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);
409 static void check_update_rusage(struct thread_data *td)
411 if (td->update_rusage) {
412 td->update_rusage = 0;
413 update_rusage_stat(td);
414 fio_mutex_up(td->rusage_sem);
419 * The main verify engine. Runs over the writes we previously submitted,
420 * reads the blocks back in, and checks the crc/md5 of the data.
422 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
424 uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
430 dprint(FD_VERIFY, "starting loop\n");
433 * sync io first and invalidate cache, to make sure we really
436 for_each_file(td, f, i) {
437 if (!fio_file_open(f))
439 if (fio_io_sync(td, f))
441 if (file_invalidate_cache(td, f))
445 check_update_rusage(td);
450 td_set_runstate(td, TD_VERIFYING);
453 while (!td->terminate) {
458 check_update_rusage(td);
460 if (runtime_exceeded(td, &td->tv_cache)) {
461 __update_tv_cache(td);
462 if (runtime_exceeded(td, &td->tv_cache)) {
468 if (flow_threshold_exceeded(td))
471 if (!td->o.experimental_verify) {
472 io_u = __get_io_u(td);
476 if (get_next_verify(td, io_u)) {
481 if (td_io_prep(td, io_u)) {
486 if (ddir_rw_sum(bytes_done) + td->o.rw_min_bs > verify_bytes)
489 while ((io_u = get_io_u(td)) != NULL) {
491 * We are only interested in the places where
492 * we wrote or trimmed IOs. Turn those into
493 * reads for verification purposes.
495 if (io_u->ddir == DDIR_READ) {
497 * Pretend we issued it for rwmix
500 td->io_issues[DDIR_READ]++;
503 } else if (io_u->ddir == DDIR_TRIM) {
504 io_u->ddir = DDIR_READ;
505 io_u->flags |= IO_U_F_TRIMMED;
507 } else if (io_u->ddir == DDIR_WRITE) {
508 io_u->ddir = DDIR_READ;
520 if (td->o.verify_async)
521 io_u->end_io = verify_io_u_async;
523 io_u->end_io = verify_io_u;
527 ret = td_io_queue(td, io_u);
529 case FIO_Q_COMPLETED:
532 clear_io_u(td, io_u);
533 } else if (io_u->resid) {
534 int bytes = io_u->xfer_buflen - io_u->resid;
540 td_verror(td, EIO, "full resid");
545 io_u->xfer_buflen = io_u->resid;
546 io_u->xfer_buf += bytes;
547 io_u->offset += bytes;
549 if (ddir_rw(io_u->ddir))
550 td->ts.short_io_u[io_u->ddir]++;
553 if (io_u->offset == f->real_file_size)
556 requeue_io_u(td, &io_u);
559 ret = io_u_sync_complete(td, io_u, bytes_done);
567 requeue_io_u(td, &io_u);
568 ret2 = td_io_commit(td);
574 td_verror(td, -ret, "td_io_queue");
578 if (break_on_this_error(td, ddir, &ret))
582 * if we can queue more, do so. but check if there are
583 * completed io_u's first. Note that we can get BUSY even
584 * without IO queued, if the system is resource starved.
586 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
587 if (full || !td->o.iodepth_batch_complete) {
588 min_events = min(td->o.iodepth_batch_complete,
591 * if the queue is full, we MUST reap at least 1 event
593 if (full && !min_events)
598 * Reap required number of io units, if any,
599 * and do the verification on them through
600 * the callback handler
602 if (io_u_queued_complete(td, min_events, bytes_done) < 0) {
606 } while (full && (td->cur_depth > td->o.iodepth_low));
612 check_update_rusage(td);
615 min_events = td->cur_depth;
618 ret = io_u_queued_complete(td, min_events, NULL);
620 cleanup_pending_aio(td);
622 td_set_runstate(td, TD_RUNNING);
624 dprint(FD_VERIFY, "exiting loop\n");
627 static int io_bytes_exceeded(struct thread_data *td)
629 unsigned long long bytes;
632 bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE];
633 else if (td_write(td))
634 bytes = td->this_io_bytes[DDIR_WRITE];
635 else if (td_read(td))
636 bytes = td->this_io_bytes[DDIR_READ];
638 bytes = td->this_io_bytes[DDIR_TRIM];
640 return bytes >= td->o.size;
644 * Main IO worker function. It retrieves io_u's to process and queues
645 * and reaps them, checking for rate and errors along the way.
647 * Returns number of bytes written and trimmed.
649 static uint64_t do_io(struct thread_data *td)
651 uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
655 if (in_ramp_time(td))
656 td_set_runstate(td, TD_RAMP);
658 td_set_runstate(td, TD_RUNNING);
660 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
661 (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td) ||
663 struct timeval comp_time;
669 check_update_rusage(td);
671 if (td->terminate || td->done)
676 if (runtime_exceeded(td, &td->tv_cache)) {
677 __update_tv_cache(td);
678 if (runtime_exceeded(td, &td->tv_cache)) {
684 if (flow_threshold_exceeded(td))
694 * Add verification end_io handler if:
695 * - Asked to verify (!td_rw(td))
696 * - Or the io_u is from our verify list (mixed write/ver)
698 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
699 ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
700 if (td->o.verify_async)
701 io_u->end_io = verify_io_u_async;
703 io_u->end_io = verify_io_u;
704 td_set_runstate(td, TD_VERIFYING);
705 } else if (in_ramp_time(td))
706 td_set_runstate(td, TD_RAMP);
708 td_set_runstate(td, TD_RUNNING);
710 ret = td_io_queue(td, io_u);
712 case FIO_Q_COMPLETED:
715 clear_io_u(td, io_u);
716 } else if (io_u->resid) {
717 int bytes = io_u->xfer_buflen - io_u->resid;
718 struct fio_file *f = io_u->file;
724 td_verror(td, EIO, "full resid");
729 io_u->xfer_buflen = io_u->resid;
730 io_u->xfer_buf += bytes;
731 io_u->offset += bytes;
733 if (ddir_rw(io_u->ddir))
734 td->ts.short_io_u[io_u->ddir]++;
736 if (io_u->offset == f->real_file_size)
739 requeue_io_u(td, &io_u);
742 if (__should_check_rate(td, DDIR_READ) ||
743 __should_check_rate(td, DDIR_WRITE) ||
744 __should_check_rate(td, DDIR_TRIM))
745 fio_gettime(&comp_time, NULL);
747 ret = io_u_sync_complete(td, io_u, bytes_done);
754 * if the engine doesn't have a commit hook,
755 * the io_u is really queued. if it does have such
756 * a hook, it has to call io_u_queued() itself.
758 if (td->io_ops->commit == NULL)
759 io_u_queued(td, io_u);
762 requeue_io_u(td, &io_u);
763 ret2 = td_io_commit(td);
773 if (break_on_this_error(td, ddir, &ret))
777 * See if we need to complete some commands. Note that we
778 * can get BUSY even without IO queued, if the system is
781 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
782 if (full || !td->o.iodepth_batch_complete) {
783 min_evts = min(td->o.iodepth_batch_complete,
786 * if the queue is full, we MUST reap at least 1 event
788 if (full && !min_evts)
791 if (__should_check_rate(td, DDIR_READ) ||
792 __should_check_rate(td, DDIR_WRITE) ||
793 __should_check_rate(td, DDIR_TRIM))
794 fio_gettime(&comp_time, NULL);
797 ret = io_u_queued_complete(td, min_evts, bytes_done);
801 } while (full && (td->cur_depth > td->o.iodepth_low));
806 if (!ddir_rw_sum(bytes_done) && !(td->io_ops->flags & FIO_NOIO))
809 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
810 if (check_min_rate(td, &comp_time, bytes_done)) {
811 if (exitall_on_terminate)
812 fio_terminate_threads(td->groupid);
813 td_verror(td, EIO, "check_min_rate");
818 if (td->o.thinktime) {
819 unsigned long long b;
821 b = ddir_rw_sum(td->io_blocks);
822 if (!(b % td->o.thinktime_blocks)) {
825 if (td->o.thinktime_spin)
826 usec_spin(td->o.thinktime_spin);
828 left = td->o.thinktime - td->o.thinktime_spin;
830 usec_sleep(td, left);
835 check_update_rusage(td);
837 if (td->trim_entries)
838 log_err("fio: %d trim entries leaked?\n", td->trim_entries);
840 if (td->o.fill_device && td->error == ENOSPC) {
849 ret = io_u_queued_complete(td, i, bytes_done);
850 if (td->o.fill_device && td->error == ENOSPC)
854 if (should_fsync(td) && td->o.end_fsync) {
855 td_set_runstate(td, TD_FSYNCING);
857 for_each_file(td, f, i) {
858 if (!fio_file_fsync(td, f))
861 log_err("fio: end_fsync failed for file %s\n",
866 cleanup_pending_aio(td);
869 * stop job if we failed doing any IO
871 if (!ddir_rw_sum(td->this_io_bytes))
874 return bytes_done[DDIR_WRITE] + bytes_done[DDIR_TRIM];
877 static void cleanup_io_u(struct thread_data *td)
879 struct flist_head *entry, *n;
882 flist_for_each_safe(entry, n, &td->io_u_freelist) {
883 io_u = flist_entry(entry, struct io_u, list);
885 flist_del(&io_u->list);
887 if (td->io_ops->io_u_free)
888 td->io_ops->io_u_free(td, io_u);
890 fio_memfree(io_u, sizeof(*io_u));
896 static int init_io_u(struct thread_data *td)
899 unsigned int max_bs, min_write;
900 int cl_align, i, max_units;
904 max_units = td->o.iodepth;
905 max_bs = td_max_bs(td);
906 min_write = td->o.min_bs[DDIR_WRITE];
907 td->orig_buffer_size = (unsigned long long) max_bs
908 * (unsigned long long) max_units;
910 if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
913 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
916 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
917 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
920 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
921 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
925 if (data_xfer && allocate_io_mem(td))
928 if (td->o.odirect || td->o.mem_align ||
929 (td->io_ops->flags & FIO_RAWIO))
930 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
934 cl_align = os_cache_line_size();
936 for (i = 0; i < max_units; i++) {
942 ptr = fio_memalign(cl_align, sizeof(*io_u));
944 log_err("fio: unable to allocate aligned memory\n");
949 memset(io_u, 0, sizeof(*io_u));
950 INIT_FLIST_HEAD(&io_u->list);
951 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
955 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
958 io_u_fill_buffer(td, io_u, min_write, max_bs);
959 if (td_write(td) && td->o.verify_pattern_bytes) {
961 * Fill the buffer with the pattern if we are
962 * going to be doing writes.
964 fill_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
969 io_u->flags = IO_U_F_FREE;
970 flist_add(&io_u->list, &td->io_u_freelist);
972 if (td->io_ops->io_u_init) {
973 int ret = td->io_ops->io_u_init(td, io_u);
976 log_err("fio: failed to init engine data: %d\n", ret);
987 static int switch_ioscheduler(struct thread_data *td)
989 char tmp[256], tmp2[128];
993 if (td->io_ops->flags & FIO_DISKLESSIO)
996 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
998 f = fopen(tmp, "r+");
1000 if (errno == ENOENT) {
1001 log_err("fio: os or kernel doesn't support IO scheduler"
1005 td_verror(td, errno, "fopen iosched");
1012 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1013 if (ferror(f) || ret != 1) {
1014 td_verror(td, errno, "fwrite");
1022 * Read back and check that the selected scheduler is now the default.
1024 ret = fread(tmp, 1, sizeof(tmp), f);
1025 if (ferror(f) || ret < 0) {
1026 td_verror(td, errno, "fread");
1031 sprintf(tmp2, "[%s]", td->o.ioscheduler);
1032 if (!strstr(tmp, tmp2)) {
1033 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1034 td_verror(td, EINVAL, "iosched_switch");
1043 static int keep_running(struct thread_data *td)
1047 if (td->o.time_based)
1054 if (td->o.size != -1ULL && ddir_rw_sum(td->io_bytes) < td->o.size) {
1058 * If the difference is less than the minimum IO size, we
1061 diff = td->o.size - ddir_rw_sum(td->io_bytes);
1062 if (diff < td_max_bs(td))
1071 static int exec_string(const char *string)
1073 int ret, newlen = strlen(string) + 1 + 8;
1076 str = malloc(newlen);
1077 sprintf(str, "sh -c %s", string);
1081 log_err("fio: exec of cmd <%s> failed\n", str);
1088 * Entry point for the thread based jobs. The process based jobs end up
1089 * here as well, after a little setup.
1091 static void *thread_main(void *data)
1093 unsigned long long elapsed;
1094 struct thread_data *td = data;
1095 pthread_condattr_t attr;
1098 if (!td->o.use_thread) {
1104 fio_local_clock_init(td->o.use_thread);
1106 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1108 INIT_FLIST_HEAD(&td->io_u_freelist);
1109 INIT_FLIST_HEAD(&td->io_u_busylist);
1110 INIT_FLIST_HEAD(&td->io_u_requeues);
1111 INIT_FLIST_HEAD(&td->io_log_list);
1112 INIT_FLIST_HEAD(&td->io_hist_list);
1113 INIT_FLIST_HEAD(&td->verify_list);
1114 INIT_FLIST_HEAD(&td->trim_list);
1115 INIT_FLIST_HEAD(&td->next_rand_list);
1116 pthread_mutex_init(&td->io_u_lock, NULL);
1117 td->io_hist_tree = RB_ROOT;
1119 pthread_condattr_init(&attr);
1120 pthread_cond_init(&td->verify_cond, &attr);
1121 pthread_cond_init(&td->free_cond, &attr);
1123 td_set_runstate(td, TD_INITIALIZED);
1124 dprint(FD_MUTEX, "up startup_mutex\n");
1125 fio_mutex_up(startup_mutex);
1126 dprint(FD_MUTEX, "wait on td->mutex\n");
1127 fio_mutex_down(td->mutex);
1128 dprint(FD_MUTEX, "done waiting on td->mutex\n");
1131 * the ->mutex mutex is now no longer used, close it to avoid
1132 * eating a file descriptor
1134 fio_mutex_remove(td->mutex);
1137 * A new gid requires privilege, so we need to do this before setting
1140 if (td->o.gid != -1U && setgid(td->o.gid)) {
1141 td_verror(td, errno, "setgid");
1144 if (td->o.uid != -1U && setuid(td->o.uid)) {
1145 td_verror(td, errno, "setuid");
1150 * If we have a gettimeofday() thread, make sure we exclude that
1151 * thread from this job
1154 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1157 * Set affinity first, in case it has an impact on the memory
1160 if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1161 td_verror(td, errno, "cpu_set_affinity");
1165 #ifdef CONFIG_LIBNUMA
1166 /* numa node setup */
1167 if (td->o.numa_cpumask_set || td->o.numa_memmask_set) {
1170 if (numa_available() < 0) {
1171 td_verror(td, errno, "Does not support NUMA API\n");
1175 if (td->o.numa_cpumask_set) {
1176 ret = numa_run_on_node_mask(td->o.numa_cpunodesmask);
1178 td_verror(td, errno, \
1179 "numa_run_on_node_mask failed\n");
1184 if (td->o.numa_memmask_set) {
1186 switch (td->o.numa_mem_mode) {
1187 case MPOL_INTERLEAVE:
1188 numa_set_interleave_mask(td->o.numa_memnodesmask);
1191 numa_set_membind(td->o.numa_memnodesmask);
1194 numa_set_localalloc();
1196 case MPOL_PREFERRED:
1197 numa_set_preferred(td->o.numa_mem_prefer_node);
1209 * May alter parameters that init_io_u() will use, so we need to
1218 if (td->o.verify_async && verify_async_init(td))
1221 if (td->ioprio_set) {
1222 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1223 td_verror(td, errno, "ioprio_set");
1228 if (td->o.cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1232 if (nice(td->o.nice) == -1 && errno != 0) {
1233 td_verror(td, errno, "nice");
1237 if (td->o.ioscheduler && switch_ioscheduler(td))
1240 if (!td->o.create_serialize && setup_files(td))
1246 if (init_random_map(td))
1249 if (td->o.exec_prerun) {
1250 if (exec_string(td->o.exec_prerun))
1254 if (td->o.pre_read) {
1255 if (pre_read_files(td) < 0)
1259 fio_gettime(&td->epoch, NULL);
1260 fio_getrusage(&td->ru_start);
1262 while (keep_running(td)) {
1263 uint64_t verify_bytes;
1265 fio_gettime(&td->start, NULL);
1266 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1267 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1268 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1270 if (td->o.ratemin[DDIR_READ] || td->o.ratemin[DDIR_WRITE] ||
1271 td->o.ratemin[DDIR_TRIM]) {
1272 memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1273 sizeof(td->bw_sample_time));
1274 memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1275 sizeof(td->bw_sample_time));
1276 memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1277 sizeof(td->bw_sample_time));
1283 prune_io_piece_log(td);
1285 verify_bytes = do_io(td);
1289 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1290 elapsed = utime_since_now(&td->start);
1291 td->ts.runtime[DDIR_READ] += elapsed;
1293 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1294 elapsed = utime_since_now(&td->start);
1295 td->ts.runtime[DDIR_WRITE] += elapsed;
1297 if (td_trim(td) && td->io_bytes[DDIR_TRIM]) {
1298 elapsed = utime_since_now(&td->start);
1299 td->ts.runtime[DDIR_TRIM] += elapsed;
1302 if (td->error || td->terminate)
1305 if (!td->o.do_verify ||
1306 td->o.verify == VERIFY_NONE ||
1307 (td->io_ops->flags & FIO_UNIDIR))
1312 fio_gettime(&td->start, NULL);
1314 do_verify(td, verify_bytes);
1316 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1318 if (td->error || td->terminate)
1322 update_rusage_stat(td);
1323 td->ts.runtime[DDIR_READ] = (td->ts.runtime[DDIR_READ] + 999) / 1000;
1324 td->ts.runtime[DDIR_WRITE] = (td->ts.runtime[DDIR_WRITE] + 999) / 1000;
1325 td->ts.runtime[DDIR_TRIM] = (td->ts.runtime[DDIR_TRIM] + 999) / 1000;
1326 td->ts.total_run_time = mtime_since_now(&td->epoch);
1327 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1328 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1329 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1331 fio_mutex_down(writeout_mutex);
1333 if (td->o.bw_log_file) {
1334 finish_log_named(td, td->bw_log,
1335 td->o.bw_log_file, "bw");
1337 finish_log(td, td->bw_log, "bw");
1340 if (td->o.lat_log_file) {
1341 finish_log_named(td, td->lat_log,
1342 td->o.lat_log_file, "lat");
1344 finish_log(td, td->lat_log, "lat");
1347 if (td->o.lat_log_file) {
1348 finish_log_named(td, td->slat_log,
1349 td->o.lat_log_file, "slat");
1351 finish_log(td, td->slat_log, "slat");
1354 if (td->o.lat_log_file) {
1355 finish_log_named(td, td->clat_log,
1356 td->o.lat_log_file, "clat");
1358 finish_log(td, td->clat_log, "clat");
1361 if (td->o.iops_log_file) {
1362 finish_log_named(td, td->iops_log,
1363 td->o.iops_log_file, "iops");
1365 finish_log(td, td->iops_log, "iops");
1368 fio_mutex_up(writeout_mutex);
1369 if (td->o.exec_postrun)
1370 exec_string(td->o.exec_postrun);
1372 if (exitall_on_terminate)
1373 fio_terminate_threads(td->groupid);
1377 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1380 if (td->o.verify_async)
1381 verify_async_exit(td);
1383 close_and_free_files(td);
1386 cgroup_shutdown(td, &cgroup_mnt);
1388 if (td->o.cpumask_set) {
1389 int ret = fio_cpuset_exit(&td->o.cpumask);
1391 td_verror(td, ret, "fio_cpuset_exit");
1395 * do this very late, it will log file closing as well
1397 if (td->o.write_iolog_file)
1398 write_iolog_close(td);
1400 fio_mutex_remove(td->rusage_sem);
1401 td->rusage_sem = NULL;
1403 td_set_runstate(td, TD_EXITED);
1404 return (void *) (uintptr_t) td->error;
1409 * We cannot pass the td data into a forked process, so attach the td and
1410 * pass it to the thread worker.
1412 static int fork_main(int shmid, int offset)
1414 struct thread_data *td;
1418 data = shmat(shmid, NULL, 0);
1419 if (data == (void *) -1) {
1427 * HP-UX inherits shm mappings?
1432 td = data + offset * sizeof(struct thread_data);
1433 ret = thread_main(td);
1435 return (int) (uintptr_t) ret;
1439 * Run over the job map and reap the threads that have exited, if any.
1441 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1442 unsigned int *m_rate)
1444 struct thread_data *td;
1445 unsigned int cputhreads, realthreads, pending;
1449 * reap exited threads (TD_EXITED -> TD_REAPED)
1451 realthreads = pending = cputhreads = 0;
1452 for_each_td(td, i) {
1456 * ->io_ops is NULL for a thread that has closed its
1459 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1468 if (td->runstate == TD_REAPED)
1470 if (td->o.use_thread) {
1471 if (td->runstate == TD_EXITED) {
1472 td_set_runstate(td, TD_REAPED);
1479 if (td->runstate == TD_EXITED)
1483 * check if someone quit or got killed in an unusual way
1485 ret = waitpid(td->pid, &status, flags);
1487 if (errno == ECHILD) {
1488 log_err("fio: pid=%d disappeared %d\n",
1489 (int) td->pid, td->runstate);
1491 td_set_runstate(td, TD_REAPED);
1495 } else if (ret == td->pid) {
1496 if (WIFSIGNALED(status)) {
1497 int sig = WTERMSIG(status);
1499 if (sig != SIGTERM && sig != SIGUSR2)
1500 log_err("fio: pid=%d, got signal=%d\n",
1501 (int) td->pid, sig);
1503 td_set_runstate(td, TD_REAPED);
1506 if (WIFEXITED(status)) {
1507 if (WEXITSTATUS(status) && !td->error)
1508 td->error = WEXITSTATUS(status);
1510 td_set_runstate(td, TD_REAPED);
1516 * thread is not dead, continue
1522 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1523 (*t_rate) -= ddir_rw_sum(td->o.rate);
1530 done_secs += mtime_since_now(&td->epoch) / 1000;
1533 if (*nr_running == cputhreads && !pending && realthreads)
1534 fio_terminate_threads(TERMINATE_ALL);
1538 * Main function for kicking off and reaping jobs, as needed.
1540 static void run_threads(void)
1542 struct thread_data *td;
1543 unsigned long spent;
1544 unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
1546 if (fio_pin_memory())
1549 if (fio_gtod_offload && fio_start_gtod_thread())
1552 fio_idle_prof_init();
1556 if (output_format == FIO_OUTPUT_NORMAL) {
1557 log_info("Starting ");
1559 log_info("%d thread%s", nr_thread,
1560 nr_thread > 1 ? "s" : "");
1564 log_info("%d process%s", nr_process,
1565 nr_process > 1 ? "es" : "");
1571 todo = thread_number;
1574 m_rate = t_rate = 0;
1576 for_each_td(td, i) {
1577 print_status_init(td->thread_number - 1);
1579 if (!td->o.create_serialize)
1583 * do file setup here so it happens sequentially,
1584 * we don't want X number of threads getting their
1585 * client data interspersed on disk
1587 if (setup_files(td)) {
1590 log_err("fio: pid=%d, err=%d/%s\n",
1591 (int) td->pid, td->error, td->verror);
1592 td_set_runstate(td, TD_REAPED);
1599 * for sharing to work, each job must always open
1600 * its own files. so close them, if we opened them
1603 for_each_file(td, f, j) {
1604 if (fio_file_open(f))
1605 td_io_close_file(td, f);
1610 /* start idle threads before io threads start to run */
1611 fio_idle_prof_start();
1616 struct thread_data *map[REAL_MAX_JOBS];
1617 struct timeval this_start;
1618 int this_jobs = 0, left;
1621 * create threads (TD_NOT_CREATED -> TD_CREATED)
1623 for_each_td(td, i) {
1624 if (td->runstate != TD_NOT_CREATED)
1628 * never got a chance to start, killed by other
1629 * thread for some reason
1631 if (td->terminate) {
1636 if (td->o.start_delay) {
1637 spent = mtime_since_genesis();
1639 if (td->o.start_delay * 1000 > spent)
1643 if (td->o.stonewall && (nr_started || nr_running)) {
1644 dprint(FD_PROCESS, "%s: stonewall wait\n",
1651 td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED);
1652 td->update_rusage = 0;
1655 * Set state to created. Thread will transition
1656 * to TD_INITIALIZED when it's done setting up.
1658 td_set_runstate(td, TD_CREATED);
1659 map[this_jobs++] = td;
1662 if (td->o.use_thread) {
1665 dprint(FD_PROCESS, "will pthread_create\n");
1666 ret = pthread_create(&td->thread, NULL,
1669 log_err("pthread_create: %s\n",
1674 ret = pthread_detach(td->thread);
1676 log_err("pthread_detach: %s",
1680 dprint(FD_PROCESS, "will fork\n");
1683 int ret = fork_main(shm_id, i);
1686 } else if (i == fio_debug_jobno)
1687 *fio_debug_jobp = pid;
1689 dprint(FD_MUTEX, "wait on startup_mutex\n");
1690 if (fio_mutex_down_timeout(startup_mutex, 10)) {
1691 log_err("fio: job startup hung? exiting.\n");
1692 fio_terminate_threads(TERMINATE_ALL);
1697 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1701 * Wait for the started threads to transition to
1704 fio_gettime(&this_start, NULL);
1706 while (left && !fio_abort) {
1707 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1712 for (i = 0; i < this_jobs; i++) {
1716 if (td->runstate == TD_INITIALIZED) {
1719 } else if (td->runstate >= TD_EXITED) {
1723 nr_running++; /* work-around... */
1729 log_err("fio: %d job%s failed to start\n", left,
1730 left > 1 ? "s" : "");
1731 for (i = 0; i < this_jobs; i++) {
1735 kill(td->pid, SIGTERM);
1741 * start created threads (TD_INITIALIZED -> TD_RUNNING).
1743 for_each_td(td, i) {
1744 if (td->runstate != TD_INITIALIZED)
1747 if (in_ramp_time(td))
1748 td_set_runstate(td, TD_RAMP);
1750 td_set_runstate(td, TD_RUNNING);
1753 m_rate += ddir_rw_sum(td->o.ratemin);
1754 t_rate += ddir_rw_sum(td->o.rate);
1756 fio_mutex_up(td->mutex);
1759 reap_threads(&nr_running, &t_rate, &m_rate);
1763 fio_server_idle_loop();
1769 while (nr_running) {
1770 reap_threads(&nr_running, &t_rate, &m_rate);
1773 fio_server_idle_loop();
1778 fio_idle_prof_stop();
1784 void wait_for_disk_thread_exit(void)
1786 fio_mutex_down(disk_thread_mutex);
1789 static void free_disk_util(void)
1791 disk_util_start_exit();
1792 wait_for_disk_thread_exit();
1793 disk_util_prune_entries();
1796 static void *disk_thread_main(void *data)
1800 fio_mutex_up(startup_mutex);
1802 while (threads && !ret) {
1803 usleep(DISK_UTIL_MSEC * 1000);
1806 ret = update_io_ticks();
1809 print_thread_status();
1812 fio_mutex_up(disk_thread_mutex);
1816 static int create_disk_util_thread(void)
1822 disk_thread_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
1824 ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
1826 fio_mutex_remove(disk_thread_mutex);
1827 log_err("Can't create disk util thread: %s\n", strerror(ret));
1831 ret = pthread_detach(disk_util_thread);
1833 fio_mutex_remove(disk_thread_mutex);
1834 log_err("Can't detatch disk util thread: %s\n", strerror(ret));
1838 dprint(FD_MUTEX, "wait on startup_mutex\n");
1839 fio_mutex_down(startup_mutex);
1840 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1844 int fio_backend(void)
1846 struct thread_data *td;
1850 if (load_profile(exec_profile))
1853 exec_profile = NULL;
1859 setup_log(&agg_io_log[DDIR_READ], 0);
1860 setup_log(&agg_io_log[DDIR_WRITE], 0);
1861 setup_log(&agg_io_log[DDIR_TRIM], 0);
1864 startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
1865 if (startup_mutex == NULL)
1867 writeout_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1868 if (writeout_mutex == NULL)
1872 create_disk_util_thread();
1874 cgroup_list = smalloc(sizeof(*cgroup_list));
1875 INIT_FLIST_HEAD(cgroup_list);
1882 __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1883 __finish_log(agg_io_log[DDIR_WRITE],
1884 "agg-write_bw.log");
1885 __finish_log(agg_io_log[DDIR_TRIM],
1886 "agg-write_bw.log");
1891 fio_options_free(td);
1894 cgroup_kill(cgroup_list);
1898 fio_mutex_remove(startup_mutex);
1899 fio_mutex_remove(writeout_mutex);
1900 fio_mutex_remove(disk_thread_mutex);