From: Toke Høiland-Jørgensen <t...@redhat.com>
Date: Fri, 29 May 2020 14:43:44 +0200
> While the other fq-based qdiscs take advantage of skb->hash and doesn't
> recompute it if it is already set, sch_cake does not.
>
> This was a deliberate choice because sch_cake hashes various parts of the
> packet header to support its advanced flow isolation modes. However,
> foregoing the use of skb->hash entirely loses a few important benefits:
>
> - When skb->hash is set by hardware, a few CPU cycles can be saved by not
> hashing again in software.
>
> - Tunnel encapsulations will generally preserve the value of skb->hash from
> before the encapsulation, which allows flow-based qdiscs to distinguish
> between flows even though the outer packet header no longer has flow
> information.
>
> It turns out that we can preserve these desirable properties in many cases,
> while still supporting the advanced flow isolation properties of sch_cake.
> This patch does so by reusing the skb->hash value as the flow_hash part of
> the hashing procedure in cake_hash() only in the following conditions:
>
> - If the skb->hash is marked as covering the flow headers (skb->l4_hash is
> set)
>
> AND
>
> - NAT header rewriting is either disabled, or did not change any values
> used for hashing. The latter is important to match local-origin packets
> such as those of a tunnel endpoint.
>
> The immediate motivation for fixing this was the recent patch to WireGuard
> to preserve the skb->hash on encapsulation. As such, this is also what I
> tested against; with this patch, added latency under load for competing
> flows drops from ~8 ms to sub-1ms on an RRUL test over a WireGuard tunnel
> going through a virtual link shaped to 1Gbps using sch_cake. This matches
> the results we saw with a similar setup using sch_fq_codel when testing the
> WireGuard patch.
>
> Fixes: 046f6fd5daef ("sched: Add Common Applications Kept Enhanced (cake)
> qdisc")
> Signed-off-by: Toke Høiland-Jørgensen <t...@redhat.com>
Applied to net-next, thanks.
