Re: [lock validator] inet6_destroy_sock(): soft-safe -> soft-unsafe lock dependency

[Ingo Molnar <mingo@elte.hu>]

* David S. Miller <davem@davemloft.net> wrote:

> From: Herbert Xu <herbert@gondor.apana.org.au>
> Date: Tue, 31 Jan 2006 21:27:58 +1100
> 
> > tcp_close is only called from process context.  The rule for sk_dst_lock
> > is that it must also only be obtained in process context.  On the other
> > hand, it is true that sk_lock can be obtained in softirq context.
> > 
> > In this particular case, sk_dst_lock is obtained by tcp_close with
> > softirqs disabled.  This is not a problem in itself since we're not
> > trying to get sk_dst_lock from a real softirq context (as opposed to
> > process context with softirq disabled).
> > 
> > I believe this warning comes about because the validator creates a
> > dependency between sk_lock and sk_dst_lock.  It then infers from this
> > dependency that in softirq contexts where sk_lock is obtained the code
> > may also attempt to obtain sk_dst_lock.
> > 
> > This inference is where the validator errs.  sk_dst_lock is never
> > (or should never be, and as far as I can see none of the traces show
> > it to do so) obtained in a real softirq context.
> 
> Herbert's analysis is correct.  This unique locking strategy is used 
> by tcp_close() because at this point it knows that every single 
> reference to this socket in the system is gone once it takes the 
> socket lock with BH disabled.
> 
> And that known invariant is why this is correct, and the locking 
> validator has no way to figure this out.

ok, thanks for the analysis! I'll fix this with an explicit hint to the 
validator.

	Ingo