Re: [PATCH bpf-next v2 4/8] xsk: register XDP sockets at bind(), and add new AF_XDP BPF helper

On Wed, Jan 20, 2021 at 12:26 PM Toke Høiland-Jørgensen [off-list ref] wrote:

This argument, however, I buy: bpf_redirect() is the single-purpose
helper for redirecting to an ifindex, bpf_redirect_xsk() is the
single-purpose helper for redirecting to an XSK, and bpf_redirect_map()
is the generic one that does both of those and more. Fair enough,
consider me convinced :)

quoted

A lot of back-and-forth for *one* if-statement, but it's kind of a
design thing for me. ;-)

Surely you don't mean to imply that you have *better* things to do with
your time than have a 10-emails-long argument over a single if
statement? ;)

After reading this thread I think I have to pour cold water on the design.

The performance blip comes from hard coded assumptions:
+ queue_id = xdp->rxq->queue_index;
+ xs = READ_ONCE(dev->_rx[queue_id].xsk);

bpf can have specialized helpers, but imo this is beyond what's reasonable.
Please move such things into the program and try to make
bpf_redirect_map faster.

Making af_xdp non-root is orthogonal. If there is actual need for that
it has to be designed thoroughly and not presented as "this helper may
help to do that".
I don't think "may" will materialize unless people actually work
toward the goal of non-root.