On 04/29/2014 10:59 PM, Joseph Gasparakis wrote:

On Tue, 29 Apr 2014, Vlad Yasevich wrote:

...

quoted

quoted

quoted

quoted

quoted

At the time SCTP tries to do checksum, it will see hw offload
support and set checksum as partial.  That will be propagated
all the way to vxlan which will set encapsulation and pass it on.
At next dev_hard_start_xmit(), we'll look at hw_enc_features,
but it won't matter since SCTP packets are never GSO.

So, we'll use the non-gso code path and still send with partial
checksum set.  Now, if the driver supports encapsulated SCTP,
then it will work.  If it doesn't, the udp checksum will be computed,
but not the sctp one.

That is exactly the way it works, dev_hard_start_xmit() is where
the (inner) unchecksummed packets will get checksummed if the Tx'ing
interface does not support inner SCTP checksumming.

No, it will not.  dev_hard_start_xmit() calls skb_checksum_help()
which only knows how to calculate TCP/UDP checksums.  It doesn't know
how to compute CRC32c for SCTP protocol.

Interesting, thanks for the clarification! Then how do things work with
non-encapsulated SCTP traffic? Where is the csumming for SCTP take place
if the NIC cannot offload the protocol?

In SCTP itself.  Look at sctp_packet_transmit().   There are also
fix-ups in netfilter to correct the checksum if the packet has
to be rerouted somewhere else.

We've tossed around the idea of making skb_checksum have pluggable
components.  Daniel started us down that path by adding checksum ops.
We just need registration and proper lookup support.

Absolutely, Vlad is right here. That's also why we currently have this
ugly xfrm test in SCTP output path for csumming. I've started something
a while ago in that direction but got distracted with other things. If
someone is faster than me, I appreciate it. ;) In case not, I've kept
it on my todo list so far. Anyway, that would be the right way to go.