Re: [PATCH net v2] bonding: Correctly support GSO ESP offload

[Cosmin Ratiu <cratiu@nvidia.com>]

On Fri, 2025-01-24 at 09:54 +0000, Hangbin Liu wrote:
> On Fri, Jan 24, 2025 at 10:57:44AM +0200, Cosmin Ratiu wrote:
> > ---
> >  drivers/net/bonding/bond_main.c | 19 ++++++++++---------
> >  1 file changed, 10 insertions(+), 9 deletions(-)
> > 
> > diff --git a/drivers/net/bonding/bond_main.c
> > b/drivers/net/bonding/bond_main.c
> > index 7b78c2bada81..e45bba240cbc 100644
> > --- a/drivers/net/bonding/bond_main.c
> > +++ b/drivers/net/bonding/bond_main.c
> > @@ -1538,17 +1538,20 @@ static netdev_features_t
> > bond_fix_features(struct net_device *dev,
> >  				 NETIF_F_HIGHDMA | NETIF_F_LRO)
> >  
> >  #define BOND_ENC_FEATURES	(NETIF_F_HW_CSUM | NETIF_F_SG | \
> > -				 NETIF_F_RXCSUM |
> > NETIF_F_GSO_SOFTWARE)
> > +				 NETIF_F_RXCSUM |
> > NETIF_F_GSO_SOFTWARE | \
> > +				 NETIF_F_GSO_PARTIAL)
> >  
> >  #define BOND_MPLS_FEATURES	(NETIF_F_HW_CSUM | NETIF_F_SG | \
> >  				 NETIF_F_GSO_SOFTWARE)
> >  
> > +#define BOND_GSO_PARTIAL_FEATURES (NETIF_F_GSO_ESP)
> > +
> >  
> >  static void bond_compute_features(struct bonding *bond)
> >  {
> > +	netdev_features_t gso_partial_features =
> > BOND_GSO_PARTIAL_FEATURES;
> >  	unsigned int dst_release_flag = IFF_XMIT_DST_RELEASE |
> >  					IFF_XMIT_DST_RELEASE_PERM;
> > -	netdev_features_t gso_partial_features = NETIF_F_GSO_ESP;
> >  	netdev_features_t vlan_features = BOND_VLAN_FEATURES;
> >  	netdev_features_t enc_features  = BOND_ENC_FEATURES;
> >  #ifdef CONFIG_XFRM_OFFLOAD
> > @@ -1582,8 +1585,9 @@ static void bond_compute_features(struct
> > bonding *bond)
> >  							 
> > BOND_XFRM_FEATURES);
> >  #endif /* CONFIG_XFRM_OFFLOAD */
> >  
> > -		if (slave->dev->hw_enc_features &
> > NETIF_F_GSO_PARTIAL)
> > -			gso_partial_features &= slave->dev-
> > >gso_partial_features;
> > +		gso_partial_features =
> > netdev_increment_features(gso_partial_features,
> > +								
> > slave->dev->gso_partial_features,
> > +								
> > BOND_GSO_PARTIAL_FEATURES);
> >  
> >  		mpls_features =
> > netdev_increment_features(mpls_features,
> >  							  slave-
> > >dev->mpls_features,
> > @@ -1598,12 +1602,8 @@ static void bond_compute_features(struct
> > bonding *bond)
> >  	}
> >  	bond_dev->hard_header_len = max_hard_header_len;
> >  
> > -	if (gso_partial_features & NETIF_F_GSO_ESP)
> > -		bond_dev->gso_partial_features |= NETIF_F_GSO_ESP;
> > -	else
> > -		bond_dev->gso_partial_features &=
> > ~NETIF_F_GSO_ESP;
> > -
> >  done:
> > +	bond_dev->gso_partial_features = gso_partial_features;
> >  	bond_dev->vlan_features = vlan_features;
> >  	bond_dev->hw_enc_features = enc_features |
> > NETIF_F_GSO_ENCAP_ALL |
> >  				    NETIF_F_HW_VLAN_CTAG_TX |
> 
>         if (!bond_has_slaves(bond))
>                 goto done;
> 
> If there is no slaves, should we add the gso_partial_features?

The other partial feature sets are added after 'done:', why not do the
same for gso_partial_features for consistency? 'gso_partial_features'
is otherwise not set anywhere else and relies on it being set to zero
when allocated in alloc_netdev_mqs. I think it's better for it to be
explicitly initialized in all cases here, like the other feature sets.

Cosmin.