Re: [PATCH bpf 0/2] Revert "xsk: support redirect to any socket bound to the same umem"

[Magnus Karlsson <magnus.karlsson@gmail.com>]

On Wed, 5 Jun 2024 at 01:03, Willem de Bruijn
<willemdebruijn.kernel@gmail.com> wrote:
>
> Magnus Karlsson wrote:
> > Revert "xsk: support redirect to any socket bound to the same umem"
> >
> > This patch introduced a potential kernel crash when multiple napi
> > instances redirect to the same AF_XDP socket. By removing the
> > queue_index check, it is possible for multiple napi instances to
> > access the Rx ring at the same time, which will result in a corrupted
> > ring state which can lead to a crash when flushing the rings in
> > __xsk_flush(). This can happen when the linked list of sockets to
> > flush gets corrupted by concurrent accesses. A quick and small fix is
> > unfortunately not possible, so let us revert this for now.
>
> This is a very useful feature, to be able to use AF_XDP sockets with
> a standard RSS nic configuration.

I completely agree.

> Not all AF_XDP use cases require the absolute highest packet rate.
>
> Can this be addressed with an optional spinlock on the RxQ, only for
> this case?

Yes, or with a MPSC ring implementation.

> If there is no simple enough fix in the short term, do you plan to
> reintroduce this in another form later?

Yuval and I are looking into a solution based around an optional
spinlock since it is easier to pull off than an MPSC ring. The
discussion is on-going on the xdp-newbies list [0], but as soon as we
have a first patch, we will post it here for review and debate.

[0] https://lore.kernel.org/xdp-newbies/8100DBDC-0B7C-49DB-9995-6027F6E63147@radware.com/