Re: [bpf-next PATCH 0/3] bpf: improvements to xdp_fwd sample

[Jesper Dangaard Brouer <brouer@redhat.com>]

On Wed, 7 Aug 2019 15:09:09 -0700
Zvi Effron <zeffron@riotgames.com> wrote:

> On Wed, Aug 7, 2019 at 6:00 AM Jesper Dangaard Brouer <brouer@redhat.com> wrote:
> >
> > Toke's devmap lookup improvement is first avail in kernel v5.3.
> > Thus, not part of XDP-tutorial yet.
> >  
> I probably missed this in an earlier email, but what are Toke's devmap
> improvements? Performance? Capability?

Toke's devmap and redirect improvements are primarily about usability.

Currently, from BPF-context (kernel-side) you cannot read the contents
of devmap (or cpumap or xskmap(AF_XDP)).  Because for devmap you get
the real pointer to the net_device ifindex, and we cannot allow you to
write/change that from BPF (kernel would likely crash or be inconsistent).

The work-around, is to keep a shadow map, that contains the "config" of
the devmap, which you check/validate against instead.  It is just a pain
to maintain this shadow map.  Toke's change allow you to read devmap
from BPF-context.  Thus, you can avoid this shadow map.

Another improvement from Toke, is that the bpf_redirect_map() helper,
now also check if the redirect index is valid in the map.  If not, then
it returns another value than XDP_REDIRECT.  You can choose the
alternative return value yourself, via "flags" e.g. XDP_PASS.  Thus,
you don't even need to check/validate devmap in your BPF-code, as it is
part of the bpf_redirect_map() call now.

 action = bpf_redirect_map(&map, &index, flags_as_xdp_value) 

The default flags used in most programs today is 0, which maps to
XDP_ABORTED.  This is sort of a small UAPI change, but for the better.
As today, the packet is dropped later, only diagnose/seen via
tracepoint xdp:xdp_redirect_map_err.

-- 
Best regards,
  Jesper Dangaard Brouer
  MSc.CS, Principal Kernel Engineer at Red Hat
  LinkedIn: http://www.linkedin.com/in/brouer