Re: [PATCH net-next v2 0/5] bpf: BPF for lightweight tunnel encapsulation

On Tue, Nov 1, 2016 at 12:11 PM, Thomas Graf [off-list ref] wrote:

On 1 November 2016 at 11:51, Tom Herbert [off-list ref] wrote:

quoted

On Tue, Nov 1, 2016 at 11:20 AM, Thomas Graf [off-list ref] wrote:

quoted

The headers cannot be extended or reduced so the offsets always remain
correct. What can happen is that the header contains invalid data.

If we can't add/remove headers then doesn't that really limit the
utility of these patches? My assumption was that BPF+LWT is needed to
allow users to define and implement their own encapsulations, EH
insertion, packet modification, etc.

I thought you were specifically referring to lwtunnel_output(). You
can extend headers at lwtunnel_xmit() but you can't for
lwtunnel_output() and lwtunnel_input(). The xmit hook is in
ip_finish_output2() where it is fine to do so. We have existing code
doing this there.

quoted

quoted

You keep saying this "middle in the stack" but the point is exactly
the same as a raw socket with IPPROTO_RAW and hdrincl, see
rawv6_sendmsg() and rawv6_send_hdrincl(). An IPv6 raw socket can feed
arbitrary garbage into dst_output(). IPv4 does some minimal sanity
checks.

What I mean is that an admin can create a BPF program that run on any
user packets (for instance default route could be set). This would be
in the path of TCP, UDP, and other protocols tightly integrated with
the stack. Packets being routed may be encapsulated, VLAN, have
checksum offload, GORed set etc. They also might be looped back in
which case the settings in skbuff become receive parameters.

You can do the same with act_pedit or cls_bpf at dev_queue_xmit()
before or after you go into the encapsulation device. This is a tool
for root, if you install a drop all ssh rule then you won't be able to
log into the box anymore. If you modify the packet and render it
invalid, the packet will be dropped.
If you attach a VLAN header while VLAN offload is already set up, then
the hardware will add another VLAN header, this is what I would
expect. I truly hope that we don't have code which crashes if we
dev_queue_xmit() garbage into any device. That would be horrible.

Do you have a specific example that could be a problem?

I believe Alexander was dealing with problems where drivers were not
properly handling IP extension headers. We regularly get reports about
driver checksum failures on edge conditions. Making a fully functional
and efficient transmit data path is nontrivial, there are many
assumptions being made some documented some not. When drivers crash we
either fix the driver or the protocol stack that is doing something
bad.

Tom