Re: [Qemu-devel] [PATCH v4] linux-user: fix to handle variably sized SIOCGSTAMP with new kernels

[Laurent Vivier]

Le 11/07/2019 à 23:05, Arnd Bergmann a écrit :
> On Thu, Jul 11, 2019 at 7:32 PM Laurent Vivier <address@hidden> wrote:
> 
>>
>> Notes:
>>     v4: [lv] timeval64 and timespec64 are { long long , long }
> 
>>
>> +STRUCT(timeval64, TYPE_LONGLONG, TYPE_LONG)
>> +
>> +STRUCT(timespec64, TYPE_LONGLONG, TYPE_LONG)
>> +
> 
> This still doesn't look right, see my earlier comment about padding
> on big-endian architectures.
> 
> Note that the in-kernel 'timespec64' is different from the uapi
> '__kernel_timespec' exported by the kernel. I also still think you may
> need to convert between SIOCGSTAMP_NEW and SIOCGSTAMP_OLD,
> e.g. when emulating a 32-bit riscv process (which only use
> SIOCGSTAMP_NEW) on a kernel that only understands
> SIOCGSTAMP_OLD.

I agree.
I'm preparing a patch always using SIOCGSTAMP and SIOCGSTAMPNS on the
host (converting the structure when needed).

I've added the SH4 variant.
I've added the sparc64 variant too: does it means sparc64 use the same
structure internally for the OLD and NEW version?
What about sparc 32bit?

For big-endian, I didn't find in the kernel where the difference is
managed: a byte swapping of the 64bit value is not enough?

Thanks,
Laurent