Re: [PATCH bpf 0/4] bpf: remove __rcu annotations from bpf_prog_array

[Stanislav Fomichev <sdf@fomichev.me>]

On 05/14, Eric Dumazet wrote:
> 
> 
> On 5/14/19 7:27 PM, Alexei Starovoitov wrote:
> 
> > what about activate_effective_progs() ?
> > I wouldn't want to lose the annotation there.
> > but then array_free will lose it?
It would not have have it because the input is the result of
bpf_prog_array_alloc() which returns kmalloc'd pointer (and
is not bound to an rcu section).

> > in some cases it's called without mutex in a destruction path.
Hm, can you point me to this place? I think I checked every path,
maybe I missed something subtle. I'll double check.

> > also how do you propose to solve different 'mtx' in
> > lockdep_is_held(&mtx)); ?
> > passing it through the call chain is imo not clean.
Every caller would know which mutex protects it. As Eric said below,
I'm adding a bunch of xxx_dereference macros that hardcode mutex, like
the existing rtnl_dereference.

> Usage of RCU api in BPF is indeed a bit strange and lacks lockdep support.
> 
> Looking at bpf_prog_array_copy_core() for example, it looks like the __rcu
> in the first argument is not needed, since the caller must have done the proper dereference already,
> and the caller knows which mutex is protecting its rcu_dereference_protected() for the writer sides.
> 
> bpf_prog_array_copy_core() should manipulate standard pointers, with no __rcu stuff.
> 
> The analogy in net/ are probably the rtnl_dereference() users.