stalld: first attempt at emulating DEADLINE using FIFO
This is a first cut at the logic to do stalled thread boosting using
SCHED_FIFO, when running on a kernel without SCHED_DEADLINE support.
The idea is that in the boost logic, we change the starving thread's
policy to SCHED_FIFO:98, then sleep up to the 'period' boundary (same
period used for DEADLINE), then change the thread back to it's original
policy and parameters.
Signed-off-by: Clark Williams <williams@redhat.com>
diff --git a/src/stalld.c b/src/stalld.c
index c9aaa36..7737723 100644
--- a/src/stalld.c
+++ b/src/stalld.c
@@ -40,6 +40,11 @@
#define MAX_WAITING_PIDS 30
/*
+ * flag indicating that the deadline scheduler is available
+ */
+int use_deadline;
+
+/*
* See kernel/sched/debug.c:print_task().
*/
struct task_info {
@@ -119,6 +124,11 @@
unsigned long config_dl_runtime = 20000;
/*
+ * FIFO priority defaults
+ */
+unsigned long config_fifo_priority = 98;
+
+/*
* control loop (time in seconds)
*/
long config_starving_threshold = 60;
@@ -692,19 +702,81 @@
return 0;
}
+void setupattr(struct sched_attr *attr)
+{
+ memset(attr, 0, sizeof(*attr));
+ attr->size = sizeof(*attr);
+ if (use_deadline) {
+ attr->sched_policy = SCHED_DEADLINE;
+ attr->sched_runtime = config_dl_runtime;
+ attr->sched_deadline = config_dl_period;
+ attr->sched_period = config_dl_period;
+ }
+ else {
+ attr->sched_policy = SCHED_FIFO;
+ attr->sched_priority = config_fifo_priority;
+ }
+}
+
+const char *get_boost_policy(void)
+{
+ if (use_deadline)
+ return "SCHED_DEADLINE";
+ return "SCHED_FIFO";
+}
+
+void normalize_timespec(struct timespec *ts)
+{
+ while (ts->tv_nsec >= NS_PER_SEC) {
+ ts->tv_nsec -= NS_PER_SEC;
+ ts->tv_sec++;
+ }
+}
+
+void get_current_time(struct timespec *now)
+{
+ int ret = clock_gettime(CLOCK_MONOTONIC, now);
+ if (ret == -1)
+ log_msg("wait_for_period: failed to get current time!");
+}
+
+void wait_for_period(struct timespec *start)
+{
+ int ret;
+ struct timespec period_end = *start;
+
+ /*
+ * if we're using SCHED_DEADLINE then just
+ * sleep for boost duration and return
+ */
+ if (use_deadline) {
+ sleep(config_boost_duration);
+ return;
+ }
+
+ /*
+ * if we're using SCHED_FIFO to boost then take care
+ * to sleep to the period end (emulate SCHED_DEADLINE)
+ */
+ period_end.tv_nsec += config_dl_period;
+ normalize_timespec(&period_end);
+ ret = clock_nanosleep(CLOCK_MONOTONIC, TIMER_ABSTIME,
+ &period_end, NULL);
+ if (ret == -1)
+ log_msg("wait_for_period: clock_nanosleep failed!: %s",
+ strerror(errno));
+}
+
int boost_starving_task(int pid)
{
int ret;
int flags = 0;
struct sched_attr new_attr;
struct sched_attr old_attr;
+ struct timespec start;
- memset(&new_attr, 0, sizeof(new_attr));
- new_attr.size = sizeof(new_attr);
- new_attr.sched_policy = SCHED_DEADLINE;
- new_attr.sched_runtime = config_dl_runtime;
- new_attr.sched_deadline = config_dl_period;
- new_attr.sched_period = config_dl_period;
+ setupattr(&new_attr);
+ get_current_time(&start);
/*
* Get the old prio, to be restored at the end of the
@@ -717,7 +789,8 @@
*/
ret = sched_setattr(pid, &new_attr, flags);
if (ret < 0) {
- log_msg("sched_setattr failed to boost pid %d: %s\n", pid, strerror(errno));
+ log_msg("sched_setattr failed to boost pid %d using %s: %s\n",
+ pid, get_boost_policy(), strerror(errno));
return 1;
}
@@ -727,7 +800,7 @@
* XXX: We might want to check if the task suspended before the
* end of the duration.
*/
- sleep(config_boost_duration);
+ wait_for_period(&start);
/*
* Restore the old priority.
@@ -1155,12 +1228,26 @@
}
}
+int check_deadline(void)
+{
+ char buffer[BUFFER_SIZE];
+
+ if (read_sched_debug(buffer, BUFFER_SIZE) == 0)
+ die("check_deadline: error reading sched_debug info");
+
+ if (get_variable_long_value(buffer, ".dl_nr_running") != -1)
+ return 1;
+ log_msg("deadline scheduler not available, falling back to fifo");
+ return 0;
+}
int main(int argc, char **argv)
{
struct cpu_info *cpus;
int nr_cpus;
+ use_deadline = check_deadline();
+
parse_args(argc, argv);
nr_cpus = sysconf(_SC_NPROCESSORS_CONF);
