Re: [RFC v2 bpf-next 2/7] drivers: net: turn on XDP features

["Toke Høiland-Jørgensen" <toke@redhat.com>]

Niklas Söderlund <niklas.soderlund@corigine.com> writes:

> On 2023-01-17 23:15:47 +0100, Toke Høiland-Jørgensen wrote:
>> Niklas Söderlund <niklas.soderlund@corigine.com> writes:
>> 
>> > Hi Toke,
>> >
>> > On 2023-01-17 22:58:57 +0100, Toke Høiland-Jørgensen wrote:
>> >> Niklas Söderlund <niklas.soderlund@corigine.com> writes:
>> >> 
>> >> > Hi Lorenzo and Marek,
>> >> >
>> >> > Thanks for your work.
>> >> >
>> >> > On 2023-01-14 16:54:32 +0100, Lorenzo Bianconi wrote:
>> >> >
>> >> > [...]
>> >> >
>> >> >> 
>> >> >> Turn 'hw-offload' feature flag on for:
>> >> >>  - netronome (nfp)
>> >> >>  - netdevsim.
>> >> >
>> >> > Is there a definition of the 'hw-offload' written down somewhere? From 
>> >> > reading this series I take it is the ability to offload a BPF program?  
>> >> 
>> >> Yeah, basically this means "allows loading and attaching programs in
>> >> XDP_MODE_HW", I suppose :)
>> >> 
>> >> > It would also be interesting to read documentation for the other flags 
>> >> > added in this series.
>> >> 
>> >> Yup, we should definitely document them :)
>> >> 
>> >> > [...]
>> >> >
>> >> >> diff --git a/drivers/net/ethernet/netronome/nfp/nfp_net_common.c 
>> >> >> b/drivers/net/ethernet/netronome/nfp/nfp_net_common.c
>> >> >> index 18fc9971f1c8..5a8ddeaff74d 100644
>> >> >> --- a/drivers/net/ethernet/netronome/nfp/nfp_net_common.c
>> >> >> +++ b/drivers/net/ethernet/netronome/nfp/nfp_net_common.c
>> >> >> @@ -2529,10 +2529,14 @@ static void nfp_net_netdev_init(struct nfp_net *nn)
>> >> >>  	netdev->features &= ~NETIF_F_HW_VLAN_STAG_RX;
>> >> >>  	nn->dp.ctrl &= ~NFP_NET_CFG_CTRL_RXQINQ;
>> >> >>  
>> >> >> +	nn->dp.netdev->xdp_features = NETDEV_XDP_ACT_BASIC |
>> >> >> +				      NETDEV_XDP_ACT_HW_OFFLOAD;
>> >> >
>> >> > If my assumption about the 'hw-offload' flag above is correct I think 
>> >> > NETDEV_XDP_ACT_HW_OFFLOAD should be conditioned on that the BPF firmware 
>> >> > flavor is in use.
>> >> >
>> >> >     nn->dp.netdev->xdp_features = NETDEV_XDP_ACT_BASIC;
>> >> >
>> >> >     if (nn->app->type->id == NFP_APP_BPF_NIC)
>> >> >         nn->dp.netdev->xdp_features |= NETDEV_XDP_ACT_HW_OFFLOAD;
>> >> >
>> >> >> +
>> >> >>  	/* Finalise the netdev setup */
>> >> >>  	switch (nn->dp.ops->version) {
>> >> >>  	case NFP_NFD_VER_NFD3:
>> >> >>  		netdev->netdev_ops = &nfp_nfd3_netdev_ops;
>> >> >> +		nn->dp.netdev->xdp_features |= NETDEV_XDP_ACT_XSK_ZEROCOPY;
>> >> >>  		break;
>> >> >>  	case NFP_NFD_VER_NFDK:
>> >> >>  		netdev->netdev_ops = &nfp_nfdk_netdev_ops;
>> >> >
>> >> > This is also a wrinkle I would like to understand. Currently NFP support 
>> >> > zero-copy on NFD3, but not for offloaded BPF programs. But with the BPF 
>> >> > firmware flavor running the device can still support zero-copy for 
>> >> > non-offloaded programs.
>> >> >
>> >> > Is it a problem that the driver advertises support for both 
>> >> > hardware-offload _and_ zero-copy at the same time, even if they can't be 
>> >> > used together but separately?
>> >> 
>> >> Hmm, so the idea with this is to only expose feature flags that are
>> >> supported "right now" (you'll note that some of the drivers turn the
>> >> REDIRECT_TARGET flag on and off at runtime). Having features that are
>> >> "supported but in a different configuration" is one of the points of
>> >> user confusion we want to clear up with the explicit flags.
>> >> 
>> >> So I guess it depends a little bit what you mean by "can't be used
>> >> together"? I believe it's possible to load two programs at the same
>> >> time, one in HW mode and one in native (driver) mode, right? In this
>> >> case, could the driver mode program use XSK zerocopy while the HW mode
>> >> program is also loaded?
>> >
>> > Exactly, this is my concern. Two programs can be loaded at the same 
>> > time, one in HW mode and one in native mode. The program in native mode 
>> > can use zero-copy at the same time as another program runs in HW mode.
>> >
>> > But the program running in HW mode can never use zero-copy.
>> 
>> Hmm, but zero-copy is an AF_XDP feature, and AFAIK offloaded programs
>> can't use AF_XDP at all? So the zero-copy "feature" is available on the
>> hardware, it's just intrinsic to that feature that it doesn't work on
>> offloaded programs?
>
> That is true, so this is indeed not an issue then. Thanks for the 
> clarification.

Cool - you're welcome :)

-Toke