Re: [PATCH bpf-next 1/5] libbpf: support XDP_SHARED_UMEM with external XDP program

[William Tu <u9012063@gmail.com>]

On Fri, Nov 08, 2019 at 08:17:53PM +0100, Magnus Karlsson wrote:
> On Fri, Nov 8, 2019 at 7:43 PM William Tu <u9012063@gmail.com> wrote:
> >
> > On Fri, Nov 08, 2019 at 07:19:18PM +0100, Magnus Karlsson wrote:
> > > On Fri, Nov 8, 2019 at 7:03 PM William Tu <u9012063@gmail.com> wrote:
> > > >
> > > > Hi Magnus,
> > > >
> > > > Thanks for the patch.
> > > >
> > > > On Thu, Nov 07, 2019 at 06:47:36PM +0100, Magnus Karlsson wrote:
> > > > > Add support in libbpf to create multiple sockets that share a single
> > > > > umem. Note that an external XDP program need to be supplied that
> > > > > routes the incoming traffic to the desired sockets. So you need to
> > > > > supply the libbpf_flag XSK_LIBBPF_FLAGS__INHIBIT_PROG_LOAD and load
> > > > > your own XDP program.
> > > > >
> > > > > Signed-off-by: Magnus Karlsson <magnus.karlsson@intel.com>
> > > > > ---
> > > > >  tools/lib/bpf/xsk.c | 27 +++++++++++++++++----------
> > > > >  1 file changed, 17 insertions(+), 10 deletions(-)
> > > > >
> > > > > diff --git a/tools/lib/bpf/xsk.c b/tools/lib/bpf/xsk.c
> > > > > index 86c1b61..8ebd810 100644
> > > > > --- a/tools/lib/bpf/xsk.c
> > > > > +++ b/tools/lib/bpf/xsk.c
> > > > > @@ -586,15 +586,21 @@ int xsk_socket__create(struct xsk_socket **xsk_ptr, const char *ifname,
> > > > >       if (!umem || !xsk_ptr || !rx || !tx)
> > > > >               return -EFAULT;
> > > > >
> > > > > -     if (umem->refcount) {
> > > > > -             pr_warn("Error: shared umems not supported by libbpf.\n");
> > > > > -             return -EBUSY;
> > > > > -     }
> > > > > -
> > > > >       xsk = calloc(1, sizeof(*xsk));
> > > > >       if (!xsk)
> > > > >               return -ENOMEM;
> > > > >
> > > > > +     err = xsk_set_xdp_socket_config(&xsk->config, usr_config);
> > > > > +     if (err)
> > > > > +             goto out_xsk_alloc;
> > > > > +
> > > > > +     if (umem->refcount &&
> > > > > +         !(xsk->config.libbpf_flags & XSK_LIBBPF_FLAGS__INHIBIT_PROG_LOAD)) {
> > > > > +             pr_warn("Error: shared umems not supported by libbpf supplied XDP program.\n");
> > > >
> > > > Why can't we use the existing default one in libbpf?
> > > > If users don't want to redistribute packet to different queue,
> > > > then they can still use the libbpf default one.
> > >
> > > Is there any point in creating two or more sockets tied to the same
> > > umem and directing all traffic to just one socket? IMHO, I believe
> >
> > When using build-in XDP, isn't the traffic being directed to its
> > own xsk on its queue? (so not just one xsk socket)
> >
> > So using build-in XDP, for example, queue1/xsk1 and queue2/xsk2, and
> > sharing one umem. Both xsk1 and xsk2 receive packets from their queue.
> 
> WIth the XDP_SHARED_UMEM flag this is not allowed. In your example,
> queue1/xsk1 and queue1/xsk2 would be allowed. All sockets need to be
> tied to the same queue id if they share a umem. In this case an XDP
> program has to decide how to distribute the packets since they all
> arrive on the same queue.
> 
> If you want queue1/xsk1 and queue2/xsk2 you need separate umems since
> it would otherwise violate the SPSC requirement or the rings. Or
> implement MPSC and SPMC queues to be used in this configuration.
> 
> > > that most users in this case would want to distribute the packets over
> > > the sockets in some way. I also think that users might be unpleasantly
> > > surprised if they create multiple sockets and all packets only get to
> > > a single socket because libbpf loaded an XDP program that makes little
> > > sense in the XDP_SHARED_UMEM case. If we force them to supply an XDP
> >
> > Do I misunderstand the code?
> > I looked at xsk_setup_xdp_prog, xsk_load_xdp_prog, and xsk_set_bpf_maps.
> > The build-in prog will distribute packets to different xsk sockets,
> > not a single socket.
> 
> True, but only for the case above (queue1/xsk1 and queue2/xsk2) where
> they have separate umems. For the queue1/xsk1 and queue1/xsk2 case, it
> would send everything to xsk1.
> 
> /Magnus

Hi Magnus,

Thanks for your explanation. Now I understand.

William