Re: [PATCH net-next] net: Remove VRF change to udp_sendmsg

[David Ahern <dsa@cumulusnetworks.com>]

On 8/31/15 1:44 PM, David Miller wrote:
> From: David Ahern <dsa@cumulusnetworks.com>
> Date: Mon, 31 Aug 2015 09:29:40 -0700
>
>> Remove the VRF change in udp_sendmsg to set the source address. The VRF
>> driver already has access to the packet on the TX path via the dst. It
>> can be used to update the source address in the header.
>>
>> Update function based on OVS.
>>
>> Cc: Tom Herbert <tom@herbertland.com>
>> Signed-off-by: David Ahern <dsa@cumulusnetworks.com>
>
> This is worse.
>
> You have the source address in the VRF driver's output routine in
> fl4.saddr, just use it as-is.

The original saddr code in vrf_xmit was just wrong. It should not have 
been there and was a leftover from early development days.

The saddr needs to be set in the dst output function (vrf_output) as I 
did in this patch otherwise the packet hits the tcpdump tap with 
potentially no source address.

>
> You're adding even more route lookups, at least the existing code
> just walks the device address less which is often cheaper than
> a full-blown route lookup.
>

TCP (and connected sockets) do not hit this lookup in vrf_output.

If anything I should be going straight to fib_table_lookup in the VRF 
driver for this new lookup to get the source address. It knows the exact 
table that should be used and hence can avoid the rules walk + local 
table miss which happens using the ip_route_xxxxx functions as well as 
the rth lookup/create which is not needed here.

Opinions before I work on another version?

David