Re: [PATCH net-next v2 0/5] bpf: BPF for lightweight tunnel encapsulation
From: Thomas Graf <tgraf@suug.ch>
Date: 2016-11-01 18:20:33
On 1 November 2016 at 09:17, Tom Herbert [off-list ref] wrote:
On Mon, Oct 31, 2016 at 5:37 PM, Thomas Graf [off-list ref] wrote:quoted
{Open question: Tom brought up the question on whether it is safe to modify the packet in artbirary ways before dst_output(). This is the equivalent to a raw socket injecting illegal headers. This v2 currently assumes that dst_output() is ready to accept invalid header values. This needs to be verified and if not the case, then raw sockets or dst_output() handlers must be fixed as well. Another option is to mark lwtunnel_output() as read-only for now.}The question might not be so much about illegal headers but whether fields in the skbuff related to the packet contents are kept correct. We have protocol, header offsets, offsets for inner protocols also, encapsulation settings, checksum status, checksum offset, checksum
The headers cannot be extended or reduced so the offsets always remain correct. What can happen is that the header contains invalid data.
complete value, vlan information. Any or all of which I believe could be turned into being incorrect if we allow the packet to be arbitrarily modified by BPF. This problem is different than raw sockets because LWT operates in the middle of the stack, the skbuff has already been set up which such things.
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. If this is a concern I'm fine with making the dst_output path read-only for now.
quoted
This series implements BPF program invocation from dst entries via the lightweight tunnels infrastructure. The BPF program can be attached to lwtunnel_input(), lwtunnel_output() or lwtunnel_xmit() and sees an L3 skb as context. input is read-only, output can write, xmit can write, push headers, and redirect. Motiviation for this work: - Restricting outgoing routes beyond what the route tuple supports - Per route accounting byond realms - Fast attachment of L2 headers where header does not require resolving L2 addresses - ILA like uses cases where L3 addresses are resolved and then routed in an async manner - Fast encapsulation + redirect. For now limited to use cases where not setting inner and outer offset/protocol is OK.Is checksum offload supported? By default, at least for Linux, we offload the outer UDP checksum in VXLAN and the other UDP encapsulations for performance.
No. UDP encap is done by setting a tunnel key through a helper and letting the encapsulation device handle this. I don't currently see a point in replicating all of that logic.