Re: [RFC] disable_irq()/enable_irq() semantics and ide-probe.c

[viro]

On Wed, Oct 08, 2003 at 07:29:10PM -0700, Linus Torvalds wrote:
> >				  If an interrupt comes during that
> > time, we'll get IRQ_INPROGRESS set and not reset until later register_irq()
> > (see handle_irq() for details).  Note that calling disable_irq() after that
> > will kill us on SMP - it will spin waiting for IRQ_INPROGRESS to go away.
> 
> Now _this_ is a bug waiting to happen. I don't think it actually happens 
> now (since anybody who does disable_irq() _will_ either have registered 
> the irq already or will do so soon, but I agree that it's just trouble 
> waiting to happen.

Ummm...  probe_hwif() is a good example of the opposite - it can fail
past the point where it disables irq and that means no register_irq()
after enable_irq() call on cleanup.

> I think the fix to that is to just add a trivial test for "if the handler
> list is empty, don't bother synchronizing" in disable_irq(), since clearly
> if the list is empty there is nothing to synchronize _with_. After all,
> the synchronization is there just to make sure no handler runs
> concurrently on another CPU.

How about
 
        action = NULL;
        if (!(status & (IRQ_DISABLED | IRQ_INPROGRESS))) {
                action = desc->action;
                status &= ~IRQ_PENDING; /* we commit to handling */
		if (likely(action))
			status |= IRQ_INPROGRESS; /* we are handling it */
        }
        desc->status = status;

in handle_irq()?

> As far as I can tell, 2.6.x is doing all the right things. Modulo the (not
> really supported) concurrent device probing, and the (not implemented) 
> atomic irq requesting.
> 
> Note that the IRQ_INPROGRESS thing was literally the bit that autodetect
> used to test, it got changed it to IRQ_WAITING to clarify the code and
> avoid bad interactions with the other uses of IRQ_INPROGRESS.
> 
> And note that we do _not_ clear IRQ_INPROGRESS on "action == NULL" very
> much on purpose: that "action == NULL" thing also happens if the IRQ is
> disabled, and we need to get the edge replay right. This is why
> request_irq() literally _needs_ to clear that bit in 2.6.x.

	See above - we shouldn't clear it on action == NULL, but we don't
need to set it, AFAICS.
 
> So the fix is to make 2.4.x do what 2.6.x does, methinks.

ObOtherFun:  There's another bogosity in quoted ide-probe.c code, according
to dwmw2 - he says that there are PCI IDE cards that get IRQ 0, so the
test for hwif->irq is b0rken.  We probably should stop overloading
->irq == 0 for "none given", but I'm not sure that we *have* a value
that would never be used as an IRQ number on all platforms...
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/