Re: 2.6.17-rc5-mm1

[Ingo Molnar <mingo@elte.hu>]

* Andrew Morton <akpm@osdl.org> wrote:

> Without having looked at it very hard, I'd venture that this is a 
> false positive - that driver uses disable_irq() to prevent reentry 
> onto that lock.

correct.

> It does that because it knows it's about to spend a long time talking 
> with the mii registers and it doesn't want to do that with interrupts 
> disabled.

i still consider it a 'quirky' locking construct, because disabling 
interrupts for a long time also disables all other devices sharing the 
same IRQ line - not nice.

Also, this is a really hard case for lockdep to detect automatically. 
(fortunately it's also relatively rare)

OTOH, the straightforward lockdep workaround would be to take the 
spinlock and thus disable all local interrupts - not too nice either. 

Albeit in some ways it's still a bit nicer conceptually than disabling 
the irq line, because other CPUs are still operational, and under 
certain locking designs [preempt-rt] spin_lock_irq() does not disable 
local interrupts.
 
Steve, can you think of any better solution? I dont have this card.

	Ingo