On Wed, Mar 20, 2019 at 12:57 PM Stanislav Fomichev <sdf@xxxxxxxxxxx> wrote:
>
> On 03/19, Willem de Bruijn wrote:
> > On Tue, Mar 19, 2019 at 6:21 PM Stanislav Fomichev <sdf@xxxxxxxxxx> wrote:
> > >
> > > Now that we have __flow_bpf_dissect which works on raw data (by
> > > constructing temporary on-stack skb), use it when doing
> > > BPF_PROG_TEST_RUN for flow dissector.
> > >
> > > This should help us catch any possible bugs due to missing shinfo on
> > > the per-cpu skb.
> > >
> > > Note that existing __skb_flow_bpf_dissect swallows L2 headers and returns
> > > nhoff=0, we need to preserve the existing behavior.
> > >
> > > Signed-off-by: Stanislav Fomichev <sdf@xxxxxxxxxx>
> > > ---
> > > net/bpf/test_run.c | 48 ++++++++++++++--------------------------------
> > > 1 file changed, 14 insertions(+), 34 deletions(-)
> > >
> >
> > > @@ -300,9 +277,13 @@ int bpf_prog_test_run_flow_dissector(struct bpf_prog *prog,
> > > preempt_disable();
> > > time_start = ktime_get_ns();
> > > for (i = 0; i < repeat; i++) {
> > > - retval = bpf_flow_dissect_skb(prog, skb,
> > > - &flow_keys_dissector,
> > > - &flow_keys);
> > > + retval = bpf_flow_dissect(prog, data, eth->h_proto, ETH_HLEN,
> > > + size, &flow_keys_dissector,
> > > + &flow_keys);
> > > + if (flow_keys.nhoff >= ETH_HLEN)
> > > + flow_keys.nhoff -= ETH_HLEN;
> > > + if (flow_keys.thoff >= ETH_HLEN)
> > > + flow_keys.thoff -= ETH_HLEN;
> >
> > why are these conditional?
> Hm, I didn't want these to be negative, because bpf flow program can set
> them to zero and clamp_flow_keys makes sure they are in a "sensible"
> range. For this particular case, I think we need to amend
> clamp_flow_keys to make sure that flow_keys.nhoff is in the range of
> initial_nhoff..hlen, not 0..hlen (and then we can drop these checks).
So, previously eth_type_trans would call with data at the network
header. Now it is called with data at the link layer. How would
__skb_flow_bpf_dissect "swallows L2 headers and returns nhoff=0"? That
sounds incorrect.
Agreed that the output should lie between nhoff and hlen, but as far
as I can tell it is always zero indexed to the data passed, here the
link layer:
if (!data) {
data = skb->data;
proto = skb_vlan_tag_present(skb) ?
skb->vlan_proto : skb->protocol;
nhoff = skb_network_offset(skb);
