Re: [PATCH RESEND bpf-next 03/15] ice: make RX checksum checking code more reusable

[Larysa Zaremba]

On Mon, May 22, 2023 at 05:51:37PM +0200, Alexander Lobakin wrote:
> From: Larysa Zaremba <larysa.zaremba@xxxxxxxxx>
> Date: Fri, 12 May 2023 17:25:55 +0200
> 
> > Previously, we only needed RX checksum flags in skb path,
> > hence all related code was written with skb in mind.
> > But with the addition of XDP hints via kfuncs to the ice driver,
> > the same logic will be needed in .xmo_() callbacks.
> > 
> > Put generic process of determining checksum status into
> > a separate function.
> > 
> > Signed-off-by: Larysa Zaremba <larysa.zaremba@xxxxxxxxx>
> > ---
> >  drivers/net/ethernet/intel/ice/ice_txrx_lib.c | 71 ++++++++++++-------
> >  1 file changed, 46 insertions(+), 25 deletions(-)
> > 
> > diff --git a/drivers/net/ethernet/intel/ice/ice_txrx_lib.c b/drivers/net/ethernet/intel/ice/ice_txrx_lib.c
> > index 1aab79dc8915..6a4fd3f3fc0a 100644
> > --- a/drivers/net/ethernet/intel/ice/ice_txrx_lib.c
> > +++ b/drivers/net/ethernet/intel/ice/ice_txrx_lib.c
> > @@ -104,17 +104,17 @@ ice_rx_hash_to_skb(struct ice_rx_ring *rx_ring,
> >  }
> >  
> >  /**
> > - * ice_rx_csum - Indicate in skb if checksum is good
> > - * @ring: the ring we care about
> > - * @skb: skb currently being received and modified
> > + * ice_rx_csum_checked - Indicates, whether hardware has checked the checksum
> 
> %CHECKSUM_UNNECESSARY means that the csum is correct / frame is not
> damaged. So "checked" is not enough I'd say, it's "verified" at least.
> OTOH that's too long already, I'd go with classic "csum_ok" :D

'csum_ok' sounds good :) 'csum_valid' if want to be fancy

> 
> >   * @rx_desc: the receive descriptor
> >   * @ptype: the packet type decoded by hardware
> > + * @csum_lvl_dst: address to put checksum level into
> > + * @ring: ring for error stats, can be NULL
> >   *
> > - * skb->protocol must be set before this function is called
> > + * Returns true, if hardware has checked the checksum.
> >   */
> > -static void
> > -ice_rx_csum(struct ice_rx_ring *ring, struct sk_buff *skb,
> > -	    union ice_32b_rx_flex_desc *rx_desc, u16 ptype)
> > +static bool
> > +ice_rx_csum_checked(union ice_32b_rx_flex_desc *rx_desc, u16 ptype,
> 
> (also const, but I guess you'll do that either way after the previous
>  mails)

OK

> 
> > +		    u8 *csum_lvl_dst, struct ice_rx_ring *ring)
> >  {
> >  	struct ice_rx_ptype_decoded decoded;
> >  	u16 rx_status0, rx_status1;
> 
> [...]
> 
> > +/**
> > + * ice_rx_csum_into_skb - Indicate in skb if checksum is good
> > + * @ring: the ring we care about
> > + * @skb: skb currently being received and modified
> > + * @rx_desc: the receive descriptor
> > + * @ptype: the packet type decoded by hardware
> > + */
> > +static void
> > +ice_rx_csum_into_skb(struct ice_rx_ring *ring, struct sk_buff *skb,
> > +		     union ice_32b_rx_flex_desc *rx_desc, u16 ptype)
> > +{
> > +	u8 csum_level = 0;
> 
> I'm not a fan of variables shorter than u32 on the stack. And since it
> gets passed by a reference, I'm not sure the compiler will inline it =\
> 
> > +
> > +	/* Start with CHECKSUM_NONE and by default csum_level = 0 */
> > +	skb->ip_summed = CHECKSUM_NONE;
> > +	skb_checksum_none_assert(skb);
> 
> Can we also remove this? Neither of these makes sense. ::ip_summed is
> always zeroed after the memset() in __build_skb_around() (somewhere
> there), while the assertion checks for `skb->ip_summed ==
> CHECKSUM_NONE`, i.e. it's *always* true here (set and check :D). It's
> some ancient pathetic rituals copied over and over again from e100
> centuries or so...

Will fix.

> 
> ...and BTW the comment is misleading, because the code doesn't zero
> ::csum_level as they claim :D
> 
> > +
> > +	/* check if Rx checksum is enabled */
> > +	if (!(ring->netdev->features & NETIF_F_RXCSUM))
> > +		return;
> > +
> > +	if (!ice_rx_csum_checked(rx_desc, ptype, &csum_level, ring))
> > +		return;
> > +
> > +	skb->ip_summed = CHECKSUM_UNNECESSARY;
> > +	skb->csum_level = csum_level;
> 
> Since csum_level is useless when ip_summed is set to NONE, what do you
> think about making the function return -1, 0, or 1 without writing
> anything by reference?
> 
> 	int csum_level;
> 
> 	csum_level = ice_rx_csum_ok(rx_desc, ptype, ring);
> 	if (csum_level < 0)
> 		return;
> 
> 	skb->ip_summed = CHECKSUM_UNNECESSARY;
> 	skb->csum_level = csum_level;
> 
> I'm not saying it's better (might be a bit at codegen), just proposing.

I think it's worth a try.

> 
> >  }
> >  
> >  /**
> > @@ -232,7 +253,7 @@ ice_process_skb_fields(struct ice_rx_ring *rx_ring,
> >  	/* modifies the skb - consumes the enet header */
> >  	skb->protocol = eth_type_trans(skb, rx_ring->netdev);
> >  
> > -	ice_rx_csum(rx_ring, skb, rx_desc, ptype);
> > +	ice_rx_csum_into_skb(rx_ring, skb, rx_desc, ptype);
> >  
> >  	if (rx_ring->ptp_rx)
> >  		ice_ptp_rx_hwts_to_skb(rx_ring, rx_desc, skb);
> 
> Thanks,
> Olek