[RFC] sched/rt: Fix RT (group) throttling with nohz_full

If nohz_full is enabled (more precisely HK_FLAG_TIMER is set), then
do_sched_rt_period_timer may be called on a housekeeping CPU,
which would not service the isolated CPU for a non-root cgroup
(requires a kernel with RT_GROUP_SCHEDULING).
This causes RT tasks in a non-root cgroup to get throttled 
indefinitely (unless throttling is disabled) once the timer has 
been moved to a housekeeping CPU.
To fix this, housekeeping CPUs now service all online CPUs 
if HK_FLAG_TIMER (nohz_full) is set.

I'm not really sure how this relates to  Mike Galbraith previous
commit e221d028bb08 ("sched,rt: fix isolated CPUs leaving root_task_group
indefinitely throttled"), (which is dated before the housekeeping changes,)
so I'm posting this as an RFC.


Signed-off-by: Jonathan Schwender <redacted>
---
 kernel/sched/rt.c | 9 ++++++++-
 1 file changed, 8 insertions(+), 1 deletion(-)

diff --git a/kernel/sched/rt.c b/kernel/sched/rt.c
index 49ec096a8aa1..3185e00b828a 100644
--- a/kernel/sched/rt.c
+++ b/kernel/sched/rt.c

@@ -865,9 +865,16 @@ static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun)
 	 * isolation is really required, the user will turn the throttle
 	 * off to kill the perturbations it causes anyway.  Meanwhile,
 	 * this maintains functionality for boot and/or troubleshooting.
+	 * If nohz_full is active and the timer was offloaded to a
+	 * housekeeping CPU, sched_rt_period_mask() will not contain
+	 * the isolated CPU. To prevent indefinite throttling of tasks
+	 * on isolated CPUs, housekeeping CPUs service all online CPUs.
 	 */
-	if (rt_b == &root_task_group.rt_bandwidth)
+	if (rt_b == &root_task_group.rt_bandwidth
+		|| (housekeeping_enabled(HK_FLAG_TIMER)
+			&& housekeeping_cpu(this_rq()->cpu, HK_FLAG_TIMER))) {
 		span = cpu_online_mask;
+	}
 #endif
 	for_each_cpu(i, span) {
 		int enqueue = 0;

-- 
2.29.2

RFC v2 changes: Fix compile error if CONFIG_SMP is not set, which was
Reported-by: kernel test robot <redacted>
I'm now using smp_processor_id() to get the ID of the CPU the timer 
is running on.


If nohz_full is enabled (more precisely HK_FLAG_TIMER is set), then
do_sched_rt_period_timer may be called on a housekeeping CPU,
which would not service the isolated CPU for a non-root cgroup
(requires a kernel with RT_GROUP_SCHEDULING).
This causes RT tasks in a non-root cgroup to get throttled
indefinitely (unless throttling is disabled) once the timer has
been moved to a housekeeping CPU.
To fix this, housekeeping CPUs now service all online CPUs
if HK_FLAG_TIMER (nohz_full) is set.

I'm not really sure how this relates to  Mike Galbraith previous
commit e221d028bb08 ("sched,rt: fix isolated CPUs leaving root_task_group
indefinitely throttled"), (which is dated before the housekeeping changes,)
so I'm posting this as an RFC.

Signed-off-by: Jonathan Schwender <redacted>
---
 kernel/sched/rt.c | 9 ++++++++-
 1 file changed, 8 insertions(+), 1 deletion(-)

diff --git a/kernel/sched/rt.c b/kernel/sched/rt.c
index 8f720b71d13d..879abb5de023 100644
--- a/kernel/sched/rt.c
+++ b/kernel/sched/rt.c

@@ -866,9 +866,16 @@ static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun)
 	 * isolation is really required, the user will turn the throttle
 	 * off to kill the perturbations it causes anyway.  Meanwhile,
 	 * this maintains functionality for boot and/or troubleshooting.
+	 * If nohz_full is active and the timer was offloaded to a
+	 * housekeeping CPU, sched_rt_period_mask() will not contain
+	 * the isolated CPU. To prevent indefinite throttling of tasks
+	 * on isolated CPUs, housekeeping CPUs service all online CPUs.
 	 */
-	if (rt_b == &root_task_group.rt_bandwidth)
+	if (rt_b == &root_task_group.rt_bandwidth
+		|| (housekeeping_enabled(HK_FLAG_TIMER)
+			&& housekeeping_cpu(smp_processor_id(), HK_FLAG_TIMER))) {
 		span = cpu_online_mask;
+	}
 #endif
 	for_each_cpu(i, span) {
 		int enqueue = 0;

-- 
2.29.2

On 2021-02-02 10:00:10 [+0100], Jonathan Schwender wrote:

If nohz_full is enabled (more precisely HK_FLAG_TIMER is set), then
do_sched_rt_period_timer may be called on a housekeeping CPU,
which would not service the isolated CPU for a non-root cgroup
(requires a kernel with RT_GROUP_SCHEDULING).
This causes RT tasks in a non-root cgroup to get throttled
indefinitely (unless throttling is disabled) once the timer has
been moved to a housekeeping CPU.
To fix this, housekeeping CPUs now service all online CPUs
if HK_FLAG_TIMER (nohz_full) is set.

This originates from
   https://lore.kernel.org/linux-rt-users/b07b6fc7-1a5f-0dcc-ca65-821de96cd8b4@gmail.com/

Could someone please take a look?

Sebastian