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)
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)) {
827 if (td->o.thinktime_spin)
828 usec_spin(td->o.thinktime_spin);
830 left = td->o.thinktime - td->o.thinktime_spin;
832 usec_sleep(td, left);
837 check_update_rusage(td);
839 if (td->trim_entries)
840 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
842 if (td->o.fill_device && td->error == ENOSPC) {
851 ret = io_u_queued_complete(td, i, bytes_done);
852 if (td->o.fill_device && td->error == ENOSPC)
856 if (should_fsync(td) && td->o.end_fsync) {
857 td_set_runstate(td, TD_FSYNCING);
859 for_each_file(td, f, i) {
860 if (!fio_file_fsync(td, f))
863 log_err("fio: end_fsync failed for file %s\n",
868 cleanup_pending_aio(td);
871 * stop job if we failed doing any IO
873 if (!ddir_rw_sum(td->this_io_bytes))
876 return bytes_done[DDIR_WRITE] + bytes_done[DDIR_TRIM];
879 static void cleanup_io_u(struct thread_data *td)
881 struct flist_head *entry, *n;
884 flist_for_each_safe(entry, n, &td->io_u_freelist) {
885 io_u = flist_entry(entry, struct io_u, list);
887 flist_del(&io_u->list);
889 if (td->io_ops->io_u_free)
890 td->io_ops->io_u_free(td, io_u);
892 fio_memfree(io_u, sizeof(*io_u));
898 static int init_io_u(struct thread_data *td)
901 unsigned int max_bs, min_write;
902 int cl_align, i, max_units;
906 max_units = td->o.iodepth;
907 max_bs = td_max_bs(td);
908 min_write = td->o.min_bs[DDIR_WRITE];
909 td->orig_buffer_size = (unsigned long long) max_bs
910 * (unsigned long long) max_units;
912 if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
916 * if we may later need to do address alignment, then add any
917 * possible adjustment here so that we don't cause a buffer
918 * overflow later. this adjustment may be too much if we get
919 * lucky and the allocator gives us an aligned address.
921 if (td->o.odirect || td->o.mem_align || (td->io_ops->flags & FIO_RAWIO))
922 td->orig_buffer_size += page_mask + td->o.mem_align;
924 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
927 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
928 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
931 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
932 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
936 if (data_xfer && allocate_io_mem(td))
939 if (td->o.odirect || td->o.mem_align ||
940 (td->io_ops->flags & FIO_RAWIO))
941 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
945 cl_align = os_cache_line_size();
947 for (i = 0; i < max_units; i++) {
953 ptr = fio_memalign(cl_align, sizeof(*io_u));
955 log_err("fio: unable to allocate aligned memory\n");
960 memset(io_u, 0, sizeof(*io_u));
961 INIT_FLIST_HEAD(&io_u->list);
962 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
966 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
969 io_u_fill_buffer(td, io_u, min_write, max_bs);
970 if (td_write(td) && td->o.verify_pattern_bytes) {
972 * Fill the buffer with the pattern if we are
973 * going to be doing writes.
975 fill_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
980 io_u->flags = IO_U_F_FREE;
981 flist_add(&io_u->list, &td->io_u_freelist);
983 if (td->io_ops->io_u_init) {
984 int ret = td->io_ops->io_u_init(td, io_u);
987 log_err("fio: failed to init engine data: %d\n", ret);
998 static int switch_ioscheduler(struct thread_data *td)
1000 char tmp[256], tmp2[128];
1004 if (td->io_ops->flags & FIO_DISKLESSIO)
1007 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
1009 f = fopen(tmp, "r+");
1011 if (errno == ENOENT) {
1012 log_err("fio: os or kernel doesn't support IO scheduler"
1016 td_verror(td, errno, "fopen iosched");
1023 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1024 if (ferror(f) || ret != 1) {
1025 td_verror(td, errno, "fwrite");
1033 * Read back and check that the selected scheduler is now the default.
1035 ret = fread(tmp, 1, sizeof(tmp), f);
1036 if (ferror(f) || ret < 0) {
1037 td_verror(td, errno, "fread");
1042 sprintf(tmp2, "[%s]", td->o.ioscheduler);
1043 if (!strstr(tmp, tmp2)) {
1044 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1045 td_verror(td, EINVAL, "iosched_switch");
1054 static int keep_running(struct thread_data *td)
1058 if (td->o.time_based)
1065 if (td->o.size != -1ULL && ddir_rw_sum(td->io_bytes) < td->o.size) {
1069 * If the difference is less than the minimum IO size, we
1072 diff = td->o.size - ddir_rw_sum(td->io_bytes);
1073 if (diff < td_max_bs(td))
1082 static int exec_string(const char *string)
1084 int ret, newlen = strlen(string) + 1 + 8;
1087 str = malloc(newlen);
1088 sprintf(str, "sh -c %s", string);
1092 log_err("fio: exec of cmd <%s> failed\n", str);
1099 * Entry point for the thread based jobs. The process based jobs end up
1100 * here as well, after a little setup.
1102 static void *thread_main(void *data)
1104 unsigned long long elapsed;
1105 struct thread_data *td = data;
1106 struct thread_options *o = &td->o;
1107 pthread_condattr_t attr;
1111 if (!o->use_thread) {
1117 fio_local_clock_init(o->use_thread);
1119 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1122 fio_server_send_start(td);
1124 INIT_FLIST_HEAD(&td->io_u_freelist);
1125 INIT_FLIST_HEAD(&td->io_u_busylist);
1126 INIT_FLIST_HEAD(&td->io_u_requeues);
1127 INIT_FLIST_HEAD(&td->io_log_list);
1128 INIT_FLIST_HEAD(&td->io_hist_list);
1129 INIT_FLIST_HEAD(&td->verify_list);
1130 INIT_FLIST_HEAD(&td->trim_list);
1131 INIT_FLIST_HEAD(&td->next_rand_list);
1132 pthread_mutex_init(&td->io_u_lock, NULL);
1133 td->io_hist_tree = RB_ROOT;
1135 pthread_condattr_init(&attr);
1136 pthread_cond_init(&td->verify_cond, &attr);
1137 pthread_cond_init(&td->free_cond, &attr);
1139 td_set_runstate(td, TD_INITIALIZED);
1140 dprint(FD_MUTEX, "up startup_mutex\n");
1141 fio_mutex_up(startup_mutex);
1142 dprint(FD_MUTEX, "wait on td->mutex\n");
1143 fio_mutex_down(td->mutex);
1144 dprint(FD_MUTEX, "done waiting on td->mutex\n");
1147 * the ->mutex mutex is now no longer used, close it to avoid
1148 * eating a file descriptor
1150 fio_mutex_remove(td->mutex);
1154 * A new gid requires privilege, so we need to do this before setting
1157 if (o->gid != -1U && setgid(o->gid)) {
1158 td_verror(td, errno, "setgid");
1161 if (o->uid != -1U && setuid(o->uid)) {
1162 td_verror(td, errno, "setuid");
1167 * If we have a gettimeofday() thread, make sure we exclude that
1168 * thread from this job
1171 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1174 * Set affinity first, in case it has an impact on the memory
1177 if (o->cpumask_set) {
1178 ret = fio_setaffinity(td->pid, o->cpumask);
1180 td_verror(td, errno, "cpu_set_affinity");
1185 #ifdef CONFIG_LIBNUMA
1186 /* numa node setup */
1187 if (o->numa_cpumask_set || o->numa_memmask_set) {
1190 if (numa_available() < 0) {
1191 td_verror(td, errno, "Does not support NUMA API\n");
1195 if (o->numa_cpumask_set) {
1196 ret = numa_run_on_node_mask(o->numa_cpunodesmask);
1198 td_verror(td, errno, \
1199 "numa_run_on_node_mask failed\n");
1204 if (o->numa_memmask_set) {
1206 switch (o->numa_mem_mode) {
1207 case MPOL_INTERLEAVE:
1208 numa_set_interleave_mask(o->numa_memnodesmask);
1211 numa_set_membind(o->numa_memnodesmask);
1214 numa_set_localalloc();
1216 case MPOL_PREFERRED:
1217 numa_set_preferred(o->numa_mem_prefer_node);
1228 if (fio_pin_memory(td))
1232 * May alter parameters that init_io_u() will use, so we need to
1241 if (o->verify_async && verify_async_init(td))
1245 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1247 td_verror(td, errno, "ioprio_set");
1252 if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1256 if (nice(o->nice) == -1 && errno != 0) {
1257 td_verror(td, errno, "nice");
1261 if (o->ioscheduler && switch_ioscheduler(td))
1264 if (!o->create_serialize && setup_files(td))
1270 if (init_random_map(td))
1273 if (o->exec_prerun && exec_string(o->exec_prerun))
1277 if (pre_read_files(td) < 0)
1281 fio_verify_init(td);
1283 fio_gettime(&td->epoch, NULL);
1284 fio_getrusage(&td->ru_start);
1286 while (keep_running(td)) {
1287 uint64_t verify_bytes;
1289 fio_gettime(&td->start, NULL);
1290 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1291 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1292 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1294 if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1295 o->ratemin[DDIR_TRIM]) {
1296 memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1297 sizeof(td->bw_sample_time));
1298 memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1299 sizeof(td->bw_sample_time));
1300 memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1301 sizeof(td->bw_sample_time));
1307 prune_io_piece_log(td);
1309 verify_bytes = do_io(td);
1313 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1314 elapsed = utime_since_now(&td->start);
1315 td->ts.runtime[DDIR_READ] += elapsed;
1317 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1318 elapsed = utime_since_now(&td->start);
1319 td->ts.runtime[DDIR_WRITE] += elapsed;
1321 if (td_trim(td) && td->io_bytes[DDIR_TRIM]) {
1322 elapsed = utime_since_now(&td->start);
1323 td->ts.runtime[DDIR_TRIM] += elapsed;
1326 if (td->error || td->terminate)
1329 if (!o->do_verify ||
1330 o->verify == VERIFY_NONE ||
1331 (td->io_ops->flags & FIO_UNIDIR))
1336 fio_gettime(&td->start, NULL);
1338 do_verify(td, verify_bytes);
1340 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1342 if (td->error || td->terminate)
1346 update_rusage_stat(td);
1347 td->ts.runtime[DDIR_READ] = (td->ts.runtime[DDIR_READ] + 999) / 1000;
1348 td->ts.runtime[DDIR_WRITE] = (td->ts.runtime[DDIR_WRITE] + 999) / 1000;
1349 td->ts.runtime[DDIR_TRIM] = (td->ts.runtime[DDIR_TRIM] + 999) / 1000;
1350 td->ts.total_run_time = mtime_since_now(&td->epoch);
1351 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1352 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1353 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1355 fio_unpin_memory(td);
1357 fio_mutex_down(writeout_mutex);
1359 if (o->bw_log_file) {
1360 finish_log_named(td, td->bw_log,
1361 o->bw_log_file, "bw");
1363 finish_log(td, td->bw_log, "bw");
1366 if (o->lat_log_file) {
1367 finish_log_named(td, td->lat_log,
1368 o->lat_log_file, "lat");
1370 finish_log(td, td->lat_log, "lat");
1373 if (o->lat_log_file) {
1374 finish_log_named(td, td->slat_log,
1375 o->lat_log_file, "slat");
1377 finish_log(td, td->slat_log, "slat");
1380 if (o->lat_log_file) {
1381 finish_log_named(td, td->clat_log,
1382 o->lat_log_file, "clat");
1384 finish_log(td, td->clat_log, "clat");
1387 if (o->iops_log_file) {
1388 finish_log_named(td, td->iops_log,
1389 o->iops_log_file, "iops");
1391 finish_log(td, td->iops_log, "iops");
1394 fio_mutex_up(writeout_mutex);
1395 if (o->exec_postrun)
1396 exec_string(o->exec_postrun);
1398 if (exitall_on_terminate)
1399 fio_terminate_threads(td->groupid);
1403 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1406 if (o->verify_async)
1407 verify_async_exit(td);
1409 close_and_free_files(td);
1412 cgroup_shutdown(td, &cgroup_mnt);
1414 if (o->cpumask_set) {
1415 int ret = fio_cpuset_exit(&o->cpumask);
1417 td_verror(td, ret, "fio_cpuset_exit");
1421 * do this very late, it will log file closing as well
1423 if (o->write_iolog_file)
1424 write_iolog_close(td);
1426 fio_mutex_remove(td->rusage_sem);
1427 td->rusage_sem = NULL;
1429 td_set_runstate(td, TD_EXITED);
1430 return (void *) (uintptr_t) td->error;
1435 * We cannot pass the td data into a forked process, so attach the td and
1436 * pass it to the thread worker.
1438 static int fork_main(int shmid, int offset)
1440 struct thread_data *td;
1444 data = shmat(shmid, NULL, 0);
1445 if (data == (void *) -1) {
1453 * HP-UX inherits shm mappings?
1458 td = data + offset * sizeof(struct thread_data);
1459 ret = thread_main(td);
1461 return (int) (uintptr_t) ret;
1465 * Run over the job map and reap the threads that have exited, if any.
1467 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1468 unsigned int *m_rate)
1470 struct thread_data *td;
1471 unsigned int cputhreads, realthreads, pending;
1475 * reap exited threads (TD_EXITED -> TD_REAPED)
1477 realthreads = pending = cputhreads = 0;
1478 for_each_td(td, i) {
1482 * ->io_ops is NULL for a thread that has closed its
1485 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1494 if (td->runstate == TD_REAPED)
1496 if (td->o.use_thread) {
1497 if (td->runstate == TD_EXITED) {
1498 td_set_runstate(td, TD_REAPED);
1505 if (td->runstate == TD_EXITED)
1509 * check if someone quit or got killed in an unusual way
1511 ret = waitpid(td->pid, &status, flags);
1513 if (errno == ECHILD) {
1514 log_err("fio: pid=%d disappeared %d\n",
1515 (int) td->pid, td->runstate);
1517 td_set_runstate(td, TD_REAPED);
1521 } else if (ret == td->pid) {
1522 if (WIFSIGNALED(status)) {
1523 int sig = WTERMSIG(status);
1525 if (sig != SIGTERM && sig != SIGUSR2)
1526 log_err("fio: pid=%d, got signal=%d\n",
1527 (int) td->pid, sig);
1529 td_set_runstate(td, TD_REAPED);
1532 if (WIFEXITED(status)) {
1533 if (WEXITSTATUS(status) && !td->error)
1534 td->error = WEXITSTATUS(status);
1536 td_set_runstate(td, TD_REAPED);
1542 * thread is not dead, continue
1548 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1549 (*t_rate) -= ddir_rw_sum(td->o.rate);
1556 done_secs += mtime_since_now(&td->epoch) / 1000;
1559 if (*nr_running == cputhreads && !pending && realthreads)
1560 fio_terminate_threads(TERMINATE_ALL);
1563 static void do_usleep(unsigned int usecs)
1565 check_for_running_stats();
1570 * Main function for kicking off and reaping jobs, as needed.
1572 static void run_threads(void)
1574 struct thread_data *td;
1575 unsigned long spent;
1576 unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
1578 if (fio_gtod_offload && fio_start_gtod_thread())
1581 fio_idle_prof_init();
1585 nr_thread = nr_process = 0;
1586 for_each_td(td, i) {
1587 if (td->o.use_thread)
1593 if (output_format == FIO_OUTPUT_NORMAL) {
1594 log_info("Starting ");
1596 log_info("%d thread%s", nr_thread,
1597 nr_thread > 1 ? "s" : "");
1601 log_info("%d process%s", nr_process,
1602 nr_process > 1 ? "es" : "");
1608 todo = thread_number;
1611 m_rate = t_rate = 0;
1613 for_each_td(td, i) {
1614 print_status_init(td->thread_number - 1);
1616 if (!td->o.create_serialize)
1620 * do file setup here so it happens sequentially,
1621 * we don't want X number of threads getting their
1622 * client data interspersed on disk
1624 if (setup_files(td)) {
1627 log_err("fio: pid=%d, err=%d/%s\n",
1628 (int) td->pid, td->error, td->verror);
1629 td_set_runstate(td, TD_REAPED);
1636 * for sharing to work, each job must always open
1637 * its own files. so close them, if we opened them
1640 for_each_file(td, f, j) {
1641 if (fio_file_open(f))
1642 td_io_close_file(td, f);
1647 /* start idle threads before io threads start to run */
1648 fio_idle_prof_start();
1653 struct thread_data *map[REAL_MAX_JOBS];
1654 struct timeval this_start;
1655 int this_jobs = 0, left;
1658 * create threads (TD_NOT_CREATED -> TD_CREATED)
1660 for_each_td(td, i) {
1661 if (td->runstate != TD_NOT_CREATED)
1665 * never got a chance to start, killed by other
1666 * thread for some reason
1668 if (td->terminate) {
1673 if (td->o.start_delay) {
1674 spent = mtime_since_genesis();
1676 if (td->o.start_delay * 1000 > spent)
1680 if (td->o.stonewall && (nr_started || nr_running)) {
1681 dprint(FD_PROCESS, "%s: stonewall wait\n",
1688 td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED);
1689 td->update_rusage = 0;
1692 * Set state to created. Thread will transition
1693 * to TD_INITIALIZED when it's done setting up.
1695 td_set_runstate(td, TD_CREATED);
1696 map[this_jobs++] = td;
1699 if (td->o.use_thread) {
1702 dprint(FD_PROCESS, "will pthread_create\n");
1703 ret = pthread_create(&td->thread, NULL,
1706 log_err("pthread_create: %s\n",
1711 ret = pthread_detach(td->thread);
1713 log_err("pthread_detach: %s",
1717 dprint(FD_PROCESS, "will fork\n");
1720 int ret = fork_main(shm_id, i);
1723 } else if (i == fio_debug_jobno)
1724 *fio_debug_jobp = pid;
1726 dprint(FD_MUTEX, "wait on startup_mutex\n");
1727 if (fio_mutex_down_timeout(startup_mutex, 10)) {
1728 log_err("fio: job startup hung? exiting.\n");
1729 fio_terminate_threads(TERMINATE_ALL);
1734 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1738 * Wait for the started threads to transition to
1741 fio_gettime(&this_start, NULL);
1743 while (left && !fio_abort) {
1744 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1749 for (i = 0; i < this_jobs; i++) {
1753 if (td->runstate == TD_INITIALIZED) {
1756 } else if (td->runstate >= TD_EXITED) {
1760 nr_running++; /* work-around... */
1766 log_err("fio: %d job%s failed to start\n", left,
1767 left > 1 ? "s" : "");
1768 for (i = 0; i < this_jobs; i++) {
1772 kill(td->pid, SIGTERM);
1778 * start created threads (TD_INITIALIZED -> TD_RUNNING).
1780 for_each_td(td, i) {
1781 if (td->runstate != TD_INITIALIZED)
1784 if (in_ramp_time(td))
1785 td_set_runstate(td, TD_RAMP);
1787 td_set_runstate(td, TD_RUNNING);
1790 m_rate += ddir_rw_sum(td->o.ratemin);
1791 t_rate += ddir_rw_sum(td->o.rate);
1793 fio_mutex_up(td->mutex);
1796 reap_threads(&nr_running, &t_rate, &m_rate);
1802 while (nr_running) {
1803 reap_threads(&nr_running, &t_rate, &m_rate);
1807 fio_idle_prof_stop();
1812 void wait_for_disk_thread_exit(void)
1814 fio_mutex_down(disk_thread_mutex);
1817 static void free_disk_util(void)
1819 disk_util_start_exit();
1820 wait_for_disk_thread_exit();
1821 disk_util_prune_entries();
1824 static void *disk_thread_main(void *data)
1828 fio_mutex_up(startup_mutex);
1830 while (threads && !ret) {
1831 usleep(DISK_UTIL_MSEC * 1000);
1834 ret = update_io_ticks();
1837 print_thread_status();
1840 fio_mutex_up(disk_thread_mutex);
1844 static int create_disk_util_thread(void)
1850 disk_thread_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
1852 ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
1854 fio_mutex_remove(disk_thread_mutex);
1855 log_err("Can't create disk util thread: %s\n", strerror(ret));
1859 ret = pthread_detach(disk_util_thread);
1861 fio_mutex_remove(disk_thread_mutex);
1862 log_err("Can't detatch disk util thread: %s\n", strerror(ret));
1866 dprint(FD_MUTEX, "wait on startup_mutex\n");
1867 fio_mutex_down(startup_mutex);
1868 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1872 int fio_backend(void)
1874 struct thread_data *td;
1878 if (load_profile(exec_profile))
1881 exec_profile = NULL;
1887 setup_log(&agg_io_log[DDIR_READ], 0, IO_LOG_TYPE_BW);
1888 setup_log(&agg_io_log[DDIR_WRITE], 0, IO_LOG_TYPE_BW);
1889 setup_log(&agg_io_log[DDIR_TRIM], 0, IO_LOG_TYPE_BW);
1892 startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
1893 if (startup_mutex == NULL)
1895 writeout_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1896 if (writeout_mutex == NULL)
1901 create_disk_util_thread();
1903 cgroup_list = smalloc(sizeof(*cgroup_list));
1904 INIT_FLIST_HEAD(cgroup_list);
1911 __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1912 __finish_log(agg_io_log[DDIR_WRITE],
1913 "agg-write_bw.log");
1914 __finish_log(agg_io_log[DDIR_TRIM],
1915 "agg-write_bw.log");
1920 fio_options_free(td);
1923 cgroup_kill(cgroup_list);
1927 fio_mutex_remove(startup_mutex);
1928 fio_mutex_remove(writeout_mutex);
1929 fio_mutex_remove(disk_thread_mutex);