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;
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,
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)
244 * get immediately available events, if any
246 r = io_u_queued_complete(td, 0, NULL);
251 * now cancel remaining active events
253 if (td->io_ops->cancel) {
257 io_u_qiter(&td->io_u_all, io_u, i) {
258 if (io_u->flags & IO_U_F_FLIGHT) {
259 r = td->io_ops->cancel(td, io_u);
267 r = io_u_queued_complete(td, td->cur_depth, NULL);
271 * Helper to handle the final sync of a file. Works just like the normal
272 * io path, just does everything sync.
274 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
276 struct io_u *io_u = __get_io_u(td);
282 io_u->ddir = DDIR_SYNC;
285 if (td_io_prep(td, io_u)) {
291 ret = td_io_queue(td, io_u);
293 td_verror(td, io_u->error, "td_io_queue");
296 } else if (ret == FIO_Q_QUEUED) {
297 if (io_u_queued_complete(td, 1, NULL) < 0)
299 } else if (ret == FIO_Q_COMPLETED) {
301 td_verror(td, io_u->error, "td_io_queue");
305 if (io_u_sync_complete(td, io_u, NULL) < 0)
307 } else if (ret == FIO_Q_BUSY) {
308 if (td_io_commit(td))
316 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
320 if (fio_file_open(f))
321 return fio_io_sync(td, f);
323 if (td_io_open_file(td, f))
326 ret = fio_io_sync(td, f);
327 td_io_close_file(td, f);
331 static inline void __update_tv_cache(struct thread_data *td)
333 fio_gettime(&td->tv_cache, NULL);
336 static inline void update_tv_cache(struct thread_data *td)
338 if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
339 __update_tv_cache(td);
342 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
344 if (in_ramp_time(td))
348 if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
354 static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
359 if (ret < 0 || td->error) {
361 enum error_type_bit eb;
366 eb = td_error_type(ddir, err);
367 if (!(td->o.continue_on_error & (1 << eb)))
370 if (td_non_fatal_error(td, eb, err)) {
372 * Continue with the I/Os in case of
375 update_error_count(td, err);
379 } else if (td->o.fill_device && err == ENOSPC) {
381 * We expect to hit this error if
382 * fill_device option is set.
389 * Stop the I/O in case of a fatal
392 update_error_count(td, err);
400 static void check_update_rusage(struct thread_data *td)
402 if (td->update_rusage) {
403 td->update_rusage = 0;
404 update_rusage_stat(td);
405 fio_mutex_up(td->rusage_sem);
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))
436 check_update_rusage(td);
441 td_set_runstate(td, TD_VERIFYING);
444 while (!td->terminate) {
449 check_update_rusage(td);
451 if (runtime_exceeded(td, &td->tv_cache)) {
452 __update_tv_cache(td);
453 if (runtime_exceeded(td, &td->tv_cache)) {
459 if (flow_threshold_exceeded(td))
462 if (!td->o.experimental_verify) {
463 io_u = __get_io_u(td);
467 if (get_next_verify(td, io_u)) {
472 if (td_io_prep(td, io_u)) {
477 if (ddir_rw_sum(bytes_done) + td->o.rw_min_bs > verify_bytes)
480 while ((io_u = get_io_u(td)) != NULL) {
482 * We are only interested in the places where
483 * we wrote or trimmed IOs. Turn those into
484 * reads for verification purposes.
486 if (io_u->ddir == DDIR_READ) {
488 * Pretend we issued it for rwmix
491 td->io_issues[DDIR_READ]++;
494 } else if (io_u->ddir == DDIR_TRIM) {
495 io_u->ddir = DDIR_READ;
496 io_u->flags |= IO_U_F_TRIMMED;
498 } else if (io_u->ddir == DDIR_WRITE) {
499 io_u->ddir = DDIR_READ;
511 if (td->o.verify_async)
512 io_u->end_io = verify_io_u_async;
514 io_u->end_io = verify_io_u;
518 ret = td_io_queue(td, io_u);
520 case FIO_Q_COMPLETED:
523 clear_io_u(td, io_u);
524 } else if (io_u->resid) {
525 int bytes = io_u->xfer_buflen - io_u->resid;
531 td_verror(td, EIO, "full resid");
536 io_u->xfer_buflen = io_u->resid;
537 io_u->xfer_buf += bytes;
538 io_u->offset += bytes;
540 if (ddir_rw(io_u->ddir))
541 td->ts.short_io_u[io_u->ddir]++;
544 if (io_u->offset == f->real_file_size)
547 requeue_io_u(td, &io_u);
550 ret = io_u_sync_complete(td, io_u, bytes_done);
558 requeue_io_u(td, &io_u);
559 ret2 = td_io_commit(td);
565 td_verror(td, -ret, "td_io_queue");
569 if (break_on_this_error(td, ddir, &ret))
573 * if we can queue more, do so. but check if there are
574 * completed io_u's first. Note that we can get BUSY even
575 * without IO queued, if the system is resource starved.
577 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
578 if (full || !td->o.iodepth_batch_complete) {
579 min_events = min(td->o.iodepth_batch_complete,
582 * if the queue is full, we MUST reap at least 1 event
584 if (full && !min_events)
589 * Reap required number of io units, if any,
590 * and do the verification on them through
591 * the callback handler
593 if (io_u_queued_complete(td, min_events, bytes_done) < 0) {
597 } while (full && (td->cur_depth > td->o.iodepth_low));
603 check_update_rusage(td);
606 min_events = td->cur_depth;
609 ret = io_u_queued_complete(td, min_events, NULL);
611 cleanup_pending_aio(td);
613 td_set_runstate(td, TD_RUNNING);
615 dprint(FD_VERIFY, "exiting loop\n");
618 static int io_bytes_exceeded(struct thread_data *td)
620 unsigned long long bytes;
623 bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE];
624 else if (td_write(td))
625 bytes = td->this_io_bytes[DDIR_WRITE];
626 else if (td_read(td))
627 bytes = td->this_io_bytes[DDIR_READ];
629 bytes = td->this_io_bytes[DDIR_TRIM];
631 return bytes >= td->o.size;
635 * Main IO worker function. It retrieves io_u's to process and queues
636 * and reaps them, checking for rate and errors along the way.
638 * Returns number of bytes written and trimmed.
640 static uint64_t do_io(struct thread_data *td)
642 uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
645 uint64_t bytes_issued = 0;
647 if (in_ramp_time(td))
648 td_set_runstate(td, TD_RAMP);
650 td_set_runstate(td, TD_RUNNING);
652 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
653 (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td) ||
655 struct timeval comp_time;
661 check_update_rusage(td);
663 if (td->terminate || td->done)
668 if (runtime_exceeded(td, &td->tv_cache)) {
669 __update_tv_cache(td);
670 if (runtime_exceeded(td, &td->tv_cache)) {
676 if (flow_threshold_exceeded(td))
679 if (bytes_issued >= (uint64_t) td->o.size)
689 * Add verification end_io handler if:
690 * - Asked to verify (!td_rw(td))
691 * - Or the io_u is from our verify list (mixed write/ver)
693 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
694 ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
695 if (td->o.verify_async)
696 io_u->end_io = verify_io_u_async;
698 io_u->end_io = verify_io_u;
699 td_set_runstate(td, TD_VERIFYING);
700 } else if (in_ramp_time(td))
701 td_set_runstate(td, TD_RAMP);
703 td_set_runstate(td, TD_RUNNING);
705 ret = td_io_queue(td, io_u);
707 case FIO_Q_COMPLETED:
710 clear_io_u(td, io_u);
711 } else if (io_u->resid) {
712 int bytes = io_u->xfer_buflen - io_u->resid;
713 struct fio_file *f = io_u->file;
715 bytes_issued += bytes;
720 td_verror(td, EIO, "full resid");
725 io_u->xfer_buflen = io_u->resid;
726 io_u->xfer_buf += bytes;
727 io_u->offset += bytes;
729 if (ddir_rw(io_u->ddir))
730 td->ts.short_io_u[io_u->ddir]++;
732 if (io_u->offset == f->real_file_size)
735 requeue_io_u(td, &io_u);
738 if (__should_check_rate(td, DDIR_READ) ||
739 __should_check_rate(td, DDIR_WRITE) ||
740 __should_check_rate(td, DDIR_TRIM))
741 fio_gettime(&comp_time, NULL);
743 ret = io_u_sync_complete(td, io_u, bytes_done);
746 bytes_issued += io_u->xfer_buflen;
751 * if the engine doesn't have a commit hook,
752 * the io_u is really queued. if it does have such
753 * a hook, it has to call io_u_queued() itself.
755 if (td->io_ops->commit == NULL)
756 io_u_queued(td, io_u);
757 bytes_issued += io_u->xfer_buflen;
760 requeue_io_u(td, &io_u);
761 ret2 = td_io_commit(td);
771 if (break_on_this_error(td, ddir, &ret))
775 * See if we need to complete some commands. Note that we
776 * can get BUSY even without IO queued, if the system is
779 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
780 if (full || !td->o.iodepth_batch_complete) {
781 min_evts = min(td->o.iodepth_batch_complete,
784 * if the queue is full, we MUST reap at least 1 event
786 if (full && !min_evts)
789 if (__should_check_rate(td, DDIR_READ) ||
790 __should_check_rate(td, DDIR_WRITE) ||
791 __should_check_rate(td, DDIR_TRIM))
792 fio_gettime(&comp_time, NULL);
795 ret = io_u_queued_complete(td, min_evts, bytes_done);
799 } while (full && (td->cur_depth > td->o.iodepth_low));
804 if (!ddir_rw_sum(bytes_done) && !(td->io_ops->flags & FIO_NOIO))
807 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
808 if (check_min_rate(td, &comp_time, bytes_done)) {
809 if (exitall_on_terminate)
810 fio_terminate_threads(td->groupid);
811 td_verror(td, EIO, "check_min_rate");
816 if (td->o.thinktime) {
817 unsigned long long b;
819 b = ddir_rw_sum(td->io_blocks);
820 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: %lu 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)
881 while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
883 if (td->io_ops->io_u_free)
884 td->io_ops->io_u_free(td, io_u);
886 fio_memfree(io_u, sizeof(*io_u));
891 io_u_rexit(&td->io_u_requeues);
892 io_u_qexit(&td->io_u_freelist);
893 io_u_qexit(&td->io_u_all);
896 static int init_io_u(struct thread_data *td)
899 unsigned int max_bs, min_write;
900 int cl_align, i, max_units;
901 int data_xfer = 1, err;
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)))
914 err += io_u_rinit(&td->io_u_requeues, td->o.iodepth);
915 err += io_u_qinit(&td->io_u_freelist, td->o.iodepth);
916 err += io_u_qinit(&td->io_u_all, td->o.iodepth);
919 log_err("fio: failed setting up IO queues\n");
924 * if we may later need to do address alignment, then add any
925 * possible adjustment here so that we don't cause a buffer
926 * overflow later. this adjustment may be too much if we get
927 * lucky and the allocator gives us an aligned address.
929 if (td->o.odirect || td->o.mem_align || (td->io_ops->flags & FIO_RAWIO))
930 td->orig_buffer_size += page_mask + td->o.mem_align;
932 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
935 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
936 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
939 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
940 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
944 if (data_xfer && allocate_io_mem(td))
947 if (td->o.odirect || td->o.mem_align ||
948 (td->io_ops->flags & FIO_RAWIO))
949 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
953 cl_align = os_cache_line_size();
955 for (i = 0; i < max_units; i++) {
961 ptr = fio_memalign(cl_align, sizeof(*io_u));
963 log_err("fio: unable to allocate aligned memory\n");
968 memset(io_u, 0, sizeof(*io_u));
969 INIT_FLIST_HEAD(&io_u->verify_list);
970 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
974 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
977 io_u_fill_buffer(td, io_u, min_write, max_bs);
978 if (td_write(td) && td->o.verify_pattern_bytes) {
980 * Fill the buffer with the pattern if we are
981 * going to be doing writes.
983 fill_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
988 io_u->flags = IO_U_F_FREE;
989 io_u_qpush(&td->io_u_freelist, io_u);
992 * io_u never leaves this stack, used for iteration of all
995 io_u_qpush(&td->io_u_all, io_u);
997 if (td->io_ops->io_u_init) {
998 int ret = td->io_ops->io_u_init(td, io_u);
1001 log_err("fio: failed to init engine data: %d\n", ret);
1012 static int switch_ioscheduler(struct thread_data *td)
1014 char tmp[256], tmp2[128];
1018 if (td->io_ops->flags & FIO_DISKLESSIO)
1021 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
1023 f = fopen(tmp, "r+");
1025 if (errno == ENOENT) {
1026 log_err("fio: os or kernel doesn't support IO scheduler"
1030 td_verror(td, errno, "fopen iosched");
1037 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1038 if (ferror(f) || ret != 1) {
1039 td_verror(td, errno, "fwrite");
1047 * Read back and check that the selected scheduler is now the default.
1049 ret = fread(tmp, 1, sizeof(tmp), f);
1050 if (ferror(f) || ret < 0) {
1051 td_verror(td, errno, "fread");
1056 sprintf(tmp2, "[%s]", td->o.ioscheduler);
1057 if (!strstr(tmp, tmp2)) {
1058 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1059 td_verror(td, EINVAL, "iosched_switch");
1068 static int keep_running(struct thread_data *td)
1072 if (td->o.time_based)
1079 if (td->o.size != -1ULL && ddir_rw_sum(td->io_bytes) < td->o.size) {
1083 * If the difference is less than the minimum IO size, we
1086 diff = td->o.size - ddir_rw_sum(td->io_bytes);
1087 if (diff < td_max_bs(td))
1096 static int exec_string(struct thread_options *o, const char *string, const char *mode)
1098 int ret, newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
1101 str = malloc(newlen);
1102 sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
1104 log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1107 log_err("fio: exec of cmd <%s> failed\n", str);
1114 * Entry point for the thread based jobs. The process based jobs end up
1115 * here as well, after a little setup.
1117 static void *thread_main(void *data)
1119 unsigned long long elapsed;
1120 struct thread_data *td = data;
1121 struct thread_options *o = &td->o;
1122 pthread_condattr_t attr;
1126 if (!o->use_thread) {
1132 fio_local_clock_init(o->use_thread);
1134 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1137 fio_server_send_start(td);
1139 INIT_FLIST_HEAD(&td->io_log_list);
1140 INIT_FLIST_HEAD(&td->io_hist_list);
1141 INIT_FLIST_HEAD(&td->verify_list);
1142 INIT_FLIST_HEAD(&td->trim_list);
1143 INIT_FLIST_HEAD(&td->next_rand_list);
1144 pthread_mutex_init(&td->io_u_lock, NULL);
1145 td->io_hist_tree = RB_ROOT;
1147 pthread_condattr_init(&attr);
1148 pthread_cond_init(&td->verify_cond, &attr);
1149 pthread_cond_init(&td->free_cond, &attr);
1151 td_set_runstate(td, TD_INITIALIZED);
1152 dprint(FD_MUTEX, "up startup_mutex\n");
1153 fio_mutex_up(startup_mutex);
1154 dprint(FD_MUTEX, "wait on td->mutex\n");
1155 fio_mutex_down(td->mutex);
1156 dprint(FD_MUTEX, "done waiting on td->mutex\n");
1159 * the ->mutex mutex is now no longer used, close it to avoid
1160 * eating a file descriptor
1162 fio_mutex_remove(td->mutex);
1166 * A new gid requires privilege, so we need to do this before setting
1169 if (o->gid != -1U && setgid(o->gid)) {
1170 td_verror(td, errno, "setgid");
1173 if (o->uid != -1U && setuid(o->uid)) {
1174 td_verror(td, errno, "setuid");
1179 * If we have a gettimeofday() thread, make sure we exclude that
1180 * thread from this job
1183 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1186 * Set affinity first, in case it has an impact on the memory
1189 if (o->cpumask_set) {
1190 ret = fio_setaffinity(td->pid, o->cpumask);
1192 td_verror(td, errno, "cpu_set_affinity");
1197 #ifdef CONFIG_LIBNUMA
1198 /* numa node setup */
1199 if (o->numa_cpumask_set || o->numa_memmask_set) {
1202 if (numa_available() < 0) {
1203 td_verror(td, errno, "Does not support NUMA API\n");
1207 if (o->numa_cpumask_set) {
1208 ret = numa_run_on_node_mask(o->numa_cpunodesmask);
1210 td_verror(td, errno, \
1211 "numa_run_on_node_mask failed\n");
1216 if (o->numa_memmask_set) {
1218 switch (o->numa_mem_mode) {
1219 case MPOL_INTERLEAVE:
1220 numa_set_interleave_mask(o->numa_memnodesmask);
1223 numa_set_membind(o->numa_memnodesmask);
1226 numa_set_localalloc();
1228 case MPOL_PREFERRED:
1229 numa_set_preferred(o->numa_mem_prefer_node);
1240 if (fio_pin_memory(td))
1244 * May alter parameters that init_io_u() will use, so we need to
1253 if (o->verify_async && verify_async_init(td))
1257 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1259 td_verror(td, errno, "ioprio_set");
1264 if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1268 if (nice(o->nice) == -1 && errno != 0) {
1269 td_verror(td, errno, "nice");
1273 if (o->ioscheduler && switch_ioscheduler(td))
1276 if (!o->create_serialize && setup_files(td))
1282 if (init_random_map(td))
1285 if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
1289 if (pre_read_files(td) < 0)
1293 fio_verify_init(td);
1295 fio_gettime(&td->epoch, NULL);
1296 fio_getrusage(&td->ru_start);
1298 while (keep_running(td)) {
1299 uint64_t verify_bytes;
1301 fio_gettime(&td->start, NULL);
1302 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1303 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1304 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1306 if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1307 o->ratemin[DDIR_TRIM]) {
1308 memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1309 sizeof(td->bw_sample_time));
1310 memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1311 sizeof(td->bw_sample_time));
1312 memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1313 sizeof(td->bw_sample_time));
1319 prune_io_piece_log(td);
1321 verify_bytes = do_io(td);
1325 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1326 elapsed = utime_since_now(&td->start);
1327 td->ts.runtime[DDIR_READ] += elapsed;
1329 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1330 elapsed = utime_since_now(&td->start);
1331 td->ts.runtime[DDIR_WRITE] += elapsed;
1333 if (td_trim(td) && td->io_bytes[DDIR_TRIM]) {
1334 elapsed = utime_since_now(&td->start);
1335 td->ts.runtime[DDIR_TRIM] += elapsed;
1338 if (td->error || td->terminate)
1341 if (!o->do_verify ||
1342 o->verify == VERIFY_NONE ||
1343 (td->io_ops->flags & FIO_UNIDIR))
1348 fio_gettime(&td->start, NULL);
1350 do_verify(td, verify_bytes);
1352 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1354 if (td->error || td->terminate)
1358 update_rusage_stat(td);
1359 td->ts.runtime[DDIR_READ] = (td->ts.runtime[DDIR_READ] + 999) / 1000;
1360 td->ts.runtime[DDIR_WRITE] = (td->ts.runtime[DDIR_WRITE] + 999) / 1000;
1361 td->ts.runtime[DDIR_TRIM] = (td->ts.runtime[DDIR_TRIM] + 999) / 1000;
1362 td->ts.total_run_time = mtime_since_now(&td->epoch);
1363 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1364 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1365 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1367 fio_unpin_memory(td);
1369 fio_mutex_down(writeout_mutex);
1371 if (o->bw_log_file) {
1372 finish_log_named(td, td->bw_log,
1373 o->bw_log_file, "bw");
1375 finish_log(td, td->bw_log, "bw");
1378 if (o->lat_log_file) {
1379 finish_log_named(td, td->lat_log,
1380 o->lat_log_file, "lat");
1382 finish_log(td, td->lat_log, "lat");
1385 if (o->lat_log_file) {
1386 finish_log_named(td, td->slat_log,
1387 o->lat_log_file, "slat");
1389 finish_log(td, td->slat_log, "slat");
1392 if (o->lat_log_file) {
1393 finish_log_named(td, td->clat_log,
1394 o->lat_log_file, "clat");
1396 finish_log(td, td->clat_log, "clat");
1399 if (o->iops_log_file) {
1400 finish_log_named(td, td->iops_log,
1401 o->iops_log_file, "iops");
1403 finish_log(td, td->iops_log, "iops");
1406 fio_mutex_up(writeout_mutex);
1407 if (o->exec_postrun)
1408 exec_string(o, o->exec_postrun, (const char *)"postrun");
1410 if (exitall_on_terminate)
1411 fio_terminate_threads(td->groupid);
1415 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1418 if (o->verify_async)
1419 verify_async_exit(td);
1421 close_and_free_files(td);
1424 cgroup_shutdown(td, &cgroup_mnt);
1426 if (o->cpumask_set) {
1427 int ret = fio_cpuset_exit(&o->cpumask);
1429 td_verror(td, ret, "fio_cpuset_exit");
1433 * do this very late, it will log file closing as well
1435 if (o->write_iolog_file)
1436 write_iolog_close(td);
1438 fio_mutex_remove(td->rusage_sem);
1439 td->rusage_sem = NULL;
1441 td_set_runstate(td, TD_EXITED);
1442 return (void *) (uintptr_t) td->error;
1447 * We cannot pass the td data into a forked process, so attach the td and
1448 * pass it to the thread worker.
1450 static int fork_main(int shmid, int offset)
1452 struct thread_data *td;
1456 data = shmat(shmid, NULL, 0);
1457 if (data == (void *) -1) {
1465 * HP-UX inherits shm mappings?
1470 td = data + offset * sizeof(struct thread_data);
1471 ret = thread_main(td);
1473 return (int) (uintptr_t) ret;
1477 * Run over the job map and reap the threads that have exited, if any.
1479 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1480 unsigned int *m_rate)
1482 struct thread_data *td;
1483 unsigned int cputhreads, realthreads, pending;
1487 * reap exited threads (TD_EXITED -> TD_REAPED)
1489 realthreads = pending = cputhreads = 0;
1490 for_each_td(td, i) {
1494 * ->io_ops is NULL for a thread that has closed its
1497 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1506 if (td->runstate == TD_REAPED)
1508 if (td->o.use_thread) {
1509 if (td->runstate == TD_EXITED) {
1510 td_set_runstate(td, TD_REAPED);
1517 if (td->runstate == TD_EXITED)
1521 * check if someone quit or got killed in an unusual way
1523 ret = waitpid(td->pid, &status, flags);
1525 if (errno == ECHILD) {
1526 log_err("fio: pid=%d disappeared %d\n",
1527 (int) td->pid, td->runstate);
1529 td_set_runstate(td, TD_REAPED);
1533 } else if (ret == td->pid) {
1534 if (WIFSIGNALED(status)) {
1535 int sig = WTERMSIG(status);
1537 if (sig != SIGTERM && sig != SIGUSR2)
1538 log_err("fio: pid=%d, got signal=%d\n",
1539 (int) td->pid, sig);
1541 td_set_runstate(td, TD_REAPED);
1544 if (WIFEXITED(status)) {
1545 if (WEXITSTATUS(status) && !td->error)
1546 td->error = WEXITSTATUS(status);
1548 td_set_runstate(td, TD_REAPED);
1554 * thread is not dead, continue
1560 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1561 (*t_rate) -= ddir_rw_sum(td->o.rate);
1568 done_secs += mtime_since_now(&td->epoch) / 1000;
1569 profile_td_exit(td);
1572 if (*nr_running == cputhreads && !pending && realthreads)
1573 fio_terminate_threads(TERMINATE_ALL);
1576 static void do_usleep(unsigned int usecs)
1578 check_for_running_stats();
1583 * Main function for kicking off and reaping jobs, as needed.
1585 static void run_threads(void)
1587 struct thread_data *td;
1588 unsigned long spent;
1589 unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
1591 if (fio_gtod_offload && fio_start_gtod_thread())
1594 fio_idle_prof_init();
1598 nr_thread = nr_process = 0;
1599 for_each_td(td, i) {
1600 if (td->o.use_thread)
1606 if (output_format == FIO_OUTPUT_NORMAL) {
1607 log_info("Starting ");
1609 log_info("%d thread%s", nr_thread,
1610 nr_thread > 1 ? "s" : "");
1614 log_info("%d process%s", nr_process,
1615 nr_process > 1 ? "es" : "");
1621 todo = thread_number;
1624 m_rate = t_rate = 0;
1626 for_each_td(td, i) {
1627 print_status_init(td->thread_number - 1);
1629 if (!td->o.create_serialize)
1633 * do file setup here so it happens sequentially,
1634 * we don't want X number of threads getting their
1635 * client data interspersed on disk
1637 if (setup_files(td)) {
1640 log_err("fio: pid=%d, err=%d/%s\n",
1641 (int) td->pid, td->error, td->verror);
1642 td_set_runstate(td, TD_REAPED);
1649 * for sharing to work, each job must always open
1650 * its own files. so close them, if we opened them
1653 for_each_file(td, f, j) {
1654 if (fio_file_open(f))
1655 td_io_close_file(td, f);
1660 /* start idle threads before io threads start to run */
1661 fio_idle_prof_start();
1666 struct thread_data *map[REAL_MAX_JOBS];
1667 struct timeval this_start;
1668 int this_jobs = 0, left;
1671 * create threads (TD_NOT_CREATED -> TD_CREATED)
1673 for_each_td(td, i) {
1674 if (td->runstate != TD_NOT_CREATED)
1678 * never got a chance to start, killed by other
1679 * thread for some reason
1681 if (td->terminate) {
1686 if (td->o.start_delay) {
1687 spent = mtime_since_genesis();
1689 if (td->o.start_delay * 1000 > spent)
1693 if (td->o.stonewall && (nr_started || nr_running)) {
1694 dprint(FD_PROCESS, "%s: stonewall wait\n",
1701 td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED);
1702 td->update_rusage = 0;
1705 * Set state to created. Thread will transition
1706 * to TD_INITIALIZED when it's done setting up.
1708 td_set_runstate(td, TD_CREATED);
1709 map[this_jobs++] = td;
1712 if (td->o.use_thread) {
1715 dprint(FD_PROCESS, "will pthread_create\n");
1716 ret = pthread_create(&td->thread, NULL,
1719 log_err("pthread_create: %s\n",
1724 ret = pthread_detach(td->thread);
1726 log_err("pthread_detach: %s",
1730 dprint(FD_PROCESS, "will fork\n");
1733 int ret = fork_main(shm_id, i);
1736 } else if (i == fio_debug_jobno)
1737 *fio_debug_jobp = pid;
1739 dprint(FD_MUTEX, "wait on startup_mutex\n");
1740 if (fio_mutex_down_timeout(startup_mutex, 10)) {
1741 log_err("fio: job startup hung? exiting.\n");
1742 fio_terminate_threads(TERMINATE_ALL);
1747 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1751 * Wait for the started threads to transition to
1754 fio_gettime(&this_start, NULL);
1756 while (left && !fio_abort) {
1757 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1762 for (i = 0; i < this_jobs; i++) {
1766 if (td->runstate == TD_INITIALIZED) {
1769 } else if (td->runstate >= TD_EXITED) {
1773 nr_running++; /* work-around... */
1779 log_err("fio: %d job%s failed to start\n", left,
1780 left > 1 ? "s" : "");
1781 for (i = 0; i < this_jobs; i++) {
1785 kill(td->pid, SIGTERM);
1791 * start created threads (TD_INITIALIZED -> TD_RUNNING).
1793 for_each_td(td, i) {
1794 if (td->runstate != TD_INITIALIZED)
1797 if (in_ramp_time(td))
1798 td_set_runstate(td, TD_RAMP);
1800 td_set_runstate(td, TD_RUNNING);
1803 m_rate += ddir_rw_sum(td->o.ratemin);
1804 t_rate += ddir_rw_sum(td->o.rate);
1806 fio_mutex_up(td->mutex);
1809 reap_threads(&nr_running, &t_rate, &m_rate);
1815 while (nr_running) {
1816 reap_threads(&nr_running, &t_rate, &m_rate);
1820 fio_idle_prof_stop();
1825 void wait_for_disk_thread_exit(void)
1827 fio_mutex_down(disk_thread_mutex);
1830 static void free_disk_util(void)
1832 disk_util_start_exit();
1833 wait_for_disk_thread_exit();
1834 disk_util_prune_entries();
1837 static void *disk_thread_main(void *data)
1841 fio_mutex_up(startup_mutex);
1843 while (threads && !ret) {
1844 usleep(DISK_UTIL_MSEC * 1000);
1847 ret = update_io_ticks();
1850 print_thread_status();
1853 fio_mutex_up(disk_thread_mutex);
1857 static int create_disk_util_thread(void)
1863 disk_thread_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
1865 ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
1867 fio_mutex_remove(disk_thread_mutex);
1868 log_err("Can't create disk util thread: %s\n", strerror(ret));
1872 ret = pthread_detach(disk_util_thread);
1874 fio_mutex_remove(disk_thread_mutex);
1875 log_err("Can't detatch disk util thread: %s\n", strerror(ret));
1879 dprint(FD_MUTEX, "wait on startup_mutex\n");
1880 fio_mutex_down(startup_mutex);
1881 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1885 int fio_backend(void)
1887 struct thread_data *td;
1891 if (load_profile(exec_profile))
1894 exec_profile = NULL;
1900 setup_log(&agg_io_log[DDIR_READ], 0, IO_LOG_TYPE_BW);
1901 setup_log(&agg_io_log[DDIR_WRITE], 0, IO_LOG_TYPE_BW);
1902 setup_log(&agg_io_log[DDIR_TRIM], 0, IO_LOG_TYPE_BW);
1905 startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
1906 if (startup_mutex == NULL)
1908 writeout_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1909 if (writeout_mutex == NULL)
1914 create_disk_util_thread();
1916 cgroup_list = smalloc(sizeof(*cgroup_list));
1917 INIT_FLIST_HEAD(cgroup_list);
1924 __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1925 __finish_log(agg_io_log[DDIR_WRITE],
1926 "agg-write_bw.log");
1927 __finish_log(agg_io_log[DDIR_TRIM],
1928 "agg-write_bw.log");
1933 fio_options_free(td);
1936 cgroup_kill(cgroup_list);
1940 fio_mutex_remove(startup_mutex);
1941 fio_mutex_remove(writeout_mutex);
1942 fio_mutex_remove(disk_thread_mutex);