Alexei Starovoitov wrote:
> On Tue, Mar 03, 2020 at 12:46:58PM +0100, Jesper Dangaard Brouer wrote:
> > The Intel based drivers (ixgbe + i40e) have implemented XDP with
> > headroom 192 bytes and not the recommended 256 bytes defined by
> > XDP_PACKET_HEADROOM. For generic-XDP, accept that this headroom
> > is also a valid size.
The reason is to fit two packets on a 4k page. The driver itself
is fairly flexible at this point. I think we should reconsider
pushing down the headroom required in the program metadata and
configuring it at runtime. At the moment the drivers are wasting
half a page for no good reason in most cases I suspect. What is the
use case for >192B headroom? I've not found an actual user who
has complained yet.
Resurrecting an old debate here so probably doesn't need to
stall this patch.
> >
> > Still for generic-XDP if headroom is less, still expand headroom to
> > XDP_PACKET_HEADROOM as this is the default in most XDP drivers.
> >
> > Tested on ixgbe with xdp_rxq_info --skb-mode and --action XDP_DROP:
> > - Before: 4,816,430 pps
> > - After : 7,749,678 pps
> > (Note that ixgbe in native mode XDP_DROP 14,704,539 pps)
> >
But why do we care about generic-XDP performance? Seems users should
just use XDP proper on ixgbe and i40e its supported.
> > Signed-off-by: Jesper Dangaard Brouer <brouer@xxxxxxxxxx>
> > ---
> > include/uapi/linux/bpf.h | 1 +
> > net/core/dev.c | 4 ++--
> > 2 files changed, 3 insertions(+), 2 deletions(-)
> >
> > diff --git a/include/uapi/linux/bpf.h b/include/uapi/linux/bpf.h
> > index 906e9f2752db..14dc4f9fb3c8 100644
> > --- a/include/uapi/linux/bpf.h
> > +++ b/include/uapi/linux/bpf.h
> > @@ -3312,6 +3312,7 @@ struct bpf_xdp_sock {
> > };
> >
> > #define XDP_PACKET_HEADROOM 256
> > +#define XDP_PACKET_HEADROOM_MIN 192
>
> why expose it in uapi?
>
> > /* User return codes for XDP prog type.
> > * A valid XDP program must return one of these defined values. All other
> > diff --git a/net/core/dev.c b/net/core/dev.c
> > index 4770dde3448d..9c941cd38b13 100644
> > --- a/net/core/dev.c
> > +++ b/net/core/dev.c
> > @@ -4518,11 +4518,11 @@ static u32 netif_receive_generic_xdp(struct sk_buff *skb,
> > return XDP_PASS;
> >
> > /* XDP packets must be linear and must have sufficient headroom
> > - * of XDP_PACKET_HEADROOM bytes. This is the guarantee that also
> > + * of XDP_PACKET_HEADROOM_MIN bytes. This is the guarantee that also
> > * native XDP provides, thus we need to do it here as well.
> > */
> > if (skb_cloned(skb) || skb_is_nonlinear(skb) ||
> > - skb_headroom(skb) < XDP_PACKET_HEADROOM) {
> > + skb_headroom(skb) < XDP_PACKET_HEADROOM_MIN) {
> > int hroom = XDP_PACKET_HEADROOM - skb_headroom(skb);
>
> this looks odd. It's comparing against 192, but doing math with 256.
> I guess that's ok, but needs a clear comment.
> How about just doing 'skb_headroom(skb) < 192' here.
> Or #define 192 right before this function with a comment about ixgbe?
Or just let ixgbe/i40e be slow? I guess I'm missing some context?
