Re: lockdep and threaded IRQs (was: ...)

[David Brownell <david-b@pacbell.net>]

On Saturday 28 February 2009, Thomas Gleixner wrote:
> On Sat, 28 Feb 2009, David Brownell wrote:
> > Got a version that applies to mainline GIT?
> 
> http://tglx.de/~tglx/patches.tar.bz2 

Got it, thanks.


> > At a quick glance it looks like these patches don't cover
> > set_irq_chained_handler(), which would be trouble since
> > __setup_irq() doesn't get called in those cases.
> 
> Hmm, I did not think about chained handlers where the demux handler
> needs to run in a thread as well. Usually demux handlers just have a
> fast path kicking the particular real handlers.

That can't work when the demux needs to access state across
I2C in order to see which "real" handlers to kick.  :)


> > They should however handle simpler cases, like I2C devices
> > that only expose one IRQ instead of needing to demux several
> > dozen IRQs going to different drivers and subsystems.
> > 
> > And, not touching lockdep, the original ugliness will still
> > be needed (re-enabling IRQs in threaded handlers).
> 
> Err ? The threaded handlers run with interrupts enabled.

Hmm, I'll have a closer look.  You changed handle_IRQ_event()
which is where the relevant IRQF_DISABLED test kicks in.  In
your updated code, that pokes any quick_check_handler() and
then maybe pokes a per-IRQ thread.

That seems to presume a hardirq-to-taskirq handoff.  But the
problem case is taskirq-to-taskirq chaining, through e.g.
what set_irq_chip_and_handler() provided.  (Details not very
amenable to brief emails, just UTSL.)

Thing is, I'm not sure a per-IRQ thread can work easily with
that chaining.  The chained IRQs can need to be handled before
the top-level IRQ gets re-enabled.  That's why the twl4030-irq
code uses just one taskirq thread for all incoming events.

(Which of course is rarely more than one at a time, so there's
little reason not to share that task between the demuxing code
and the events being demuxed.  Interrupts that need processing
via I2C/SPI/etc are more or less by definition not frequent or
performance-critical.)

- Dave