Re: [PATCH 0/3] lockdep: Allow checking a read-only lock

From: Peter Zijlstra
Date: Wed Jan 17 2018 - 11:33:51 EST


On Wed, Jan 17, 2018 at 07:14:11AM -0800, Matthew Wilcox wrote:
> From: Matthew Wilcox <mawilcox@xxxxxxxxxxxxx>
>
> I am not for one moment suggesting that the concept of a read-only lock
> makes sense. You can't sensibly put one in ROM or in read-only mappings.
> What does make sense is some APIs want to specify a const pointer to
> indicate that they do not modify the object being pointed to. One example
> we have of this today is in the networking stack; tcp_md5_do_lookup takes
> a const struct sock * argument and wants to ensure that the caller either
> took the socket lock or the rcu lock.
>
> At the moment, tcp_md5_do_lookup() is actually lying to its callers;
> lockdep_sock_is_held() casts away the constness of the pointer because
> lockdep actually does modify the lock when checking whether it's held
> (under rare and unnecessary conditions).
>
> Fix this situation by (patch 1) only assigning a lock key on registration,
> not on check, (patch 2) marking the pointers in the lockdep check path
> as const and (patch 3) converting a few of the callers to themselves
> be const, removing the nasty hack in lockdep_sock_is_held().
>

Seems OK.

Acked-by: Peter Zijlstra (Intel) <peterz@xxxxxxxxxxxxx>

Ingo can you make that happen?