On Wed, Dec 2, 2020 at 5:32 PM Cong Wang <xiyou.wangcong@xxxxxxxxx> wrote:
>
> On Wed, Dec 2, 2020 at 5:10 PM Jakub Kicinski <kuba@xxxxxxxxxx> wrote:
> >
> > On Tue, 1 Dec 2020 11:44:38 -0800 Cong Wang wrote:
> > > From: Dongdong Wang <wangdongdong@xxxxxxxxxxxxx>
> > >
> > > The per-cpu bpf_redirect_info is shared among all skb_do_redirect()
> > > and BPF redirect helpers. Callers on RX path are all in BH context,
> > > disabling preemption is not sufficient to prevent BH interruption.
> > >
> > > In production, we observed strange packet drops because of the race
> > > condition between LWT xmit and TC ingress, and we verified this issue
> > > is fixed after we disable BH.
> > >
> > > Although this bug was technically introduced from the beginning, that
> > > is commit 3a0af8fd61f9 ("bpf: BPF for lightweight tunnel infrastructure"),
> > > at that time call_rcu() had to be call_rcu_bh() to match the RCU context.
> > > So this patch may not work well before RCU flavor consolidation has been
> > > completed around v5.0.
> > >
> > > Update the comments above the code too, as call_rcu() is now BH friendly.
> > >
> > > Cc: Thomas Graf <tgraf@xxxxxxx>
> > > Cc: bpf@xxxxxxxxxxxxxxx
> > > Reviewed-by: Cong Wang <cong.wang@xxxxxxxxxxxxx>
> > > Signed-off-by: Dongdong Wang <wangdongdong@xxxxxxxxxxxxx>
> > > ---
> > > net/core/lwt_bpf.c | 8 ++++----
> > > 1 file changed, 4 insertions(+), 4 deletions(-)
> > >
> > > diff --git a/net/core/lwt_bpf.c b/net/core/lwt_bpf.c
> > > index 7d3438215f32..4f3cb7c15ddf 100644
> > > --- a/net/core/lwt_bpf.c
> > > +++ b/net/core/lwt_bpf.c
> > > @@ -39,12 +39,11 @@ static int run_lwt_bpf(struct sk_buff *skb, struct bpf_lwt_prog *lwt,
> > > {
> > > int ret;
> > >
> > > - /* Preempt disable is needed to protect per-cpu redirect_info between
> > > - * BPF prog and skb_do_redirect(). The call_rcu in bpf_prog_put() and
> > > - * access to maps strictly require a rcu_read_lock() for protection,
> > > - * mixing with BH RCU lock doesn't work.
> > > + /* Preempt disable and BH disable are needed to protect per-cpu
> > > + * redirect_info between BPF prog and skb_do_redirect().
> > > */
> > > preempt_disable();
> > > + local_bh_disable();
> >
> > Why not remove the preempt_disable()? Disabling BH must also disable
> > preemption AFAIK.
>
> It seems RT kernel still needs preempt disable:
No. It's the opposite. When we did RT+bpf changes we missed this function.
It should be migrate_disable here instead of preempt_disable.
I don't know what local_bh_disable() maps to in RT.
Since it's used in many other places it's fine to use it here to
prevent this race.
