static struct fio_mutex *startup_mutex;
static volatile int fio_abort;
static int exit_value;
+static struct itimerval itimer;
+static pthread_t gtod_thread;
struct io_log *agg_io_log[2];
#define TERMINATE_ALL (-1)
#define JOB_START_TIMEOUT (5 * 1000)
-static inline void td_set_runstate(struct thread_data *td, int runstate)
+void td_set_runstate(struct thread_data *td, int runstate)
{
if (td->runstate == runstate)
return;
- dprint(FD_PROCESS, "pid=%d: runstate %d -> %d\n", td->pid, td->runstate,
- runstate);
+ dprint(FD_PROCESS, "pid=%d: runstate %d -> %d\n", (int) td->pid,
+ td->runstate, runstate);
td->runstate = runstate;
}
for_each_td(td, i) {
if (group_id == TERMINATE_ALL || groupid == td->groupid) {
dprint(FD_PROCESS, "setting terminate on %s/%d\n",
- td->o.name, td->pid);
+ td->o.name, (int) td->pid);
td->terminate = 1;
td->o.start_delay = 0;
}
}
-static void sig_handler(int sig)
+static void status_timer_arm(void)
{
- if (!threads)
- return;
+ itimer.it_value.tv_sec = 0;
+ itimer.it_value.tv_usec = DISK_UTIL_MSEC * 1000;
+ setitimer(ITIMER_REAL, &itimer, NULL);
+}
- switch (sig) {
- case SIGALRM:
+static void sig_alrm(int fio_unused sig)
+{
+ if (threads) {
update_io_ticks();
- disk_util_timer_arm();
print_thread_status();
- break;
- default:
+ status_timer_arm();
+ }
+}
+
+/*
+ * Happens on thread runs with ctrl-c, ignore our own SIGQUIT
+ */
+static void sig_quit(int sig)
+{
+}
+
+static void sig_int(int sig)
+{
+ if (threads) {
printf("\nfio: terminating on signal %d\n", sig);
fflush(stdout);
terminate_threads(TERMINATE_ALL);
- break;
}
}
+static void sig_ill(int fio_unused sig)
+{
+ if (!threads)
+ return;
+
+ log_err("fio: illegal instruction. your cpu does not support "
+ "the sse4.2 instruction for crc32c\n");
+ terminate_threads(TERMINATE_ALL);
+ exit(4);
+}
+
+static void set_sig_handlers(void)
+{
+ struct sigaction act;
+
+ memset(&act, 0, sizeof(act));
+ act.sa_handler = sig_alrm;
+ act.sa_flags = SA_RESTART;
+ sigaction(SIGALRM, &act, NULL);
+
+ memset(&act, 0, sizeof(act));
+ act.sa_handler = sig_int;
+ act.sa_flags = SA_RESTART;
+ sigaction(SIGINT, &act, NULL);
+
+ memset(&act, 0, sizeof(act));
+ act.sa_handler = sig_ill;
+ act.sa_flags = SA_RESTART;
+ sigaction(SIGILL, &act, NULL);
+
+ memset(&act, 0, sizeof(act));
+ act.sa_handler = sig_quit;
+ act.sa_flags = SA_RESTART;
+ sigaction(SIGQUIT, &act, NULL);
+}
+
+static inline int should_check_rate(struct thread_data *td)
+{
+ /*
+ * No minimum rate set, always ok
+ */
+ if (!td->o.ratemin && !td->o.rate_iops_min)
+ return 0;
+
+ return 1;
+}
+
/*
* Check if we are above the minimum rate given.
*/
unsigned long spent;
unsigned long rate;
- /*
- * No minimum rate set, always ok
- */
- if (!td->o.ratemin && !td->o.rate_iops_min)
- return 0;
-
/*
* allow a 2 second settle period in the beginning
*/
*/
static void cleanup_pending_aio(struct thread_data *td)
{
- struct list_head *entry, *n;
+ struct flist_head *entry, *n;
struct io_u *io_u;
int r;
* now cancel remaining active events
*/
if (td->io_ops->cancel) {
- list_for_each_safe(entry, n, &td->io_u_busylist) {
- io_u = list_entry(entry, struct io_u, list);
+ flist_for_each_safe(entry, n, &td->io_u_busylist) {
+ io_u = flist_entry(entry, struct io_u, list);
/*
* if the io_u isn't in flight, then that generally
return 0;
}
+static inline void update_tv_cache(struct thread_data *td)
+{
+ if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
+ fio_gettime(&td->tv_cache, NULL);
+}
+
/*
* The main verify engine. Runs over the writes we previously submitted,
* reads the blocks back in, and checks the crc/md5 of the data.
io_u = NULL;
while (!td->terminate) {
- int ret2;
+ int ret2, full;
io_u = __get_io_u(td);
if (!io_u)
break;
- if (runtime_exceeded(td, &io_u->start_time)) {
+ update_tv_cache(td);
+
+ if (runtime_exceeded(td, &td->tv_cache)) {
put_io_u(td, io_u);
td->terminate = 1;
break;
* zero read, fail
*/
if (!bytes) {
- td_verror(td, ENODATA, "full resid");
+ td_verror(td, EIO, "full resid");
put_io_u(td, io_u);
break;
}
* if we can queue more, do so. but check if there are
* completed io_u's first.
*/
- min_events = 0;
- if (queue_full(td) || ret == FIO_Q_BUSY) {
- if (td->cur_depth >= td->o.iodepth_low)
- min_events = td->cur_depth - td->o.iodepth_low;
- if (!min_events)
+ full = queue_full(td) || ret == FIO_Q_BUSY;
+ if (full || !td->o.iodepth_batch_complete) {
+ min_events = td->o.iodepth_batch_complete;
+ if (full && !min_events)
min_events = 1;
- }
- /*
- * Reap required number of io units, if any, and do the
- * verification on them through the callback handler
- */
- if (io_u_queued_complete(td, min_events) < 0)
+ do {
+ /*
+ * Reap required number of io units, if any,
+ * and do the verification on them through
+ * the callback handler
+ */
+ if (io_u_queued_complete(td, min_events) < 0) {
+ ret = -1;
+ break;
+ }
+ } while (full && (td->cur_depth > td->o.iodepth_low));
+ }
+ if (ret < 0)
break;
}
*/
static void do_io(struct thread_data *td)
{
- struct timeval s;
unsigned long usec;
unsigned int i;
int ret = 0;
- td_set_runstate(td, TD_RUNNING);
+ if (in_ramp_time(td))
+ td_set_runstate(td, TD_RAMP);
+ else
+ td_set_runstate(td, TD_RUNNING);
while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) {
struct timeval comp_time;
long bytes_done = 0;
int min_evts = 0;
struct io_u *io_u;
- int ret2;
+ int ret2, full;
if (td->terminate)
break;
if (!io_u)
break;
- memcpy(&s, &io_u->start_time, sizeof(s));
+ update_tv_cache(td);
- if (runtime_exceeded(td, &s)) {
+ if (runtime_exceeded(td, &td->tv_cache)) {
put_io_u(td, io_u);
td->terminate = 1;
break;
if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ) {
io_u->end_io = verify_io_u;
td_set_runstate(td, TD_VERIFYING);
- } else
+ } else if (in_ramp_time(td))
+ td_set_runstate(td, TD_RAMP);
+ else
td_set_runstate(td, TD_RUNNING);
ret = td_io_queue(td, io_u);
* zero read, fail
*/
if (!bytes) {
- td_verror(td, ENODATA, "full resid");
+ td_verror(td, EIO, "full resid");
put_io_u(td, io_u);
break;
}
requeue_io_u(td, &io_u);
} else {
sync_done:
- fio_gettime(&comp_time, NULL);
+ if (should_check_rate(td))
+ fio_gettime(&comp_time, NULL);
+
bytes_done = io_u_sync_complete(td, io_u);
if (bytes_done < 0)
ret = bytes_done;
/*
* See if we need to complete some commands
*/
- if (queue_full(td) || ret == FIO_Q_BUSY) {
- min_evts = 0;
- if (td->cur_depth >= td->o.iodepth_low)
- min_evts = td->cur_depth - td->o.iodepth_low;
- if (!min_evts)
+ full = queue_full(td) || ret == FIO_Q_BUSY;
+ if (full || !td->o.iodepth_batch_complete) {
+ min_evts = td->o.iodepth_batch_complete;
+ if (full && !min_evts)
min_evts = 1;
- fio_gettime(&comp_time, NULL);
- bytes_done = io_u_queued_complete(td, min_evts);
- if (bytes_done < 0)
- break;
+
+ if (should_check_rate(td))
+ fio_gettime(&comp_time, NULL);
+
+ do {
+ ret = io_u_queued_complete(td, min_evts);
+ if (ret <= 0)
+ break;
+
+ bytes_done += ret;
+ } while (full && (td->cur_depth > td->o.iodepth_low));
}
+ if (ret < 0)
+ break;
if (!bytes_done)
continue;
* of completions except the very first one which may look
* a little bursty
*/
- usec = utime_since(&s, &comp_time);
+ if (!in_ramp_time(td) && should_check_rate(td)) {
+ usec = utime_since(&td->tv_cache, &comp_time);
- rate_throttle(td, usec, bytes_done);
+ rate_throttle(td, usec, bytes_done);
- if (check_min_rate(td, &comp_time)) {
- if (exitall_on_terminate)
- terminate_threads(td->groupid);
- td_verror(td, ENODATA, "check_min_rate");
- break;
+ if (check_min_rate(td, &comp_time)) {
+ if (exitall_on_terminate)
+ terminate_threads(td->groupid);
+ td_verror(td, EIO, "check_min_rate");
+ break;
+ }
}
if (td->o.thinktime) {
static void cleanup_io_u(struct thread_data *td)
{
- struct list_head *entry, *n;
+ struct flist_head *entry, *n;
struct io_u *io_u;
- list_for_each_safe(entry, n, &td->io_u_freelist) {
- io_u = list_entry(entry, struct io_u, list);
+ flist_for_each_safe(entry, n, &td->io_u_freelist) {
+ io_u = flist_entry(entry, struct io_u, list);
- list_del(&io_u->list);
+ flist_del(&io_u->list);
free(io_u);
}
return 1;
io_u = malloc(sizeof(*io_u));
memset(io_u, 0, sizeof(*io_u));
- INIT_LIST_HEAD(&io_u->list);
+ INIT_FLIST_HEAD(&io_u->list);
if (!(td->io_ops->flags & FIO_NOIO)) {
io_u->buf = p + max_bs * i;
io_u->index = i;
io_u->flags = IO_U_F_FREE;
- list_add(&io_u->list, &td->io_u_freelist);
+ flist_add(&io_u->list, &td->io_u_freelist);
}
- io_u_init_timeout();
-
return 0;
}
return 0;
}
-static int clear_io_state(struct thread_data *td)
+static void reset_io_counters(struct thread_data *td)
{
- struct fio_file *f;
- unsigned int i;
- int ret;
-
td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
td->zone_bytes = 0;
*/
if (td->o.time_based || td->o.loops)
td->nr_done_files = 0;
+}
+
+void reset_all_stats(struct thread_data *td)
+{
+ struct timeval tv;
+ int i;
+
+ reset_io_counters(td);
+
+ for (i = 0; i < 2; i++) {
+ td->io_bytes[i] = 0;
+ td->io_blocks[i] = 0;
+ td->io_issues[i] = 0;
+ td->ts.total_io_u[i] = 0;
+ }
+
+ fio_gettime(&tv, NULL);
+ memcpy(&td->epoch, &tv, sizeof(tv));
+ memcpy(&td->start, &tv, sizeof(tv));
+}
+
+static int clear_io_state(struct thread_data *td)
+{
+ struct fio_file *f;
+ unsigned int i;
+ int ret;
+
+ reset_io_counters(td);
close_files(td);
td->pid = getpid();
- dprint(FD_PROCESS, "jobs pid=%d started\n", td->pid);
+ dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
- INIT_LIST_HEAD(&td->io_u_freelist);
- INIT_LIST_HEAD(&td->io_u_busylist);
- INIT_LIST_HEAD(&td->io_u_requeues);
- INIT_LIST_HEAD(&td->io_log_list);
- INIT_LIST_HEAD(&td->io_hist_list);
+ INIT_FLIST_HEAD(&td->io_u_freelist);
+ INIT_FLIST_HEAD(&td->io_u_busylist);
+ INIT_FLIST_HEAD(&td->io_u_requeues);
+ INIT_FLIST_HEAD(&td->io_log_list);
+ INIT_FLIST_HEAD(&td->io_hist_list);
td->io_hist_tree = RB_ROOT;
td_set_runstate(td, TD_INITIALIZED);
goto err;
}
+ if (td->o.gtod_cpu) {
+ if (fio_getaffinity(td->pid, &td->o.cpumask) == -1) {
+ td_verror(td, errno, "cpu_get_affinity");
+ goto err;
+ }
+ fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
+ if (fio_setaffinity(td) == -1) {
+ td_verror(td, errno, "cpu_set_affinity");
+ goto err;
+ }
+ }
+
if (td->ioprio_set) {
if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
td_verror(td, errno, "ioprio_set");
}
fio_gettime(&td->epoch, NULL);
- memcpy(&td->timeout_end, &td->epoch, sizeof(td->epoch));
getrusage(RUSAGE_SELF, &td->ts.ru_start);
runtime[0] = runtime[1] = 0;
while (keep_running(td)) {
fio_gettime(&td->start, NULL);
memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
+ memcpy(&td->tv_cache, &td->start, sizeof(td->start));
if (td->o.ratemin)
memcpy(&td->lastrate, &td->ts.stat_sample_time,
td->ts.io_bytes[0] = td->io_bytes[0];
td->ts.io_bytes[1] = td->io_bytes[1];
- if (td->ts.bw_log)
- finish_log(td, td->ts.bw_log, "bw");
- if (td->ts.slat_log)
- finish_log(td, td->ts.slat_log, "slat");
- if (td->ts.clat_log)
- finish_log(td, td->ts.clat_log, "clat");
+ if (td->ts.bw_log) {
+ if (td->o.bw_log_file)
+ finish_log_named(td, td->ts.bw_log, td->o.bw_log_file, "bw");
+ else
+ finish_log(td, td->ts.bw_log, "bw");
+ }
+ if (td->ts.slat_log) {
+ if (td->o.lat_log_file)
+ finish_log_named(td, td->ts.slat_log, td->o.lat_log_file, "clat");
+ else
+ finish_log(td, td->ts.slat_log, "slat");
+ }
+ if (td->ts.clat_log) {
+ if (td->o.lat_log_file)
+ finish_log_named(td, td->ts.clat_log, td->o.lat_log_file, "clat");
+ else
+ finish_log(td, td->ts.clat_log, "clat");
+ }
if (td->o.exec_postrun) {
if (system(td->o.exec_postrun) < 0)
log_err("fio: postrun %s failed\n", td->o.exec_postrun);
err:
if (td->error)
- printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error,
+ printf("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
td->verror);
close_and_free_files(td);
close_ioengine(td);
ret = waitpid(td->pid, &status, flags);
if (ret < 0) {
if (errno == ECHILD) {
- log_err("fio: pid=%d disappeared %d\n", td->pid,
- td->runstate);
+ log_err("fio: pid=%d disappeared %d\n",
+ (int) td->pid, td->runstate);
td_set_runstate(td, TD_REAPED);
goto reaped;
}
if (sig != SIGQUIT)
log_err("fio: pid=%d, got signal=%d\n",
- td->pid, sig);
+ (int) td->pid, sig);
td_set_runstate(td, TD_REAPED);
goto reaped;
}
terminate_threads(TERMINATE_ALL);
}
+static void *gtod_thread_main(void *data)
+{
+ fio_mutex_up(startup_mutex);
+
+ /*
+ * As long as we have jobs around, update the clock. It would be nice
+ * to have some way of NOT hammering that CPU with gettimeofday(),
+ * but I'm not sure what to use outside of a simple CPU nop to relax
+ * it - we don't want to lose precision.
+ */
+ while (threads) {
+ fio_gtod_update();
+ nop;
+ }
+
+ return NULL;
+}
+
+static int fio_start_gtod_thread(void)
+{
+ if (pthread_create(>od_thread, NULL, gtod_thread_main, NULL)) {
+ perror("Can't create gtod thread");
+ return 1;
+ }
+ if (pthread_detach(gtod_thread) < 0) {
+ perror("Can't detatch gtod thread");
+ return 1;
+ }
+
+ fio_mutex_down(startup_mutex);
+ return 0;
+}
+
/*
* Main function for kicking off and reaping jobs, as needed.
*/
if (fio_pin_memory())
return;
+ if (fio_gtod_offload && fio_start_gtod_thread())
+ return;
+
if (!terse_output) {
printf("Starting ");
if (nr_thread)
fflush(stdout);
}
- signal(SIGINT, sig_handler);
- signal(SIGALRM, sig_handler);
+ set_sig_handlers();
todo = thread_number;
nr_running = 0;
if (setup_files(td)) {
exit_value++;
if (td->error)
- log_err("fio: pid=%d, err=%d/%s\n", td->pid,
- td->error, td->verror);
+ log_err("fio: pid=%d, err=%d/%s\n",
+ (int) td->pid, td->error, td->verror);
td_set_runstate(td, TD_REAPED);
todo--;
} else {
if (td->runstate != TD_INITIALIZED)
continue;
- td_set_runstate(td, TD_RUNNING);
+ if (in_ramp_time(td))
+ td_set_runstate(td, TD_RAMP);
+ else
+ td_set_runstate(td, TD_RUNNING);
nr_running++;
nr_started--;
m_rate += td->o.ratemin;
set_genesis_time();
- disk_util_timer_arm();
+ status_timer_arm();
run_threads();
projects / fio.git / blobdiff