Re: 2.6.19-rc6: known regressions (v4)

[Pavel Emelianov <xemul@openvz.org>]

> I really think this is wrong.
> 
> The original patch was wrong, and the _real_ problem is in __do_IRQ() that 
> got the desc->lock too early.
> 
> I _think_ the correct fix is to simply revert the broken commit, and fix 
> the _one_ place that called "misnote_interrupt()" with the lock held.
> 
> Something like this..
> 
> I also think that the real fix will be to move the whole
> 
> 	if (!noirqdebug)
> 		note_interrupt(irq, desc, action_ret);
> 
> 
> into handle_IRQ_event itself, since every caller (except for 
> "misrouted_irq()" itself, and that should probably be done separately) 
> should always do it. Right now we have a lot of people that just do
> 
> 	action_ret = handle_IRQ_event(irq, action);
> 	if (!noirqdebug)
> 		note_interrupt(irq, desc, action_ret);
> 
> explicitly.
> 
> The only thing that keeps us from doing that is that we don't pass in 
> "desc", but we should just do that.
> 
> But in the meantime, this appears to be the minimal fix. Can people please 
> test and verify?

This works for me, but is this normal that desc's fields are
modified non-atomically in note_interrupt()?

And one more thing - report_bad_irq() traverses desc->action
list without any locking either.