Re: [Xenomai-help] kernel oopses when killing realtime task

[Philippe Gerum <rpm@xenomai.org>]

On Tue, 2010-10-26 at 21:33 +0200, Jan Kiszka wrote:
> Am 26.10.2010 07:22, Jan Kiszka wrote:
> > Will come up with two patches for stable, one for I-pipe and one for
> > Xenomai, later today. Then we can discuss which cases I'm missing.
> 
> While meditating over my approach (which turned out to be less trivial
> as expected - of course), I also reconsidered your current patches. The
> concerns I had (forwarding of spurious IRQ to Linux) turned out to be
> harmless (Linux will ignore such few spurious events).
> 

That is not even an issue if you consider the sequence to be
xnarch_disable_irq then ipipe_control (new version, doing a critical
entry to flip the irq mode).

> Still, the approach to sync via shutting down the line for the current
> domain before xnintr_irq_detach doesn't work for us. It only works if
> xnintr_irq_detach actually detaches from the line, but it breaks if
> there are users remaining.
> 
> We need intrlock to check if we are the last user while removing
> ourselves from the list. And we cannot postpone line detaching after the
> critical section as we may otherwise race with the next registration on
> that line. IOW, I don't see how to solve the issue without moving the
> drain after the detach and making the detach safer instead.
> 
> Do you agree?
> 

I agree this is not trivial, for sure. To keep things simple, I would
introduce a new "teardown" flag to freeze the descriptor, thus avoiding
further attachments, while xnintr_detach can probe the shared list for
lingering users, and eventually call xnarch_disable_irq
+xnarch_ignore_irq+xnarch_release_irq in sequence with all locks
dropped, if empty.

The only adverse effect I can see ATM would be some concurrent caller of
xnintr_detach() blocked on the teardown flag on another CPU, albeit it
_could_ have joined the bandwagon, attaching the irq, in case the shared
list proved to remain active (and thus xnarch_release_irq was not
called). But this may also look like a simple way to prevent live
locking of interrupt descriptors. YMMV.

> Jan
> 

-- 
Philippe.