> So how about something like this (untested! not necessarily very well 
quoted text
> thought through either!)
> 
> Basic notion: the only thing that serializes the IRQ_INPROGRESS flag is 
> the descriptor lock. And we don't have to (or even want to!) hold it while 
> waiting for the thing, but we want to *have*held*it* in between whatever 
> we're synchronizing with.
> 
> The internal irq handler functions already held the lock when they did 
> whatever they need to serialize - and they are possibly performance 
> critical too - so they use the "internal" function that doesn't get the 
> lock unnecessarily again.

That may do the trick as the read can't cross the spin_lock (it can
cross spin_unlock but not lock). Advantage over adding a barrier to
handle_IRQ_event() is that it keeps the overhead to the slow path
(synchronize_irq).

Note that I didn't actually experience a problem here. I just came upon
that by accident while thinking about a similar issue I have with
napi_synchronize().

Looks good to me on a first glance (unfortunately a bit ugly but heh).

Ben.


-
unsubscribe notice
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/