Re: [PATCH] cobalt/pipe: Fix invalid wait context in xnpipe_wakeup_proc()

[Philippe Gerum <philippe.gerum@gmail.com>]

Jan Kiszka <jan.kiszka@siemens.com> writes:

> On 29.08.25 17:59, Florian Bezdeka wrote:
>> Recent Dovetail versions complained about an invalid wait context in
>> xnpipe_wakeup_proc(). It turned out that this warning is correct as this
>> function is called from the inband IRQ context (while syncing the IRQ
>> log) with hard IRQs enabled and inband stage stalled.
>> 
>> The call graph looks like that:
>> xnpipe_wakeup_proc()
>>     wake_up_interruptible()
>>         __wake_up_common_lock()
>> 	    spin_lock_irqsave() <-- might sleep on PREEMPT_RT
>> 
>
> So, the warning is issued independent of PREEMPT_RT being enabled but
> the bug would be limited to that configuration, right?
>
> Since when is the kernel complaining? It is wrong since PREEMPT_RT is
> available, officially 6.12 then. Then this should go into 3.3-stable as
> well - noted.
>
>> As confirmed by Philippe in the discussion linked below, we have to
>> migrate the wakeup code from using a synthetic IRQ to irq_work to address
>> that issue.
>> 
>
> Hmm, can't we use something that is a thread when needed and an IRQ
> otherwise?
>

Unless PREEMPT_RT is enabled, we can't guarantee threaded handling, but
we can always guarantee irq-based handling by setting IRQ_WORK_HARD_IRQ
for the irq_work struct.

-- 
Philippe.