On 2021-09-30 07:59:59 -0700, Jakub Kicinski wrote:

On Thu, 30 Sep 2021 16:46:34 +0200 Niklas Söderlund wrote:

quoted

When the MTU is changed after the program is offloaded the check in 
nfp_bpf_check_mtu() is consulted and as it checks the MTU differently 
and fails the change. Maybe we should align this the other way around 
and update the check in nfp_bpf_check_mtu() to match the one in 
nfp_net_bpf_load()?

That sounds reasonable. Although I don't remember how reliable the
max_pkt_offset logic is in practice (whether it's actually capable 
of finding the max offset for realistic programs or it's mostly going
to be set to MAX).

quoted

On a side note the check in nfp_net_bpf_load() allows for BPF programs 
to be offloaded that do access data beyond the CMT size limit provided 
the MTU is set below the CMT threshold value.

Right, because of variable length offsets verifier will not be able to
estimate max_pkt_offset.

Thanks, this made the design click for me.

quoted

There should be no real harm in this as the verifier forces bounds
check so with a MTU small enough it should never happen. But maybe we
should add a check for this too to prevent such a program to be
loaded in the first place.

-- 
Regards,
Niklas Söderlund