Re: [PATCH 2/2] selftests/sched_ext: Add lazy preemption tests

[Tejun Heo]

Hello, Andrea.

This is an AI review. The tests pass here under vng in lazy, full and none
modes. The timing below was measured with the target CPU on nohz_full.

On Tue, Sep 15, 2026 at 09:45:13PM +0200, Andrea Righi wrote:
> +void BPF_STRUCT_OPS(kick_stopping, struct task_struct *p, bool runnable)
> +{
> +	if (p->pid == victim_pid && state == KICK_STATE_RESCHED)
> +		state = KICK_STATE_DONE;
> +}

In the kick and tick scenarios the victim is the only runnable task on the
target CPU, so when its slice is cleared the kernel keeps running it with a
refilled slice and ops.stopping() is never called. DONE only arrives when some
other task pinned to that CPU wakes up. With the target CPU on nohz_full here,
the first such wakeup was a kworker 2.86s after the victim started, against
the 3s wait. Without nohz_full the scenarios still spend about half a second
each waiting for one. SCX_OPS_ENQ_LAST with the SCX_ENQ_LAST kick as in
nohz_tick.bpf.c would put the victim through ops.stopping() on its own.

> +SEC("fexit/__resched_curr")
> +int BPF_PROG(kick_need_resched, struct rq *rq, int tif)
> +{
> +	struct task_struct *task = BPF_CORE_READ(rq, curr);
> +
> +	if (!task || BPF_CORE_READ(task, pid) != victim_pid ||
> +	    BPF_CORE_READ(rq, cpu) != target_cpu || state != KICK_STATE_QUEUED)
> +		return 0;

This latches the first resched of the victim whatever the cause. A higher
class wakeup on the target CPU inside the window, which is the whole 20ms
slice in TICK_EXPIRY, records a full slice and fails observation_valid().
Skipping observations with a non-zero slice would be more robust.

> +	SCX_EQ(lazy_wait.resched_tif, immediate.resched_tif);
> +	SCX_GT(lazy_wait.nr_wait_callbacks, 0);

DONE is set from ops.stopping() before the context switch and the balance
callback runs after it, so this read can race the increment. Poll for it.

> +	file = fopen("/sys/kernel/debug/sched/preempt", "r");
> +	if (!file) {
> +#if defined(CONFIG_PREEMPT_LAZY) && !defined(CONFIG_PREEMPT_DYNAMIC)
> +		return 1;
> +#else
> +		return -1;
> +#endif

This reflects the build tree, not the running kernel. Also, none and voluntary
return -1 and skip the comparison, but the TIFs must match in those modes too.

> +	skel->rodata->scenario = scenario;
> +	if (kick__load(skel))
> +		goto out;

If __resched_curr() isn't in BTF, this fails too and kick_invalid FAILs along
with the other four although it doesn't use the tracer. Skipping when the
target is missing would be better.

> +	if (!start_gated_worker(&victim) ||
> +	    !wait_for_counter(&skel->bss->nr_lazy_victim_running, 1,
> +			      PHASE_TIMEOUT_MS)) {
> +		SCX_ERR("Lazy-enqueue victim was not scheduled");
> +		goto out;
> +	}
> +	usleep(100000);

Neither lazy phase checks that the tick actually stopped before the request,
so they pass with the tick running the whole time.

> Extend the NO_HZ_FULL test with infinite-slice victims. Verify that lazy
> enqueue and kick requests restart a stopped tick and make forward
> progress. Keep invalid kick flag coverage and skip modes not exposed by
> the running kernel.

kick.c is new, nothing is kept. The controller mask and timeout changes to
nohz_tick aren't mentioned either.

Thanks.

--
tejun