On Fri, 5 Jun 2026 05:16:54 +0200 Marek Vasut wrote:

On 6/5/26 4:35 AM, Jakub Kicinski wrote:

quoted

On Thu,  4 Jun 2026 13:41:19 +0200 Paolo Abeni wrote:  

quoted

[Severity: High]
Is it intentional to enforce an 8-bit type constraint for bandwidth and DSP
equalizer tunables?

This heavily couples the global API to the 8-bit register layout of the
KSZ87xx hardware without defining generic units like Hz for bandwidth.  

AFAIK this is the one comment from AI that survives scrutiny.
Seems unlikely we'd need more bits, but then again - why not.
u32 should be our default for integer values, u8 is more for
booleans.  

Why waste memory space if we are already sure the space would not be used.

Waste what memory? We are talking about a syscall argument effectively.

My point is that this is adding the arguments as u8 in
net/ethtool/ioctl.c, so as part of the global uAPI.

The driver can still store these as u8, since any value outside of that
range is invalid for ksz87xx, and the input is properly validated. 
But we shouldn't constrain the value "in transit", at the uAPI level.
Another driver may need a larger range (however unlikely that is).