Re: [PATCH bpf-next 03/12] net: Add IFLA_XDP_EGRESS for XDP programs in the egress path

On Tue, 28 Jan 2020 15:13:43 +0100, Jesper Dangaard Brouer wrote:

On Sun, 26 Jan 2020 14:17:01 -0800
Jakub Kicinski [off-list ref] wrote:

quoted

On Sun, 26 Jan 2020 13:49:33 +0100, Jesper Dangaard Brouer wrote:  

quoted

Yes, please. I want this NIC TX hook to see both SKBs and xdp_frames.    

Any pointers on what for? Unless we see actual use cases there's
a justifiable concern of the entire thing just being an application of
"We can solve any problem by introducing an extra level of indirection."  

I have two use-cases:

(1) For XDP easier handling of interface specific setting on egress,
e.g. pushing a VLAN-id, instead of having to figure this out in RX hook.
(I think this is also David Ahern's use-case)

Is it really useful to have a hook before multi-buffer frames are
possible and perhaps TSO? The local TCP performance is going to tank
with XDP enabled otherwise.

(2) I want this egress XDP hook to have the ability to signal
backpressure. Today we have BQL in most drivers (which is essential to
avoid bufferbloat). For XDP_REDIRECT we don't, which we must solve.

For use-case(2), we likely need a BPF-helper calling netif_tx_stop_queue(),
or a return code that can stop the queue towards the higher layers.

Agreed, although for that use case, I'm not sure if non-XDP frames 
have to pass trough the hook. Hard to tell as the current patches 
don't touch on this use case.