Re: [PATCH 2/2] selftests/sched_ext: Add class_switch_stall test

[Andrea Righi <arighi@nvidia.com>]

Hi Michal,

On Mon, Oct 19, 2026 at 12:41:46PM +0000, Michal Blaszczyk wrote:
> Add a test for a task which leaves SCHED_EXT while it's waiting. Its
> scheduler keeps the task waiting on a DSQ after it wakes up. The test
> switches the task to SCHED_OTHER while it's waiting, lets it sleep outside
> of SCHED_EXT for three times the watchdog timeout, then switches it back
> and wakes it up. The scheduler keeps it waiting again, for longer than the
> watchdog interval but shorter than the timeout, which the watchdog must not
> report as a stall.
> 
> Without the previous patch, the watchdog charges the task for the time it
> spent outside of SCHED_EXT and aborts the scheduler.
> 
> Assisted-by: LLM
> Signed-off-by: Michal Blaszczyk <michalblk@google.com>

...

> diff --git a/tools/testing/selftests/sched_ext/class_switch_stall.c b/tools/testing/selftests/sched_ext/class_switch_stall.c
> new file mode 100644
> index 0000000000000..489d23c777f05
> --- /dev/null
> +++ b/tools/testing/selftests/sched_ext/class_switch_stall.c
> @@ -0,0 +1,240 @@
> +// SPDX-License-Identifier: GPL-2.0
> +/*
> + * Check that a task which leaves SCHED_EXT while it's waiting doesn't carry
> + * that wait over to its next stay on SCHED_EXT. If it did, the watchdog would
> + * count the time the task spent outside SCHED_EXT as waiting and abort the
> + * scheduler.
> + *
> + * Copyright (c) 2026 Google LLC.
> + */
> +#define _GNU_SOURCE
> +
> +#include <bpf/bpf.h>
> +#include <poll.h>
> +#include <sched.h>
> +#include <scx/common.h>
> +#include <signal.h>
> +#include <stdio.h>
> +#include <string.h>
> +#include <sys/wait.h>
> +#include <unistd.h>
> +#include "class_switch_stall.bpf.skel.h"
> +#include "scx_test.h"
> +
> +#ifndef SCHED_EXT
> +#define SCHED_EXT 7
> +#endif
> +
> +/* class_switch_stall_ops.timeout_ms, the watchdog checks every half of it */
> +#define TIMEOUT_MS 1000
> +/* longer than the watchdog interval, but shorter than the timeout */
> +#define PARK_MS 750
> +/* time spent outside of SCHED_EXT, a stale wait would exceed the timeout */
> +#define OUTSIDE_MS (3 * TIMEOUT_MS)

Could we increase the watchdog timeout to 3s and the park duration to
2s, updating timeout_ms accordingly?

With the current values, there is only about 250ms of headroom for timer
delivery, dispatch, and the child actually getting CPU time before it
becomes eligible for a watchdog stall The first phase also depends on
userspace switching the parked child out within 1s.

I haven't observed a false failure, so this is a robustness suggestion
(I'm thinking at loaded CI VMs) rather than an observed problem. A 2s
park would still span the nominal 1.5s watchdog interval while leaving
more headroom.

Thanks,
-Andrea