Possible issue with multipath when used in bridge mode

[joseph-henry]

Ticket created for tracking purposes

[andrejbinder]

I re-read through all the code and I believe I have found what is causing the problem. When receiving a packet we set the packet flowId to ZT_QOS_NO_FLOW and then only if we hit a packet that is of the VERB_FRAME type we look into its guts and set the flowId based on the actual protocol source/destination port numbers.

The catch is that packets that have been bridged are not VERB_FRAME but they are VERB_EXT_FRAME which contains no code to set the flowId correctly. This is why in-flow assignments do not work for packets from bridged hosts.

Out-flow assignments work perfectly for those because the flowId calculation is done already in Switch::onLocalEthernet of switch.cpp regardless of whether this is a VERB_FRAME or VERB_EXT_FRAME packet.

[andrejbinder]

I have also found what looks to me like a typo in the header length discriminator for VERB_FRAME type packets (it was populating flowId only when the packet was big enough to contain a VERB_EXT_FRAME header and at least some payload. I think this was meant to match against VERB_FRAME header size instead. This could make small (VERB_FRAME) packets not to have flowId populated.

[andrejbinder]

I put together a small patch to fix both of the issues above:

ab.patch

Quick testing seems to show that this has fixed the issue and in-flows now work correctly even for bridged packets. Please review the patch and consider upstreaming.

[joseph-henry]

This looks good and seems to work. I've made a PR out of it (#2306) but if you'd rather make your own PR that's fine too. Just let us know what you'd prefer and we can upstream it.