[RFC] net: selftests: Adding TX checksum offload validation

[RFC] net: selftests: Adding TX checksum offload validation

[Oleksij Rempel]

Hello everyone,

While working with the smsc95xx driver, I identified a need for better
validation of the driver and hardware TX checksum offloading capabilities. I
believe a generic test suite for this would benefit other drivers as well.

The generic selftest framework in net/core/selftests.c seems like the ideal
location. It already contains a test for the RX checksum path, so adding
validation for the TX path feels like a natural extension.

Here is the list of test cases I propose to add:
- TX csum offload, IPv4, TCP, Standard MTU Packet
- TX csum offload, IPv4, UDP, Standard MTU Packet
- TX csum offload, IPv4, ICMP, Standard Payload
- TX csum offload, IPv4, TCP, Minimal Size Packet (1-byte payload)
- TX csum offload, IPv4, UDP, Minimal Size Packet (1-byte payload)
- TX csum offload, IPv4, UDP, Zero-Checksum Payload (Verify checksum becomes
                              0xFFFF)
- TX csum offload, IPv4, TCP, With Single VLAN Tag
- TX csum offload, IPv4, TCP, With Double VLAN Tag (Q-in-Q)
- TX csum offload, IPv6, TCP, Standard MTU Packet
- TX csum offload, IPv6, UDP, Standard MTU Packet

The implementation for these tests would involve preparing an skb with the
corresponding L3/L4 headers, flagging it with CHECKSUM_PARTIAL, and sending it
through the PHY loopback. The test would pass if the received frame has a
valid checksum.

As a related question on driver implementation:
The documentation suggests that the older flags NETIF_F_IP_CSUM and
NETIF_F_IPV6_CSUM are being superseded by the more generic NETIF_F_HW_CSUM.
When the network stack sends a packet with skb->ip_summed = CHECKSUM_PARTIAL,
the driver is responsible for ensuring the final checksum is correct, either by
offloading the calculation to the device or by falling back to a software
function like skb_checksum_help().

Is this understanding correct, and is relying on CHECKSUM_PARTIAL as the
primary mechanism for requesting TX offload the recommended practice for modern
network drivers?

Thanks,
Oleksij
-- 
Pengutronix e.K.                           |                             |
Steuerwalder Str. 21                       | http://www.pengutronix.de/  |
31137 Hildesheim, Germany                  | Phone: +49-5121-206917-0    |
Amtsgericht Hildesheim, HRA 2686           | Fax:   +49-5121-206917-5555 |

Re: [RFC] net: selftests: Adding TX checksum offload validation

[Jakub Kicinski]

On Tue, 16 Sep 2025 11:12:02 +0200 Oleksij Rempel wrote:
> While working with the smsc95xx driver, I identified a need for better
> validation of the driver and hardware TX checksum offloading capabilities. I
> believe a generic test suite for this would benefit other drivers as well.
> 
> The generic selftest framework in net/core/selftests.c seems like the ideal
> location. It already contains a test for the RX checksum path, so adding
> validation for the TX path feels like a natural extension.
> 
> Here is the list of test cases I propose to add:
> - TX csum offload, IPv4, TCP, Standard MTU Packet
> - TX csum offload, IPv4, UDP, Standard MTU Packet
> - TX csum offload, IPv4, ICMP, Standard Payload
> - TX csum offload, IPv4, TCP, Minimal Size Packet (1-byte payload)
> - TX csum offload, IPv4, UDP, Minimal Size Packet (1-byte payload)
> - TX csum offload, IPv4, UDP, Zero-Checksum Payload (Verify checksum becomes
>                               0xFFFF)
> - TX csum offload, IPv4, TCP, With Single VLAN Tag
> - TX csum offload, IPv4, TCP, With Double VLAN Tag (Q-in-Q)
> - TX csum offload, IPv6, TCP, Standard MTU Packet
> - TX csum offload, IPv6, UDP, Standard MTU Packet

The in-kernel tests are best for things which are hard to trigger from
user space. Have you seen tools/testing/selftests/drivers/net/hw/csum.py
and tools/testing/selftests/net/lib/csum.c ?