Re: [RFC 0/5] add XDP support to mt76x2e/mt76x0e drivers

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

On Wed, 28 Nov 2018 13:36:26 +0100
Toke Høiland-Jørgensen <toke@toke.dk> wrote:

> Lorenzo Bianconi <lorenzo.bianconi@redhat.com> writes:
> 
> >> Lorenzo Bianconi <lorenzo.bianconi@redhat.com> writes:
> >>   
> >> > This series is intended as a playground to start experimenting/developing
> >> > with XDP/eBPF over WiFi and collect ideas/concerns about it.
> >> > Introduce XDP support to mt76x2e/mt76x0e drivers. Currently supported
> >> > actions are:
> >> > - XDP_PASS
> >> > - XDP_ABORTED
> >> > - XDP_DROP
> >> > Introduce ndo_bpf mac80211 callback in order to to load a bpf
> >> > program into low level driver XDP rx hook.
> >> > This series has been tested through a simple bpf program (available here:
> >> > https://github.com/LorenzoBianconi/bpf-workspace/tree/master/mt76_xdp_stats)
> >> > used to count frame types received by the device.
> >> > Possible eBPF use cases could be:
> >> > - implement new statistics through bpf maps
> >> > - implement fast packet filtering (e.g in monitor mode)
> >> > - ...  
> >
> > Hi Kalle,
> >  
> >> 
> >> This is most likely a stupid question, but why do this in the driver and
> >> not in mac80211 so that all drivers could benefit from it? I guess there
> >> are reasons for that, I just can't figure that out.  
> 
> XDP achieves its speedup by running the eBPF program inside the driver
> NAPI loop, before the kernel even touches the data in any other capacity
> (and in particular, before it allocates an SKB). Which kinda means the
> hook needs to be in the driver... Could be a fallback in mac80211,
> though; although we'd have to figure out how that interacts with Generic
> XDP.
> 
> > This is an early stage implementation, at this point I would collect
> > other people opinions/concerns about using bpf/xdp directly on 802.11
> > frames.  
> 
> Thanks for looking into this!
> 
> I have two concerns with running XDP on 802.11 frames:
> 
> 1. It makes it more difficult to add other XDP actions (such as
>    REDIRECT), as the XDP program would then have to make sure that the
>    outer packet headers are removed before, say, redirecting the packet
>    out of an ethernet interface. Also, if we do add redirect, we would
>    be bypassing mac80211 entirely; to what extent would that mess up
>    internal state?
> 
> 2. UI consistency; suddenly, the user needs to know which kind of
>    frames to expect, and XDP program reuse becomes more difficult. This
>    may be unavoidable given the nature of XDP, but some thought needs to
>    go into this. Especially since we wouldn't necessarily be consistent
>    between WiFi drivers (there are fullmac devices that remove 802.11
>    headers before sending up the frame, right?).
> 
> 
> Adding in Jesper; maybe he has some thoughts on this?

Today XDP assumes the frame is an Ethernet frame.  With WiFi I guess
this assumption change, right?
  I worry a bit about this, as XDP is all about performance, and I don't
want to add performance regressions, by requiring all XDP programs or
core-code to having to check-frame-type before proceeding. That said, I
do think it is doable, without adding performance regressions.

Option #1 is to move the check-frame-type to setup time.  By either
having frame-type be part of eBPF prog, or supply frame-type as option
XDP attach call.  And then reject attaching XDP prog to a device, where
the expected frame-type does not match.

Option#2, leave it up to eBPF-programmer if they want to add runtime
checks.  By extending xdp_rxq_info with frame-type (default to
Ethernet), which allow the eBPF-programmer choose to write a generic
XDP program that both work on Ethernet and WiFi, or skip-check as they
know this will e.g. only run on Wifi.  (Note xdp_rxq_info is static
read-only info per RX-queue, will all Wifi frames have same frame-type?.


Also consider what happens in case of XDP_REDIRECT, from a Wifi NIC to
an Ethernet NIC.  It would of-cause be cool to get this working cross,
Wifi-Ethernet.

Option#3 is to say, Wifi XDP is so different that we should create a
new (enum) bpf_prog_type.  And then still see if we can leverage some
of the same core-code (as long as it doesn't slowdown performance).

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