[PATCH 0/2] arm64/sve: Fix mutating register endianness on big-endian

Alan Hayward Alan.Hayward@arm.com
Wed Jun 12 10:59:00 GMT 2019



> On 12 Jun 2019, at 11:40, Alex Bennée <alex.bennee@linaro.org> wrote:
> 
> 
> Alan Hayward <Alan.Hayward@arm.com> writes:
> 
>>> On 7 Jun 2019, at 16:48, Dave Martin <Dave.Martin@arm.com> wrote:
>>> 
>>> On Fri, Jun 07, 2019 at 10:38:58AM +0100, Will Deacon wrote:
>>>> On Thu, Jun 06, 2019 at 05:44:53PM +0100, Dave Martin wrote:
>>>>> By inspection while debugging something else, I noticed that the byte
>>>>> order of FPSIMD V-register stores and SVE Z-register stores is not the
>>>>> same when running on big-endian.
>>>>> 
>>>>> This is not properly taken into account when moving between the FPSIMD
>>>>> and SVE register views inside the kernel, resulting in the bytes of a
>>>>> V-register getting spontaneously reversed in some situations, from
>>>>> userspace's point of view.  The signal frame and ptrace interface are
>>>>> also affected.  The KVM ABI forbids mixing the two views and so should
>>>>> not be affected.
>>>>> 
>>>>> See patch 2 for details.
>>>>> 
>>>>> Patch 1 does some trivial preparatory refactoring.
>>>> 
>>>> Sorry to be a pain, but would you be able to flip this series round so that
>>>> the fix doesn't depend on the refactoring, please? That way we can put it
>>>> into stable without the dependency.
>>>> 
>>>>> gdb may or may not be affected by this, depending on how it uses the
>>>>> NT_PRFPREG and NT_ARM_SVE regsets.  I'll leave it to the developers to
>>>>> assess that.
>>>> 
>>>> Wouldn't this be easy enough to test?
>>> 
>>> So, gdb works OK on big-endian but weird stuff happening on both with
>>> and without the fix.
>>> 
>>> There are places in the gdb code itself where it is likely missing
>>> endianness conversions, but I need to follow up with the gdb folks to
>>> clarify whether my patch is missing something…
>> 
>> (I added the SVE support for GDB).
>> 
>> I’ve tried these changes out myself using GDB.
>> With your changes everything looks good, apart from:
>> * GDB gets it wrong when the ptrace sve structure contains a fpsimd.
>> * I need to do some testing around sigcontexts, but again I think GDB
>>  will need a slight change.
>> I’ll get some patches together for GDB.
> 
> Where is the latest state of SVE support for GDB? I really should check
> the QEMU gdbstub does the correct things for SVE registers but I was
> waiting for upstream gdb support.

SVE is supported in GDB 8.2.

If your program starts changing vector lengths whilst the process is running,
then you’ll run into issues. Fixed for GDB in HEAD, but not for remote stubs.
Let me know if that’s a problem for you - I’ve yet to find real uses cases
for needing it.


Alan.

> 
> --
> Alex Bennée



More information about the Gdb mailing list