Re: drivers/net/ethernet/sfc/selftest.c:48:16: warning: field ip within 'struct efx_loopback_payload::(anonymous at drivers/net/ethernet/sfc/selftest.c:46:2)' is less aligned than 'struct iphdr' and is usually due to 'struct efx_loopback_payload::(anonymous a...

[Yujie Liu]

On Wed, Nov 22, 2023 at 04:15:49PM +0000, Edward Cree wrote:
> On 21/11/2023 21:25, kernel test robot wrote:
> > Hi Edward,
> > 
> > FYI, the error/warning still remains.
> 
> As I've argued previously, this is a false positive / compiler bug,
>  and there is no way to resolve it without making the code strictly
>  worse.
> 
> This:
> >>> drivers/net/ethernet/sfc/selftest.c:48:16: warning: field ip within 'struct efx_loopback_payload::(anonymous at drivers/net/ethernet/sfc/selftest.c:46:2)' is less aligned than 'struct iphdr' and is usually due to 'struct efx_loopback_payload::(anonymous at drivers/net/ethernet/sfc/selftest.c:46:2)' being packed, which can lead to unaligned accesses [-Wunaligned-access]
>  is complaining about alignment within an anonymous struct, which
>  only ever appears embedded within a larger struct in a way which
>  maintains the correct alignment.
> 
> #ifdef RANT
> 
> Indeed, the only way we even *could* create an unaligned access
>  out of this code would be via a declaration like
>     typeof(*(((struct efx_loopback_payload *)0)->packet)) bad;
>  because *the struct is anonymous*.  And if that happened, the
>  bad declaration would be the place to warn, both because it's
>  incredibly ugly and because it's the place that's actually
>  wrong.  The struct definition itself is entirely *fine*.
> The compiler should be able to detect that, and if it's not smart
>  enough to do so then it shouldn't be trying to warn in the first
>  place.  Quoth Linus[1]:
> 
>  "And if the compiler isn't good enough to do it, then the compiler
>   shouldn't be warning about something that it hasn't got a clue about."
> 
> The anonymous struct has to be there so that we can placate the
>  memcpy hardening, and it has to contain a struct iphdr at a
>  4n+2 offset because that's what shape the on-the-wire packet
>  *is*.  To avoid the warning we would need to lose __packed and
>  memcpy all of the members in and out of the buffer individually
>  to explicitly-calculated offsets, which is worse code.
> 
> #endif
> 
> Either fix the compiler to not warn, or fix your automation to
>  ignore this instance of the warning.

Hi Edward,

Thanks a lot for the detailed explanation. We've configured the robot
to ignore the "unaligned access" warning for this specific case.

Best Regards,
Yujie

> -ed
> 
> [1]: https://yarchive.net/comp/linux/gcc.html#13
>