On Tue, May 14, 2019 at 7:11 PM Stanislav Fomichev <sdf@xxxxxxxxxxx> wrote:
>
> On 05/14, Alexei Starovoitov wrote:
> > On Tue, May 14, 2019 at 10:53 AM Stanislav Fomichev <sdf@xxxxxxxxxxx> wrote:
> > >
> > > Existing __rcu annotations don't add anything to the safety.
> >
> > what do you mean?
> > BPF_PROG_RUN_ARRAY derefs these pointers under rcu.
> And I'm not removing them from the struct definitions, I'm removing __rcu
> from the helpers' arguments only. Because those helpers are always called
> with the mutex and don't need it. To reiterate: rcu_dereference_protected
> is enough to get a pointer (from __rcu annotated) for the duration
> of the mutex, helpers can operate on the non-annotated (dereferenced) prog
> array.
>
> Read section still does the following (BPF_PROG_RUN_ARRAY):
>
> rcu_read_lock();
> p = rcu_dereference(__rcu'd progs);
> while (p) {}
> rcu_read_unlock();
>
> And write sections do:
>
> mutex_lock(&mtx);
> p = rcu_dereference_protected(__rcu'd progs, lockdep_is_held(&mtx);
> // ^^^ does rcu_dereference in the mutex protected section
> bpf_prog_array_length(p);
> bpf_prog_array_copy_to_user(p, ...);
> bpf_prog_array_delete_safe(p);
> bpf_prog_array_copy_info(p);
> bpf_prog_array_copy(p, ...);
> bpf_prog_array_free(p);
what about activate_effective_progs() ?
I wouldn't want to lose the annotation there.
but then array_free will lose it?
in some cases it's called without mutex in a destruction path.
also how do you propose to solve different 'mtx' in
lockdep_is_held(&mtx)); ?
passing it through the call chain is imo not clean.
I wonder what others think about this whole discussion.
> // ^^^ all these helpers are consistent already with or
> // without __rcu annotation because we hold a mutex and
> // guarantee no concurrent updates, so __rcu annotations
> // for their input arguments is not needed.
> mutex_unlock(&mtx);
