On 3/17/26 08:23, Christian Loehle wrote:
> On 3/16/26 22:26, Christian Loehle wrote:
>> On 3/16/26 17:46, Tejun Heo wrote:
>>> Hello,
>>>
>>> On Mon, Mar 16, 2026 at 10:02:48AM +0000, Christian Loehle wrote:
>>>> @@ -5686,11 +5718,20 @@ static void kick_cpus_irq_workfn(struct irq_work *irq_work)
>>>>  		 * task is picked subsequently. The latter is necessary to break
>>>>  		 * the wait when $cpu is taken by a higher sched class.
>>>>  		 */
>>>> -		if (cpu != cpu_of(this_rq))
>>>> +		if (cpu != this_cpu)
>>>>  			smp_cond_load_acquire(wait_kick_sync, VAL != ksyncs[cpu]);
>>>
>>> Given that irq_work is executed at the end of IRQ handling, we can just
>>> reschedule the irq work when the condition is not met (or separate that out
>>> into its own irq_work). That way, I think we can avoid the global lock.
>>>
>> I'll go poke at it some more, but I think it's not guaranteed that B actually
>> advances kick_sync if A keeps kicking. At least not if the handling is in HARD irqwork?
>> Or what would the separated out irq work do differently?
> 
> So in my particular example I do the SCX_KICK_WAIT in ops.enqueue(), which is fair, but
> I don't think we can delay calling that until we've advanced our local kick_sync and if we
> don't we end up in the deadlock, even if e.g. we separate out the retry (and make that lazy),
> because then the local CPU is able to continuously issue new kicks (which will have to
> be handled by the non-retry path) without advancing it's own kick_sync.
> The closest thing to that I can get working is separating out the SCX_KICK_WAIT entirely and
> make that lazy. In practice though that would realistically make the SCX_KICK_WAIT latency
> most likely a lot higher than with the global lock, is that what you had in mind?
> Or am I missing something here?

Something like the attached, for completeness