RFC: extern illegal instruction trap and trap RDCYCLE

[Ben Dooks]

On 01/10/2024 10:41, Alexandre Ghiti wrote:
> Hi Atish, Ben,
> 
> On 19/09/2024 01:23, Atish Kumar Patra wrote:
>> On Wed, Sep 18, 2024 at 2:57 AM Ben Dooks <ben.dooks at codethink.co.uk> 
>> wrote:
>>> On 18/09/2024 10:55, Atish Patra wrote:
>>>> On 9/17/24 7:01 AM, Ben Dooks wrote:
>>>>> On 17/09/2024 14:46, Palmer Dabbelt wrote:
>>>>>> On Tue, 17 Sep 2024 06:08:50 PDT (-0700), ben.dooks at codethink.co.uk
>>>>>> wrote:
>>>>>>> This is a RFC series to change how the illegal instruction trap
>>>>>>> is handled and then how to trap RDCYCLE and emulate it with RDTIME
>>>>>>> instead.
>>>>>> Only 1/3 made it to lore for me, not sure if it's just stuck 
>>>>>> somewhere.
>>>>>>
>>>>>>> I did this when we found multiple libraries using RDCYCLE and
>>>>>>> upgrading multiple runners to newer kernels caused many problems
>>>>>> OK, so I think we're kind of just stuck with RDCYCLE then -- it was
>>>>>> part of the base ISA when we merged the port, and every time it
>>>>>> disappears we end up breaking userspace.
>>>>> Yes, it was an annoyance, we managed to track down all the users and
>>>>> patch out.
>>>>>
>>>>> I thought this might be a useful idea for more generic type, so might
>>>>> look at updating to have a table of instruction masks to call.
>>>>>
>>>>>> I'm not sure what exactly the right way to do this is: IIRC there's
>>>>>> some perf-related hooks for this, but there's also systems that just
>>>>>> don't implement the RDCYCLE instruction at all and thus we'll need
>>>>>> some sort of emulation for those.
>>>>> I couldn't get the PMU driver to allow it, not sure if there was an
>>>>> issue higher up or some other issue?
>>>>>
>>>> Here is the way to properly allow it via the driver.
>>>>
>>>> https://www.kernel.org/doc/Documentation/admin-guide/sysctl/kernel.rst
>>>>
>>>> Check the perf_user_access
>>> Didn't work for me, not sure why
>>>
>> +Alexandre Ghiti
>>
>> Hmm. That's weird. Maybe some bug crept in recently. Can you provide
>> some more details
>> on the platform, kernel version and procedure followed to change it ?
> 
> 
> So I have just tried perf_user_access on 6.11 and it works fine, meaning 
> setting the "legacy" behaviour (ie 2) allows a userspace application to 
> directly access RDCYCLE.
> 
> Can you provide more details Ben?

I've gone through and checked this again, I think it is very likely
that I got the enable as 1 instead of 2. It works on QEMU, I've not
got access to the runners that we upgraded and we've fixed the issue
by upgrading libabsl to a later version.

Thanks.

-- 
Ben Dooks				http://www.codethink.co.uk/
Senior Engineer				Codethink - Providing Genius

https://www.codethink.co.uk/privacy.html