[RFC PATCH net-next] net: Modify CSUM capability check for USO

[RFC PATCH net-next] net: Modify CSUM capability check for USO

[Radharapu Rakesh]

Some devices supporting USO have only ipv4 checksum enabled,
but USO gets disabled because both ipv4 and ipv6 functionality is checked.

IPv4 or IPv6 or both checksums, can be utilized with UDP segmentation
offload. Separate checks for both IPv4 and IPv6 checksums.

Fixes: 2b2bc3bab158 ("net: Make USO depend on CSUM offload")
Signed-off-by: Radharapu Rakesh <rakesh.radharapu@amd.com>
---
 net/core/dev.c | 8 +++++---
 1 file changed, 5 insertions(+), 3 deletions(-)

diff --git a/net/core/dev.c b/net/core/dev.c
index 1cb134ff7327..a22f8f6e2ed1 100644
--- a/net/core/dev.c
+++ b/net/core/dev.c
@@ -10465,11 +10465,13 @@ static void netdev_sync_lower_features(struct net_device *upper,
 
 static bool netdev_has_ip_or_hw_csum(netdev_features_t features)
 {
-	netdev_features_t ip_csum_mask = NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM;
-	bool ip_csum = (features & ip_csum_mask) == ip_csum_mask;
+	netdev_features_t ipv4_csum_mask = NETIF_F_IP_CSUM;
+	netdev_features_t ipv6_csum_mask = NETIF_F_IPV6_CSUM;
+	bool ipv4_csum = (features & ipv4_csum_mask) == ipv4_csum_mask;
+	bool ipv6_csum = (features & ipv6_csum_mask) == ipv6_csum_mask;
 	bool hw_csum = features & NETIF_F_HW_CSUM;
 
-	return ip_csum || hw_csum;
+	return ipv4_csum || ipv6_csum || hw_csum;
 }
 
 static netdev_features_t netdev_fix_features(struct net_device *dev,
-- 
2.34.1

Re: [RFC PATCH net-next] net: Modify CSUM capability check for USO

[Paolo Abeni]

On 3/12/25 12:54 PM, Radharapu Rakesh wrote:
>  net/core/dev.c | 8 +++++---
>  1 file changed, 5 insertions(+), 3 deletions(-)
> 
> diff --git a/net/core/dev.c b/net/core/dev.c
> index 1cb134ff7327..a22f8f6e2ed1 100644
> --- a/net/core/dev.c
> +++ b/net/core/dev.c
> @@ -10465,11 +10465,13 @@ static void netdev_sync_lower_features(struct net_device *upper,
>  
>  static bool netdev_has_ip_or_hw_csum(netdev_features_t features)
>  {
> -	netdev_features_t ip_csum_mask = NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM;
> -	bool ip_csum = (features & ip_csum_mask) == ip_csum_mask;
> +	netdev_features_t ipv4_csum_mask = NETIF_F_IP_CSUM;
> +	netdev_features_t ipv6_csum_mask = NETIF_F_IPV6_CSUM;
> +	bool ipv4_csum = (features & ipv4_csum_mask) == ipv4_csum_mask;
> +	bool ipv6_csum = (features & ipv6_csum_mask) == ipv6_csum_mask;
>  	bool hw_csum = features & NETIF_F_HW_CSUM;
>  
> -	return ip_csum || hw_csum;
> +	return ipv4_csum || ipv6_csum || hw_csum;
>  }

The above will additionally affect TLS offload, and will likely break
i.e. USO over IPv6 traffic landing on devices supporting only USO over
IPv4, unless such devices additionally implement a suitable
ndo_features_check().

Such situation will be quite bug prone, I'm unsure we want this kind of
change - even without looking at the TLS side of it.

/P

RE: [RFC PATCH net-next] net: Modify CSUM capability check for USO

[Radharapu, Rakesh]

[AMD Official Use Only - AMD Internal Distribution Only]

> -----Original Message-----
> From: Paolo Abeni <pabeni@redhat.com>
> Sent: Wednesday, March 12, 2025 9:44 PM
> To: Radharapu, Rakesh <Rakesh.Radharapu@amd.com>; git (AMD-Xilinx)
> <git@amd.com>; davem@davemloft.net; edumazet@google.com;
> kuba@kernel.org; horms@kernel.org; kuniyu@amazon.com;
> bigeasy@linutronix.de
> Cc: netdev@vger.kernel.org; linux-kernel@vger.kernel.org; Katakam, Harini
> <harini.katakam@amd.com>; Pandey, Radhey Shyam
> <radhey.shyam.pandey@amd.com>; Simek, Michal
> <michal.simek@amd.com>
> Subject: Re: [RFC PATCH net-next] net: Modify CSUM capability check for USO
>
> Caution: This message originated from an External Source. Use proper
> caution when opening attachments, clicking links, or responding.
>
>
> On 3/12/25 12:54 PM, Radharapu Rakesh wrote:
> >  net/core/dev.c | 8 +++++---
> >  1 file changed, 5 insertions(+), 3 deletions(-)
> >
> > diff --git a/net/core/dev.c b/net/core/dev.c index
> > 1cb134ff7327..a22f8f6e2ed1 100644
> > --- a/net/core/dev.c
> > +++ b/net/core/dev.c
> > @@ -10465,11 +10465,13 @@ static void
> > netdev_sync_lower_features(struct net_device *upper,
> >
> >  static bool netdev_has_ip_or_hw_csum(netdev_features_t features)  {
> > -     netdev_features_t ip_csum_mask = NETIF_F_IP_CSUM |
> NETIF_F_IPV6_CSUM;
> > -     bool ip_csum = (features & ip_csum_mask) == ip_csum_mask;
> > +     netdev_features_t ipv4_csum_mask = NETIF_F_IP_CSUM;
> > +     netdev_features_t ipv6_csum_mask = NETIF_F_IPV6_CSUM;
> > +     bool ipv4_csum = (features & ipv4_csum_mask) == ipv4_csum_mask;
> > +     bool ipv6_csum = (features & ipv6_csum_mask) == ipv6_csum_mask;
> >       bool hw_csum = features & NETIF_F_HW_CSUM;
> >
> > -     return ip_csum || hw_csum;
> > +     return ipv4_csum || ipv6_csum || hw_csum;
> >  }
>
> The above will additionally affect TLS offload, and will likely break i.e. USO
> over IPv6 traffic landing on devices supporting only USO over IPv4, unless
> such devices additionally implement a suitable ndo_features_check().
>
> Such situation will be quite bug prone, I'm unsure we want this kind of change
> - even without looking at the TLS side of it.
>
> /P
Thanks for your review. I understand that this will lead to an issue.
We have a device that supports only IPv4 CSUM and are unable to enable the
USO feature because of this check. Can you please let me know if splitting
GSO feature for IPv4 and IPv6 would be helpful? That way corresponding
CSUM offloads can be checked. But this would be a major change.
Will appreciate any other suggestions.

Regards,
Rakesh

RE: [RFC PATCH net-next] net: Modify CSUM capability check for USO

[Willem de Bruijn]

Radharapu, Rakesh wrote:
> [AMD Official Use Only - AMD Internal Distribution Only]
> 
> > -----Original Message-----
> > From: Paolo Abeni <pabeni@redhat.com>
> > Sent: Wednesday, March 12, 2025 9:44 PM
> > To: Radharapu, Rakesh <Rakesh.Radharapu@amd.com>; git (AMD-Xilinx)
> > <git@amd.com>; davem@davemloft.net; edumazet@google.com;
> > kuba@kernel.org; horms@kernel.org; kuniyu@amazon.com;
> > bigeasy@linutronix.de
> > Cc: netdev@vger.kernel.org; linux-kernel@vger.kernel.org; Katakam, Harini
> > <harini.katakam@amd.com>; Pandey, Radhey Shyam
> > <radhey.shyam.pandey@amd.com>; Simek, Michal
> > <michal.simek@amd.com>
> > Subject: Re: [RFC PATCH net-next] net: Modify CSUM capability check for USO
> >
> > Caution: This message originated from an External Source. Use proper
> > caution when opening attachments, clicking links, or responding.
> >
> >
> > On 3/12/25 12:54 PM, Radharapu Rakesh wrote:
> > >  net/core/dev.c | 8 +++++---
> > >  1 file changed, 5 insertions(+), 3 deletions(-)
> > >
> > > diff --git a/net/core/dev.c b/net/core/dev.c index
> > > 1cb134ff7327..a22f8f6e2ed1 100644
> > > --- a/net/core/dev.c
> > > +++ b/net/core/dev.c
> > > @@ -10465,11 +10465,13 @@ static void
> > > netdev_sync_lower_features(struct net_device *upper,
> > >
> > >  static bool netdev_has_ip_or_hw_csum(netdev_features_t features)  {
> > > -     netdev_features_t ip_csum_mask = NETIF_F_IP_CSUM |
> > NETIF_F_IPV6_CSUM;
> > > -     bool ip_csum = (features & ip_csum_mask) == ip_csum_mask;
> > > +     netdev_features_t ipv4_csum_mask = NETIF_F_IP_CSUM;
> > > +     netdev_features_t ipv6_csum_mask = NETIF_F_IPV6_CSUM;
> > > +     bool ipv4_csum = (features & ipv4_csum_mask) == ipv4_csum_mask;
> > > +     bool ipv6_csum = (features & ipv6_csum_mask) == ipv6_csum_mask;
> > >       bool hw_csum = features & NETIF_F_HW_CSUM;
> > >
> > > -     return ip_csum || hw_csum;
> > > +     return ipv4_csum || ipv6_csum || hw_csum;
> > >  }
> >
> > The above will additionally affect TLS offload, and will likely break i.e. USO
> > over IPv6 traffic landing on devices supporting only USO over IPv4, unless
> > such devices additionally implement a suitable ndo_features_check().
> >
> > Such situation will be quite bug prone, I'm unsure we want this kind of change
> > - even without looking at the TLS side of it.
> >
> > /P
> Thanks for your review. I understand that this will lead to an issue.
> We have a device that supports only IPv4 CSUM and are unable to enable the
> USO feature because of this check. Can you please let me know if splitting
> GSO feature for IPv4 and IPv6 would be helpful? That way corresponding
> CSUM offloads can be checked. But this would be a major change.
> Will appreciate any other suggestions.

Splitting NETIF_F_GSO_UDP_L4 would incur significant churn.

Since this is a limitation of a specific device, can you instead
advertise the feature, but for IPv6 packets drop the flag in
ndo_features_check?

RE: [RFC PATCH net-next] net: Modify CSUM capability check for USO

[Radharapu, Rakesh]

[AMD Official Use Only - AMD Internal Distribution Only]

> -----Original Message-----
> From: Willem de Bruijn <willemdebruijn.kernel@gmail.com>
> Sent: Sunday, March 16, 2025 7:35 AM
> To: Radharapu, Rakesh <Rakesh.Radharapu@amd.com>; Paolo Abeni
> <pabeni@redhat.com>; git (AMD-Xilinx) <git@amd.com>;
> davem@davemloft.net; edumazet@google.com; kuba@kernel.org;
> horms@kernel.org; kuniyu@amazon.com; bigeasy@linutronix.de
> Cc: netdev@vger.kernel.org; linux-kernel@vger.kernel.org; Katakam, Harini
> <harini.katakam@amd.com>; Pandey, Radhey Shyam
> <radhey.shyam.pandey@amd.com>; Simek, Michal
> <michal.simek@amd.com>
> Subject: RE: [RFC PATCH net-next] net: Modify CSUM capability check for USO
>
> Caution: This message originated from an External Source. Use proper
> caution when opening attachments, clicking links, or responding.
>
>
> Radharapu, Rakesh wrote:
> > [AMD Official Use Only - AMD Internal Distribution Only]
> >
> > > -----Original Message-----
> > > From: Paolo Abeni <pabeni@redhat.com>
> > > Sent: Wednesday, March 12, 2025 9:44 PM
> > > To: Radharapu, Rakesh <Rakesh.Radharapu@amd.com>; git (AMD-Xilinx)
> > > <git@amd.com>; davem@davemloft.net; edumazet@google.com;
> > > kuba@kernel.org; horms@kernel.org; kuniyu@amazon.com;
> > > bigeasy@linutronix.de
> > > Cc: netdev@vger.kernel.org; linux-kernel@vger.kernel.org; Katakam,
> > > Harini <harini.katakam@amd.com>; Pandey, Radhey Shyam
> > > <radhey.shyam.pandey@amd.com>; Simek, Michal
> <michal.simek@amd.com>
> > > Subject: Re: [RFC PATCH net-next] net: Modify CSUM capability check
> > > for USO
> > >
> > > Caution: This message originated from an External Source. Use proper
> > > caution when opening attachments, clicking links, or responding.
> > >
> > >
> > > On 3/12/25 12:54 PM, Radharapu Rakesh wrote:
> > > >  net/core/dev.c | 8 +++++---
> > > >  1 file changed, 5 insertions(+), 3 deletions(-)
> > > >
> > > > diff --git a/net/core/dev.c b/net/core/dev.c index
> > > > 1cb134ff7327..a22f8f6e2ed1 100644
> > > > --- a/net/core/dev.c
> > > > +++ b/net/core/dev.c
> > > > @@ -10465,11 +10465,13 @@ static void
> > > > netdev_sync_lower_features(struct net_device *upper,
> > > >
> > > >  static bool netdev_has_ip_or_hw_csum(netdev_features_t features)  {
> > > > -     netdev_features_t ip_csum_mask = NETIF_F_IP_CSUM |
> > > NETIF_F_IPV6_CSUM;
> > > > -     bool ip_csum = (features & ip_csum_mask) == ip_csum_mask;
> > > > +     netdev_features_t ipv4_csum_mask = NETIF_F_IP_CSUM;
> > > > +     netdev_features_t ipv6_csum_mask = NETIF_F_IPV6_CSUM;
> > > > +     bool ipv4_csum = (features & ipv4_csum_mask) ==
> ipv4_csum_mask;
> > > > +     bool ipv6_csum = (features & ipv6_csum_mask) ==
> > > > + ipv6_csum_mask;
> > > >       bool hw_csum = features & NETIF_F_HW_CSUM;
> > > >
> > > > -     return ip_csum || hw_csum;
> > > > +     return ipv4_csum || ipv6_csum || hw_csum;
> > > >  }
> > >
> > > The above will additionally affect TLS offload, and will likely
> > > break i.e. USO over IPv6 traffic landing on devices supporting only
> > > USO over IPv4, unless such devices additionally implement a suitable
> ndo_features_check().
> > >
> > > Such situation will be quite bug prone, I'm unsure we want this kind
> > > of change
> > > - even without looking at the TLS side of it.
> > >
> > > /P
> > Thanks for your review. I understand that this will lead to an issue.
> > We have a device that supports only IPv4 CSUM and are unable to enable
> > the USO feature because of this check. Can you please let me know if
> > splitting GSO feature for IPv4 and IPv6 would be helpful? That way
> > corresponding CSUM offloads can be checked. But this would be a major
> change.
> > Will appreciate any other suggestions.
>
> Splitting NETIF_F_GSO_UDP_L4 would incur significant churn.
>
> Since this is a limitation of a specific device, can you instead advertise the
> feature, but for IPv6 packets drop the flag in ndo_features_check?

Yes, Thanks for your reply. I'm making the recommended modification
to the driver and will also monitor performance numbers.
If everything goes well this RFC patch can be dropped.

Regards,
Rakesh