Stanislav Fomichev wrote:
> On 07/20, Zaremba, Larysa wrote:
> > On Thu, Jul 20, 2023 at 08:14:52AM -0700, Alexei Starovoitov wrote:
> > > On Thu, Jul 20, 2023 at 2:47 AM Zaremba, Larysa
> > > <larysa.zaremba@xxxxxxxxx> wrote:
> > > >
> > > > On Wed, Jul 19, 2023 at 05:51:17PM -0400, Willem de Bruijn wrote:
> > > > > Alexei Starovoitov wrote:
> > > > > > On Wed, Jul 19, 2023 at 08:37:26PM +0200, Larysa Zaremba wrote:
> > > > > > > Implement .xmo_rx_csum callback to allow XDP code to determine,
> > > > > > > whether HW has validated any checksums.
> > > > > > >
> > > > > > > Signed-off-by: Larysa Zaremba <larysa.zaremba@xxxxxxxxx>
> > > > > > > ---
> > > > > > > drivers/net/ethernet/intel/ice/ice_txrx_lib.c | 29 +++++++++++++++++++
> > > > > > > 1 file changed, 29 insertions(+)
> > > > > > >
> > > > > > > diff --git a/drivers/net/ethernet/intel/ice/ice_txrx_lib.c b/drivers/net/ethernet/intel/ice/ice_txrx_lib.c
> > > > > > > index 54685d0747aa..6647a7e55ac8 100644
> > > > > > > --- a/drivers/net/ethernet/intel/ice/ice_txrx_lib.c
> > > > > > > +++ b/drivers/net/ethernet/intel/ice/ice_txrx_lib.c
> > > > > > > @@ -660,8 +660,37 @@ static int ice_xdp_rx_vlan_tag(const struct xdp_md *ctx, u16 *vlan_tci,
> > > > > > > return 0;
> > > > > > > }
> > > > > > >
> > > > > > > +/**
> > > > > > > + * ice_xdp_rx_csum_lvl - Get level, at which HW has checked the checksum
> > > > > > > + * @ctx: XDP buff pointer
> > > > > > > + * @csum_status: destination address
> > > > > > > + * @csum_info: destination address
> > > > > > > + *
> > > > > > > + * Copy HW checksum level (if was checked) to the destination address.
> > > > > > > + */
> > > > > > > +static int ice_xdp_rx_csum(const struct xdp_md *ctx,
> > > > > > > + enum xdp_csum_status *csum_status,
> > > > > > > + union xdp_csum_info *csum_info)
> > > > > > > +{
> > > > > > > + const struct ice_xdp_buff *xdp_ext = (void *)ctx;
> > > > > > > + const union ice_32b_rx_flex_desc *eop_desc;
> > > > > > > + enum ice_rx_csum_status status;
> > > > > > > + u16 ptype;
> > > > > > > +
> > > > > > > + eop_desc = xdp_ext->pkt_ctx.eop_desc;
> > > > > > > + ptype = ice_get_ptype(eop_desc);
> > > > > > > +
> > > > > > > + status = ice_get_rx_csum_status(eop_desc, ptype);
> > > > > > > + if (status & ICE_RX_CSUM_NONE)
> > > > > > > + return -ENODATA;
> > > > > > > +
> > > > > > > + *csum_status = ice_rx_csum_lvl(status) + 1;
> >
> > I'll duplicate an improved version of this line from another thread in case it
> > could help with the comprehension during review:
> >
> > *csum_status = XDP_CHECKSUM_VALID_LVL0 + ice_rx_csum_lvl(status);
> >
> > > > > > > + return 0;
> > > > > > > +}
> > > > > >
> > > > > > and xdp_csum_info from previous patch left uninitialized?
> > > > > > What was the point adding it then?
> > > > >
> > > > > I suppose this driver only returns CHECKSUM_NONE or
> > > > > CHECKSUM_UNNECESSARY? Also based on a grep of the driver dir.
> > > > >
> > > >
> > > > Yes, correct, current ice HW cannot produce complete checksum,
> > > > so only CHECKSUM_UNNECESSARY for known protocols, CHECKSUM_NONE otherwise,
> > > > nothing to initialize csum_info with in either case.
> > > >
> > > > xdp_csum_info is initialized in veth implementation though, but only
> > > > csum_start/offset, so complete XDP checksum has no users in this patchset.
> > > > Is this a problem?
> > > >
> > > > In previous version I had CHECKSUM_UNNECESSARY-only kfunc, but I think everyone
> > > > has agreed, csum hint kfunc should give more comprehensive output.
> > >
> > > csum kfunc supposed to be generic.
> > > If for ICE it fills in one argument and for veth another then the whole
> > > idea of generic api is not working.
> >
> > Both ice and veth fill in the csum_status, the need to fill in the csum_info is
> > determined by the status. I don not see a problem with that.
> >
> > Maybe you have an issue with putting a valid checksum number into a status
> > instead of info? Please clarify.
>
> +1, that seems to match skb interface
>
> Regarding 'generic api not working' in general: I think we've discussed
> that with this 'flexible' kfunc format we can allow non-generic kfuncs for
> specific devices if we think that it makes sense to
> differentiate/experiment/etc. Do you think it makes sense to go
> non-generic route here?
I think we should expose the standard CHECKSUM_* behavior.
The current encoding captures all four types. Which type is returned
is not necessarily defined only by the device. Some devices can
return CHECKSUM_NONE for some packets, CHECKSUM_PARTIAL for others
and CHECKSUM_UNNECESSARY for a third set (e.g., mlx4). Specializing
the return type to the device would not simplify the struct.
