Re: [PATCH net-next] net: add annotation for sock_{lock,unlock}_fast

[Paolo Abeni <pabeni@redhat.com>]

Hello,

On Tue, 2020-11-17 at 17:58 +0100, Luc Van Oostenryck wrote:
> On Tue, Nov 17, 2020 at 09:38:45AM +0100, Paolo Abeni wrote:
> > Hello,
> > 
> > Thank you for the feedback!
> > 
> > On Mon, 2020-11-16 at 23:27 +0100, Luc Van Oostenryck wrote:
> > > > @@ -1606,10 +1607,12 @@ bool lock_sock_fast(struct sock *sk);
> > > >   */
> > > >  static inline void unlock_sock_fast(struct sock *sk, bool slow)
> > > >  {
> > > > -	if (slow)
> > > > +	if (slow) {
> > > >  		release_sock(sk);
> > > > -	else
> > > > +		__release(&sk->sk_lock.slock);
> > > 
> > > The correct solution would be to annotate the declaration of
> > > release_sock() with '__releases(&sk->sk_lock.slock)'.
> > 
> > If I add such annotation to release_sock(), I'll get several sparse
> > warnings for context imbalance (on each lock_sock()/release_sock()
> > pair), unless I also add an '__acquires()' annotation to lock_sock(). 
> > 
> > The above does not look correct to me ?!? When release_sock() completes
> > the socket spin lock is not held.
> 
> Yes, that's fine, but I suppose it somehow releases the mutex that
> is taken in lock_sock_fast() when returning true, right?

Well, it has mutex semantics, but does not really acquire any mutex.

> > The annotation added above is
> > somewhat an artifact to let unlock_sock_fast() matches lock_sock_fast()
> > from sparse perspective. I intentionally avoided changing
> > the release_sock() annotation to avoid introducing more artifacts.
> > 
> > The proposed schema is not 100% accurate, as it will also allow e.g. a
> > really-not-fitting bh_lock_sock()/unlock_sock_fast() pair, but I could
> > not come-up with anything better.
> > 
> > Can we go with the schema I proposed?
> 
> Well, I suppose it's a first step.
> But can you then add a '__releases(...)' to unlock_sock_fast()?
> It's not needed by sparse because it's an inline function and sparse
> can then deduce it but it will help to see the pairing with
> lock_sock_fast() is OK.

Ok, I'll send a v2 with such annotation.

Thanks!

Paolo