2049 fio_mutex_remove(startup_mutex);
2050 fio_mutex_remove(disk_thread_mutex);