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"
56 static pthread_t disk_util_thread;
57 static struct fio_mutex *disk_thread_mutex;
58 static struct fio_mutex *startup_mutex;
59 static struct fio_mutex *writeout_mutex;
60 static struct flist_head *cgroup_list;
61 static char *cgroup_mnt;
62 static int exit_value;
63 static volatile int fio_abort;
65 struct io_log *agg_io_log[DDIR_RWDIR_CNT];
68 unsigned int thread_number = 0;
69 unsigned int stat_number = 0;
70 unsigned int nr_process = 0;
71 unsigned int nr_thread = 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,
185 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
186 if (rate < ratemin ||
187 bytes < td->rate_bytes[ddir]) {
188 log_err("%s: min rate %u not met, got"
189 " %luKB/sec\n", td->o.name,
196 * checks iops specified rate
198 if (iops < rate_iops) {
199 log_err("%s: min iops rate %u not met\n",
200 td->o.name, rate_iops);
203 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
204 if (rate < rate_iops_min ||
205 iops < td->rate_blocks[ddir]) {
206 log_err("%s: min iops rate %u not met,"
207 " got %lu\n", td->o.name,
208 rate_iops_min, rate);
214 td->rate_bytes[ddir] = bytes;
215 td->rate_blocks[ddir] = iops;
216 memcpy(&td->lastrate[ddir], now, sizeof(*now));
220 static int check_min_rate(struct thread_data *td, struct timeval *now,
221 uint64_t *bytes_done)
225 if (bytes_done[DDIR_READ])
226 ret |= __check_min_rate(td, now, DDIR_READ);
227 if (bytes_done[DDIR_WRITE])
228 ret |= __check_min_rate(td, now, DDIR_WRITE);
229 if (bytes_done[DDIR_TRIM])
230 ret |= __check_min_rate(td, now, DDIR_TRIM);
236 * When job exits, we can cancel the in-flight IO if we are using async
237 * io. Attempt to do so.
239 static void cleanup_pending_aio(struct thread_data *td)
241 struct flist_head *entry, *n;
246 * get immediately available events, if any
248 r = io_u_queued_complete(td, 0, NULL);
253 * now cancel remaining active events
255 if (td->io_ops->cancel) {
256 flist_for_each_safe(entry, n, &td->io_u_busylist) {
257 io_u = flist_entry(entry, struct io_u, list);
260 * if the io_u isn't in flight, then that generally
261 * means someone leaked an io_u. complain but fix
262 * it up, so we don't stall here.
264 if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
265 log_err("fio: non-busy IO on busy list\n");
268 r = td->io_ops->cancel(td, io_u);
276 r = io_u_queued_complete(td, td->cur_depth, NULL);
280 * Helper to handle the final sync of a file. Works just like the normal
281 * io path, just does everything sync.
283 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
285 struct io_u *io_u = __get_io_u(td);
291 io_u->ddir = DDIR_SYNC;
294 if (td_io_prep(td, io_u)) {
300 ret = td_io_queue(td, io_u);
302 td_verror(td, io_u->error, "td_io_queue");
305 } else if (ret == FIO_Q_QUEUED) {
306 if (io_u_queued_complete(td, 1, NULL) < 0)
308 } else if (ret == FIO_Q_COMPLETED) {
310 td_verror(td, io_u->error, "td_io_queue");
314 if (io_u_sync_complete(td, io_u, NULL) < 0)
316 } else if (ret == FIO_Q_BUSY) {
317 if (td_io_commit(td))
325 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
329 if (fio_file_open(f))
330 return fio_io_sync(td, f);
332 if (td_io_open_file(td, f))
335 ret = fio_io_sync(td, f);
336 td_io_close_file(td, f);
340 static inline void __update_tv_cache(struct thread_data *td)
342 fio_gettime(&td->tv_cache, NULL);
345 static inline void update_tv_cache(struct thread_data *td)
347 if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
348 __update_tv_cache(td);
351 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
353 if (in_ramp_time(td))
357 if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
363 static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
368 if (ret < 0 || td->error) {
370 enum error_type_bit eb;
375 eb = td_error_type(ddir, err);
376 if (!(td->o.continue_on_error & (1 << eb)))
379 if (td_non_fatal_error(td, eb, err)) {
381 * Continue with the I/Os in case of
384 update_error_count(td, err);
388 } else if (td->o.fill_device && err == ENOSPC) {
390 * We expect to hit this error if
391 * fill_device option is set.
398 * Stop the I/O in case of a fatal
401 update_error_count(td, err);
410 * The main verify engine. Runs over the writes we previously submitted,
411 * reads the blocks back in, and checks the crc/md5 of the data.
413 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
415 uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
421 dprint(FD_VERIFY, "starting loop\n");
424 * sync io first and invalidate cache, to make sure we really
427 for_each_file(td, f, i) {
428 if (!fio_file_open(f))
430 if (fio_io_sync(td, f))
432 if (file_invalidate_cache(td, f))
439 td_set_runstate(td, TD_VERIFYING);
442 while (!td->terminate) {
448 if (runtime_exceeded(td, &td->tv_cache)) {
449 __update_tv_cache(td);
450 if (runtime_exceeded(td, &td->tv_cache)) {
456 if (flow_threshold_exceeded(td))
459 if (!td->o.experimental_verify) {
460 io_u = __get_io_u(td);
464 if (get_next_verify(td, io_u)) {
469 if (td_io_prep(td, io_u)) {
474 if (ddir_rw_sum(bytes_done) + td->o.rw_min_bs > verify_bytes)
477 while ((io_u = get_io_u(td)) != NULL) {
479 * We are only interested in the places where
480 * we wrote or trimmed IOs. Turn those into
481 * reads for verification purposes.
483 if (io_u->ddir == DDIR_READ) {
485 * Pretend we issued it for rwmix
488 td->io_issues[DDIR_READ]++;
491 } else if (io_u->ddir == DDIR_TRIM) {
492 io_u->ddir = DDIR_READ;
493 io_u->flags |= IO_U_F_TRIMMED;
495 } else if (io_u->ddir == DDIR_WRITE) {
496 io_u->ddir = DDIR_READ;
508 if (td->o.verify_async)
509 io_u->end_io = verify_io_u_async;
511 io_u->end_io = verify_io_u;
515 ret = td_io_queue(td, io_u);
517 case FIO_Q_COMPLETED:
520 clear_io_u(td, io_u);
521 } else if (io_u->resid) {
522 int bytes = io_u->xfer_buflen - io_u->resid;
528 td_verror(td, EIO, "full resid");
533 io_u->xfer_buflen = io_u->resid;
534 io_u->xfer_buf += bytes;
535 io_u->offset += bytes;
537 if (ddir_rw(io_u->ddir))
538 td->ts.short_io_u[io_u->ddir]++;
541 if (io_u->offset == f->real_file_size)
544 requeue_io_u(td, &io_u);
547 ret = io_u_sync_complete(td, io_u, bytes_done);
555 requeue_io_u(td, &io_u);
556 ret2 = td_io_commit(td);
562 td_verror(td, -ret, "td_io_queue");
566 if (break_on_this_error(td, ddir, &ret))
570 * if we can queue more, do so. but check if there are
571 * completed io_u's first. Note that we can get BUSY even
572 * without IO queued, if the system is resource starved.
574 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
575 if (full || !td->o.iodepth_batch_complete) {
576 min_events = min(td->o.iodepth_batch_complete,
579 * if the queue is full, we MUST reap at least 1 event
581 if (full && !min_events)
586 * Reap required number of io units, if any,
587 * and do the verification on them through
588 * the callback handler
590 if (io_u_queued_complete(td, min_events, bytes_done) < 0) {
594 } while (full && (td->cur_depth > td->o.iodepth_low));
601 min_events = td->cur_depth;
604 ret = io_u_queued_complete(td, min_events, NULL);
606 cleanup_pending_aio(td);
608 td_set_runstate(td, TD_RUNNING);
610 dprint(FD_VERIFY, "exiting loop\n");
613 static int io_bytes_exceeded(struct thread_data *td)
615 unsigned long long bytes;
618 bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE];
619 else if (td_write(td))
620 bytes = td->this_io_bytes[DDIR_WRITE];
621 else if (td_read(td))
622 bytes = td->this_io_bytes[DDIR_READ];
624 bytes = td->this_io_bytes[DDIR_TRIM];
626 return bytes >= td->o.size;
630 * Main IO worker function. It retrieves io_u's to process and queues
631 * and reaps them, checking for rate and errors along the way.
633 * Returns number of bytes written and trimmed.
635 static uint64_t do_io(struct thread_data *td)
637 uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
641 if (in_ramp_time(td))
642 td_set_runstate(td, TD_RAMP);
644 td_set_runstate(td, TD_RUNNING);
646 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
647 (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td) ||
649 struct timeval comp_time;
655 if (td->terminate || td->done)
660 if (runtime_exceeded(td, &td->tv_cache)) {
661 __update_tv_cache(td);
662 if (runtime_exceeded(td, &td->tv_cache)) {
668 if (flow_threshold_exceeded(td))
678 * Add verification end_io handler if:
679 * - Asked to verify (!td_rw(td))
680 * - Or the io_u is from our verify list (mixed write/ver)
682 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
683 ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
684 if (td->o.verify_async)
685 io_u->end_io = verify_io_u_async;
687 io_u->end_io = verify_io_u;
688 td_set_runstate(td, TD_VERIFYING);
689 } else if (in_ramp_time(td))
690 td_set_runstate(td, TD_RAMP);
692 td_set_runstate(td, TD_RUNNING);
694 ret = td_io_queue(td, io_u);
696 case FIO_Q_COMPLETED:
699 clear_io_u(td, io_u);
700 } else if (io_u->resid) {
701 int bytes = io_u->xfer_buflen - io_u->resid;
702 struct fio_file *f = io_u->file;
708 td_verror(td, EIO, "full resid");
713 io_u->xfer_buflen = io_u->resid;
714 io_u->xfer_buf += bytes;
715 io_u->offset += bytes;
717 if (ddir_rw(io_u->ddir))
718 td->ts.short_io_u[io_u->ddir]++;
720 if (io_u->offset == f->real_file_size)
723 requeue_io_u(td, &io_u);
726 if (__should_check_rate(td, DDIR_READ) ||
727 __should_check_rate(td, DDIR_WRITE) ||
728 __should_check_rate(td, DDIR_TRIM))
729 fio_gettime(&comp_time, NULL);
731 ret = io_u_sync_complete(td, io_u, bytes_done);
738 * if the engine doesn't have a commit hook,
739 * the io_u is really queued. if it does have such
740 * a hook, it has to call io_u_queued() itself.
742 if (td->io_ops->commit == NULL)
743 io_u_queued(td, io_u);
746 requeue_io_u(td, &io_u);
747 ret2 = td_io_commit(td);
757 if (break_on_this_error(td, ddir, &ret))
761 * See if we need to complete some commands. Note that we
762 * can get BUSY even without IO queued, if the system is
765 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
766 if (full || !td->o.iodepth_batch_complete) {
767 min_evts = min(td->o.iodepth_batch_complete,
770 * if the queue is full, we MUST reap at least 1 event
772 if (full && !min_evts)
775 if (__should_check_rate(td, DDIR_READ) ||
776 __should_check_rate(td, DDIR_WRITE) ||
777 __should_check_rate(td, DDIR_TRIM))
778 fio_gettime(&comp_time, NULL);
781 ret = io_u_queued_complete(td, min_evts, bytes_done);
785 } while (full && (td->cur_depth > td->o.iodepth_low));
790 if (!ddir_rw_sum(bytes_done) && !(td->io_ops->flags & FIO_NOIO))
793 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
794 if (check_min_rate(td, &comp_time, bytes_done)) {
795 if (exitall_on_terminate)
796 fio_terminate_threads(td->groupid);
797 td_verror(td, EIO, "check_min_rate");
802 if (td->o.thinktime) {
803 unsigned long long b;
805 b = ddir_rw_sum(td->io_blocks);
806 if (!(b % td->o.thinktime_blocks)) {
809 if (td->o.thinktime_spin)
810 usec_spin(td->o.thinktime_spin);
812 left = td->o.thinktime - td->o.thinktime_spin;
814 usec_sleep(td, left);
819 if (td->trim_entries)
820 log_err("fio: %d trim entries leaked?\n", td->trim_entries);
822 if (td->o.fill_device && td->error == ENOSPC) {
831 ret = io_u_queued_complete(td, i, bytes_done);
832 if (td->o.fill_device && td->error == ENOSPC)
836 if (should_fsync(td) && td->o.end_fsync) {
837 td_set_runstate(td, TD_FSYNCING);
839 for_each_file(td, f, i) {
840 if (!fio_file_fsync(td, f))
843 log_err("fio: end_fsync failed for file %s\n",
848 cleanup_pending_aio(td);
851 * stop job if we failed doing any IO
853 if (!ddir_rw_sum(td->this_io_bytes))
856 return bytes_done[DDIR_WRITE] + bytes_done[DDIR_TRIM];
859 static void cleanup_io_u(struct thread_data *td)
861 struct flist_head *entry, *n;
864 flist_for_each_safe(entry, n, &td->io_u_freelist) {
865 io_u = flist_entry(entry, struct io_u, list);
867 flist_del(&io_u->list);
869 if (td->io_ops->io_u_free)
870 td->io_ops->io_u_free(td, io_u);
872 fio_memfree(io_u, sizeof(*io_u));
878 static int init_io_u(struct thread_data *td)
881 unsigned int max_bs, min_write;
882 int cl_align, i, max_units;
886 max_units = td->o.iodepth;
887 max_bs = td_max_bs(td);
888 min_write = td->o.min_bs[DDIR_WRITE];
889 td->orig_buffer_size = (unsigned long long) max_bs
890 * (unsigned long long) max_units;
892 if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
895 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
898 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
899 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
902 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
903 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
907 if (data_xfer && allocate_io_mem(td))
910 if (td->o.odirect || td->o.mem_align ||
911 (td->io_ops->flags & FIO_RAWIO))
912 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
916 cl_align = os_cache_line_size();
918 for (i = 0; i < max_units; i++) {
924 ptr = fio_memalign(cl_align, sizeof(*io_u));
926 log_err("fio: unable to allocate aligned memory\n");
931 memset(io_u, 0, sizeof(*io_u));
932 INIT_FLIST_HEAD(&io_u->list);
933 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
937 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
940 io_u_fill_buffer(td, io_u, min_write, max_bs);
941 if (td_write(td) && td->o.verify_pattern_bytes) {
943 * Fill the buffer with the pattern if we are
944 * going to be doing writes.
946 fill_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
951 io_u->flags = IO_U_F_FREE;
952 flist_add(&io_u->list, &td->io_u_freelist);
954 if (td->io_ops->io_u_init) {
955 int ret = td->io_ops->io_u_init(td, io_u);
958 log_err("fio: failed to init engine data: %d\n", ret);
969 static int switch_ioscheduler(struct thread_data *td)
971 char tmp[256], tmp2[128];
975 if (td->io_ops->flags & FIO_DISKLESSIO)
978 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
980 f = fopen(tmp, "r+");
982 if (errno == ENOENT) {
983 log_err("fio: os or kernel doesn't support IO scheduler"
987 td_verror(td, errno, "fopen iosched");
994 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
995 if (ferror(f) || ret != 1) {
996 td_verror(td, errno, "fwrite");
1004 * Read back and check that the selected scheduler is now the default.
1006 ret = fread(tmp, 1, sizeof(tmp), f);
1007 if (ferror(f) || ret < 0) {
1008 td_verror(td, errno, "fread");
1013 sprintf(tmp2, "[%s]", td->o.ioscheduler);
1014 if (!strstr(tmp, tmp2)) {
1015 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1016 td_verror(td, EINVAL, "iosched_switch");
1025 static int keep_running(struct thread_data *td)
1029 if (td->o.time_based)
1036 if (td->o.size != -1ULL && ddir_rw_sum(td->io_bytes) < td->o.size) {
1040 * If the difference is less than the minimum IO size, we
1043 diff = td->o.size - ddir_rw_sum(td->io_bytes);
1044 if (diff < td_max_bs(td))
1053 static int exec_string(const char *string)
1055 int ret, newlen = strlen(string) + 1 + 8;
1058 str = malloc(newlen);
1059 sprintf(str, "sh -c %s", string);
1063 log_err("fio: exec of cmd <%s> failed\n", str);
1070 * Entry point for the thread based jobs. The process based jobs end up
1071 * here as well, after a little setup.
1073 static void *thread_main(void *data)
1075 unsigned long long elapsed;
1076 struct thread_data *td = data;
1077 pthread_condattr_t attr;
1080 if (!td->o.use_thread) {
1086 fio_local_clock_init(td->o.use_thread);
1088 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1090 INIT_FLIST_HEAD(&td->io_u_freelist);
1091 INIT_FLIST_HEAD(&td->io_u_busylist);
1092 INIT_FLIST_HEAD(&td->io_u_requeues);
1093 INIT_FLIST_HEAD(&td->io_log_list);
1094 INIT_FLIST_HEAD(&td->io_hist_list);
1095 INIT_FLIST_HEAD(&td->verify_list);
1096 INIT_FLIST_HEAD(&td->trim_list);
1097 INIT_FLIST_HEAD(&td->next_rand_list);
1098 pthread_mutex_init(&td->io_u_lock, NULL);
1099 td->io_hist_tree = RB_ROOT;
1101 pthread_condattr_init(&attr);
1102 pthread_cond_init(&td->verify_cond, &attr);
1103 pthread_cond_init(&td->free_cond, &attr);
1105 td_set_runstate(td, TD_INITIALIZED);
1106 dprint(FD_MUTEX, "up startup_mutex\n");
1107 fio_mutex_up(startup_mutex);
1108 dprint(FD_MUTEX, "wait on td->mutex\n");
1109 fio_mutex_down(td->mutex);
1110 dprint(FD_MUTEX, "done waiting on td->mutex\n");
1113 * the ->mutex mutex is now no longer used, close it to avoid
1114 * eating a file descriptor
1116 fio_mutex_remove(td->mutex);
1119 * A new gid requires privilege, so we need to do this before setting
1122 if (td->o.gid != -1U && setgid(td->o.gid)) {
1123 td_verror(td, errno, "setgid");
1126 if (td->o.uid != -1U && setuid(td->o.uid)) {
1127 td_verror(td, errno, "setuid");
1132 * If we have a gettimeofday() thread, make sure we exclude that
1133 * thread from this job
1136 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1139 * Set affinity first, in case it has an impact on the memory
1142 if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1143 td_verror(td, errno, "cpu_set_affinity");
1147 #ifdef CONFIG_LIBNUMA
1148 /* numa node setup */
1149 if (td->o.numa_cpumask_set || td->o.numa_memmask_set) {
1152 if (numa_available() < 0) {
1153 td_verror(td, errno, "Does not support NUMA API\n");
1157 if (td->o.numa_cpumask_set) {
1158 ret = numa_run_on_node_mask(td->o.numa_cpunodesmask);
1160 td_verror(td, errno, \
1161 "numa_run_on_node_mask failed\n");
1166 if (td->o.numa_memmask_set) {
1168 switch (td->o.numa_mem_mode) {
1169 case MPOL_INTERLEAVE:
1170 numa_set_interleave_mask(td->o.numa_memnodesmask);
1173 numa_set_membind(td->o.numa_memnodesmask);
1176 numa_set_localalloc();
1178 case MPOL_PREFERRED:
1179 numa_set_preferred(td->o.numa_mem_prefer_node);
1191 * May alter parameters that init_io_u() will use, so we need to
1200 if (td->o.verify_async && verify_async_init(td))
1203 if (td->ioprio_set) {
1204 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1205 td_verror(td, errno, "ioprio_set");
1210 if (td->o.cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1214 if (nice(td->o.nice) == -1 && errno != 0) {
1215 td_verror(td, errno, "nice");
1219 if (td->o.ioscheduler && switch_ioscheduler(td))
1222 if (!td->o.create_serialize && setup_files(td))
1228 if (init_random_map(td))
1231 if (td->o.exec_prerun) {
1232 if (exec_string(td->o.exec_prerun))
1236 if (td->o.pre_read) {
1237 if (pre_read_files(td) < 0)
1241 fio_gettime(&td->epoch, NULL);
1242 fio_getrusage(&td->ru_start);
1244 while (keep_running(td)) {
1245 uint64_t verify_bytes;
1247 fio_gettime(&td->start, NULL);
1248 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1249 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1250 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1252 if (td->o.ratemin[DDIR_READ] || td->o.ratemin[DDIR_WRITE] ||
1253 td->o.ratemin[DDIR_TRIM]) {
1254 memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1255 sizeof(td->bw_sample_time));
1256 memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1257 sizeof(td->bw_sample_time));
1258 memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1259 sizeof(td->bw_sample_time));
1265 prune_io_piece_log(td);
1267 verify_bytes = do_io(td);
1271 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1272 elapsed = utime_since_now(&td->start);
1273 td->ts.runtime[DDIR_READ] += elapsed;
1275 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1276 elapsed = utime_since_now(&td->start);
1277 td->ts.runtime[DDIR_WRITE] += elapsed;
1279 if (td_trim(td) && td->io_bytes[DDIR_TRIM]) {
1280 elapsed = utime_since_now(&td->start);
1281 td->ts.runtime[DDIR_TRIM] += elapsed;
1284 if (td->error || td->terminate)
1287 if (!td->o.do_verify ||
1288 td->o.verify == VERIFY_NONE ||
1289 (td->io_ops->flags & FIO_UNIDIR))
1294 fio_gettime(&td->start, NULL);
1296 do_verify(td, verify_bytes);
1298 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1300 if (td->error || td->terminate)
1304 update_rusage_stat(td);
1305 td->ts.runtime[DDIR_READ] = (td->ts.runtime[DDIR_READ] + 999) / 1000;
1306 td->ts.runtime[DDIR_WRITE] = (td->ts.runtime[DDIR_WRITE] + 999) / 1000;
1307 td->ts.runtime[DDIR_TRIM] = (td->ts.runtime[DDIR_TRIM] + 999) / 1000;
1308 td->ts.total_run_time = mtime_since_now(&td->epoch);
1309 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1310 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1311 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1313 fio_mutex_down(writeout_mutex);
1315 if (td->o.bw_log_file) {
1316 finish_log_named(td, td->bw_log,
1317 td->o.bw_log_file, "bw");
1319 finish_log(td, td->bw_log, "bw");
1322 if (td->o.lat_log_file) {
1323 finish_log_named(td, td->lat_log,
1324 td->o.lat_log_file, "lat");
1326 finish_log(td, td->lat_log, "lat");
1329 if (td->o.lat_log_file) {
1330 finish_log_named(td, td->slat_log,
1331 td->o.lat_log_file, "slat");
1333 finish_log(td, td->slat_log, "slat");
1336 if (td->o.lat_log_file) {
1337 finish_log_named(td, td->clat_log,
1338 td->o.lat_log_file, "clat");
1340 finish_log(td, td->clat_log, "clat");
1343 if (td->o.iops_log_file) {
1344 finish_log_named(td, td->iops_log,
1345 td->o.iops_log_file, "iops");
1347 finish_log(td, td->iops_log, "iops");
1350 fio_mutex_up(writeout_mutex);
1351 if (td->o.exec_postrun)
1352 exec_string(td->o.exec_postrun);
1354 if (exitall_on_terminate)
1355 fio_terminate_threads(td->groupid);
1359 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1362 if (td->o.verify_async)
1363 verify_async_exit(td);
1365 close_and_free_files(td);
1368 cgroup_shutdown(td, &cgroup_mnt);
1370 if (td->o.cpumask_set) {
1371 int ret = fio_cpuset_exit(&td->o.cpumask);
1373 td_verror(td, ret, "fio_cpuset_exit");
1377 * do this very late, it will log file closing as well
1379 if (td->o.write_iolog_file)
1380 write_iolog_close(td);
1382 td_set_runstate(td, TD_EXITED);
1383 return (void *) (uintptr_t) td->error;
1388 * We cannot pass the td data into a forked process, so attach the td and
1389 * pass it to the thread worker.
1391 static int fork_main(int shmid, int offset)
1393 struct thread_data *td;
1397 data = shmat(shmid, NULL, 0);
1398 if (data == (void *) -1) {
1406 * HP-UX inherits shm mappings?
1411 td = data + offset * sizeof(struct thread_data);
1412 ret = thread_main(td);
1414 return (int) (uintptr_t) ret;
1418 * Run over the job map and reap the threads that have exited, if any.
1420 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1421 unsigned int *m_rate)
1423 struct thread_data *td;
1424 unsigned int cputhreads, realthreads, pending;
1428 * reap exited threads (TD_EXITED -> TD_REAPED)
1430 realthreads = pending = cputhreads = 0;
1431 for_each_td(td, i) {
1435 * ->io_ops is NULL for a thread that has closed its
1438 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1447 if (td->runstate == TD_REAPED)
1449 if (td->o.use_thread) {
1450 if (td->runstate == TD_EXITED) {
1451 td_set_runstate(td, TD_REAPED);
1458 if (td->runstate == TD_EXITED)
1462 * check if someone quit or got killed in an unusual way
1464 ret = waitpid(td->pid, &status, flags);
1466 if (errno == ECHILD) {
1467 log_err("fio: pid=%d disappeared %d\n",
1468 (int) td->pid, td->runstate);
1470 td_set_runstate(td, TD_REAPED);
1474 } else if (ret == td->pid) {
1475 if (WIFSIGNALED(status)) {
1476 int sig = WTERMSIG(status);
1478 if (sig != SIGTERM && sig != SIGUSR2)
1479 log_err("fio: pid=%d, got signal=%d\n",
1480 (int) td->pid, sig);
1482 td_set_runstate(td, TD_REAPED);
1485 if (WIFEXITED(status)) {
1486 if (WEXITSTATUS(status) && !td->error)
1487 td->error = WEXITSTATUS(status);
1489 td_set_runstate(td, TD_REAPED);
1495 * thread is not dead, continue
1501 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1502 (*t_rate) -= ddir_rw_sum(td->o.rate);
1509 done_secs += mtime_since_now(&td->epoch) / 1000;
1512 if (*nr_running == cputhreads && !pending && realthreads)
1513 fio_terminate_threads(TERMINATE_ALL);
1517 * Main function for kicking off and reaping jobs, as needed.
1519 static void run_threads(void)
1521 struct thread_data *td;
1522 unsigned long spent;
1523 unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
1525 if (fio_pin_memory())
1528 if (fio_gtod_offload && fio_start_gtod_thread())
1531 fio_idle_prof_init();
1535 if (output_format == FIO_OUTPUT_NORMAL) {
1536 log_info("Starting ");
1538 log_info("%d thread%s", nr_thread,
1539 nr_thread > 1 ? "s" : "");
1543 log_info("%d process%s", nr_process,
1544 nr_process > 1 ? "es" : "");
1550 todo = thread_number;
1553 m_rate = t_rate = 0;
1555 for_each_td(td, i) {
1556 print_status_init(td->thread_number - 1);
1558 if (!td->o.create_serialize)
1562 * do file setup here so it happens sequentially,
1563 * we don't want X number of threads getting their
1564 * client data interspersed on disk
1566 if (setup_files(td)) {
1569 log_err("fio: pid=%d, err=%d/%s\n",
1570 (int) td->pid, td->error, td->verror);
1571 td_set_runstate(td, TD_REAPED);
1578 * for sharing to work, each job must always open
1579 * its own files. so close them, if we opened them
1582 for_each_file(td, f, j) {
1583 if (fio_file_open(f))
1584 td_io_close_file(td, f);
1589 /* start idle threads before io threads start to run */
1590 fio_idle_prof_start();
1595 struct thread_data *map[REAL_MAX_JOBS];
1596 struct timeval this_start;
1597 int this_jobs = 0, left;
1600 * create threads (TD_NOT_CREATED -> TD_CREATED)
1602 for_each_td(td, i) {
1603 if (td->runstate != TD_NOT_CREATED)
1607 * never got a chance to start, killed by other
1608 * thread for some reason
1610 if (td->terminate) {
1615 if (td->o.start_delay) {
1616 spent = mtime_since_genesis();
1618 if (td->o.start_delay * 1000 > spent)
1622 if (td->o.stonewall && (nr_started || nr_running)) {
1623 dprint(FD_PROCESS, "%s: stonewall wait\n",
1631 * Set state to created. Thread will transition
1632 * to TD_INITIALIZED when it's done setting up.
1634 td_set_runstate(td, TD_CREATED);
1635 map[this_jobs++] = td;
1638 if (td->o.use_thread) {
1641 dprint(FD_PROCESS, "will pthread_create\n");
1642 ret = pthread_create(&td->thread, NULL,
1645 log_err("pthread_create: %s\n",
1650 ret = pthread_detach(td->thread);
1652 log_err("pthread_detach: %s",
1656 dprint(FD_PROCESS, "will fork\n");
1659 int ret = fork_main(shm_id, i);
1662 } else if (i == fio_debug_jobno)
1663 *fio_debug_jobp = pid;
1665 dprint(FD_MUTEX, "wait on startup_mutex\n");
1666 if (fio_mutex_down_timeout(startup_mutex, 10)) {
1667 log_err("fio: job startup hung? exiting.\n");
1668 fio_terminate_threads(TERMINATE_ALL);
1673 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1677 * Wait for the started threads to transition to
1680 fio_gettime(&this_start, NULL);
1682 while (left && !fio_abort) {
1683 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1688 for (i = 0; i < this_jobs; i++) {
1692 if (td->runstate == TD_INITIALIZED) {
1695 } else if (td->runstate >= TD_EXITED) {
1699 nr_running++; /* work-around... */
1705 log_err("fio: %d job%s failed to start\n", left,
1706 left > 1 ? "s" : "");
1707 for (i = 0; i < this_jobs; i++) {
1711 kill(td->pid, SIGTERM);
1717 * start created threads (TD_INITIALIZED -> TD_RUNNING).
1719 for_each_td(td, i) {
1720 if (td->runstate != TD_INITIALIZED)
1723 if (in_ramp_time(td))
1724 td_set_runstate(td, TD_RAMP);
1726 td_set_runstate(td, TD_RUNNING);
1729 m_rate += ddir_rw_sum(td->o.ratemin);
1730 t_rate += ddir_rw_sum(td->o.rate);
1732 fio_mutex_up(td->mutex);
1735 reap_threads(&nr_running, &t_rate, &m_rate);
1739 fio_server_idle_loop();
1745 while (nr_running) {
1746 reap_threads(&nr_running, &t_rate, &m_rate);
1749 fio_server_idle_loop();
1754 fio_idle_prof_stop();
1760 void wait_for_disk_thread_exit(void)
1762 fio_mutex_down(disk_thread_mutex);
1765 static void free_disk_util(void)
1767 disk_util_start_exit();
1768 wait_for_disk_thread_exit();
1769 disk_util_prune_entries();
1772 static void *disk_thread_main(void *data)
1776 fio_mutex_up(startup_mutex);
1778 while (threads && !ret) {
1779 usleep(DISK_UTIL_MSEC * 1000);
1782 ret = update_io_ticks();
1785 print_thread_status();
1788 fio_mutex_up(disk_thread_mutex);
1792 static int create_disk_util_thread(void)
1798 disk_thread_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
1800 ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
1802 fio_mutex_remove(disk_thread_mutex);
1803 log_err("Can't create disk util thread: %s\n", strerror(ret));
1807 ret = pthread_detach(disk_util_thread);
1809 fio_mutex_remove(disk_thread_mutex);
1810 log_err("Can't detatch disk util thread: %s\n", strerror(ret));
1814 dprint(FD_MUTEX, "wait on startup_mutex\n");
1815 fio_mutex_down(startup_mutex);
1816 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1820 int fio_backend(void)
1822 struct thread_data *td;
1826 if (load_profile(exec_profile))
1829 exec_profile = NULL;
1835 setup_log(&agg_io_log[DDIR_READ], 0);
1836 setup_log(&agg_io_log[DDIR_WRITE], 0);
1837 setup_log(&agg_io_log[DDIR_TRIM], 0);
1840 startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
1841 if (startup_mutex == NULL)
1843 writeout_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1844 if (writeout_mutex == NULL)
1848 create_disk_util_thread();
1850 cgroup_list = smalloc(sizeof(*cgroup_list));
1851 INIT_FLIST_HEAD(cgroup_list);
1858 __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1859 __finish_log(agg_io_log[DDIR_WRITE],
1860 "agg-write_bw.log");
1861 __finish_log(agg_io_log[DDIR_TRIM],
1862 "agg-write_bw.log");
1867 fio_options_free(td);
1870 cgroup_kill(cgroup_list);
1874 fio_mutex_remove(startup_mutex);
1875 fio_mutex_remove(writeout_mutex);
1876 fio_mutex_remove(disk_thread_mutex);