On Thu, 2017-02-02 at 10:36 -0800, Tom Herbert wrote: > It's going to be a problem with a whole host of application level > protocols especially those run over UDP. QUIC is a great example. The > actual protocol will probably only ever run in userspace, but it is > inevitable that we want to provide targeted kernel support for packet > steering. filtering, GRO/GSO if they have such things. Instead of > implementing this in a specialized QUIC module, it will most likely > make everyone happier to add these in a generic protocol-agnostic way. > From QUIC POV they want to minimize any dependencies on the kernel and > be able to iterate quickly, from a kernel POV we really don't want to > have to explicitly support an endless stream of protocols like this.
Small note here : Google does not intend to add QUIC knowledge anywhere in the kernel. SO_ATTACH_REUSEPORT_EBPF can be used if someone wants to take a deep look at QUIC header (and presumably hash the connection-id) to select one socket among the group. Thanks.
