On Fri, 2018-07-13 at 16:41 +0200, Daniel Borkmann wrote:
> On 07/13/2018 04:26 PM, Paolo Abeni wrote:
> > On Fri, 2018-07-13 at 16:08 +0200, Daniel Borkmann wrote:
> > > On 07/13/2018 11:55 AM, Paolo Abeni wrote:
> > > > Each lockless action currently does its own RCU locking in ->act().
> > > > This is allows using plain RCU accessor, even if the context
> > > > is really RCU BH.
> > > >
> > > > This change drops the per action RCU lock, replace the accessors
> > > > with _bh variant, cleans up a bit the surronding code and documents
> > > > the RCU status in the relevant header.
> > > > No functional nor performance change is intended.
> > > >
> > > > The goal of this patch is clarifying that the RCU critical section
> > > > used by the tc actions extends up to the classifier's caller.
> > > >
> > > > Signed-off-by: Paolo Abeni <pab...@redhat.com>
> > >
> > > [...]
> > > > diff --git a/net/sched/act_bpf.c b/net/sched/act_bpf.c
> > > > index 06f743d8ed41..ac20266460c0 100644
> > > > --- a/net/sched/act_bpf.c
> > > > +++ b/net/sched/act_bpf.c
> > > > @@ -45,8 +45,7 @@ static int tcf_bpf(struct sk_buff *skb, const struct
> > > > tc_action *act,
> > > > tcf_lastuse_update(&prog->tcf_tm);
> > > > bstats_cpu_update(this_cpu_ptr(prog->common.cpu_bstats), skb);
> > > >
> > > > - rcu_read_lock();
> > > > - filter = rcu_dereference(prog->filter);
> > > > + filter = rcu_dereference_bh(prog->filter);
> > > > if (at_ingress) {
> > > > __skb_push(skb, skb->mac_len);
> > > > bpf_compute_data_pointers(skb);
> > > > @@ -56,7 +55,6 @@ static int tcf_bpf(struct sk_buff *skb, const struct
> > > > tc_action *act,
> > > > bpf_compute_data_pointers(skb);
> > > > filter_res = BPF_PROG_RUN(filter, skb);
> > > > }
> > > > - rcu_read_unlock();
> > >
> > > This conversion is not correct, BPF itself relies on RCU but not RCU-bh
> > > flavor.
> > > You might probably see a splat if you do e.g. a map lookup with this
> > > change in
> > > interpreter mode on tx side.
> >
> > Thank you for your review.
> >
> > I actually tested with lockdep, and lockdep is happy about it.
> >
> > The not so nice fact is that many TC modules already use plain RCU
> > primitives in the control path (call_rcu, kfree_rcu, etc.) and
> > rcu_derefence_bh() in the datapath (e.g. all the classifiers). AFACS,
> > despite the mix, this use is safe.
>
> Hmm, so out of __dev_queue_xmit() we do the RCU-bh read-side. We call
> into sch_handle_egress() which calls into tcf_classify() which may be
> a matchall one e.g. mall_classify(). It invokes tcf_exts_exec() that
> does the a->ops->act() which is the tcf_bpf() from here. If you then
> call a helper like bpf_map_lookup_elem(), there's a WARN_ON_ONCE() for
> !rcu_read_lock_held() since all of BPF is under normal RCU flavor. Why
> would that not trigger?
Because the basic sample I used did not call any other ebpf helper
beyond bpf_redirect(), I guess.
I see rcu_read_lock() is still needed here, thanks.
As a side note, after:
rcu_read_lock_bh()
rcu_read_lock();
both rcu_dereference() and rcu_derefernce_bh() are considered fine by
lockdep.
Cheers,
Paolo
