Re: [PATCH net-next 12/18] IB/mlx5: Add kernel offload flow-tag

On Tue, Jun 21, 2016 at 5:18 AM, Alexei Starovoitov
[off-list ref] wrote:

On Sat, Jun 18, 2016 at 01:31:26AM +0300, Saeed Mahameed wrote:

quoted

We simply want to selectively be able to see RoCE/RDMA ETH standard
traffic in tcpdump, for diagnostic purposes.
so in order to not overwhelm the kernel TCP/IP stack with this
traffic, this patch in particular
configures ConnectX4 hardware to tag those packets, so in downstream
patches mlx5 netdevice will mark the SKBs of those packets
to skip the TCP/IP stack and go only to tcpdump.

such 'feature' doesn't make sense to me.
'not overwhelming' kernel, but to 'overwhelm' userspace tcpdump?
Kernel can drop packets way faster than userspace, so such bypass
scheme has no prartical usage other than building a first step towards
complete dpdk-like bypass.

Alexei , I don't understand your concern.
We already have a full/complete working dpdk bypass solution in
userspace nothing extra is required from the kernel.

We just want to see this traffic and any other rdma traffic in tcpdump
or other standard sniffing tools.

Anyway we brainstormed this internally today and we don't like the
"skb->protocol = 0xffff" solution,
we will suggest a plugin for libpcap in user space to extend libpcap
ability to sniff RDMA/raw eth traffic.
This way userspace RDMA traffic will be sniffed also via a userspace
RDMA channel.

I will ask Dave to drop this series.