[PATCHv2] ARM64: Add AT_ARM64_MIDR to the aux vector

[Pinski, Andrew]

> On Sep 2, 2015, at 1:19 AM, Mark Rutland <mark.rutland at arm.com> wrote:
> 
>> On Tue, Sep 01, 2015 at 05:51:44PM +0100, pinskia at gmail.com wrote:
>> 
>> 
>> 
>> 
>>> On Sep 2, 2015, at 12:33 AM, Mark Rutland <mark.rutland at arm.com> wrote:
>>> 
>>> Hi,
>>> 
>>>> On Sat, Aug 29, 2015 at 07:46:22PM +0100, Andrew Pinski wrote:
>>>> It is useful to pass down MIDR register down to userland if all of
>>>> the online cores are all the same type.  This adds AT_ARM64_MIDR
>>>> aux vector type and passes down the midr system register.
>>>> 
>>>> This is alternative to MIDR_EL1 part of
>>>> http://lists.infradead.org/pipermail/linux-arm-kernel/2015-July/358995.html.
>>>> It allows for faster access to midr_el1 than going through a trap and
>>>> does not exist if the set of cores are not the same.
>>> 
>>> I'm not sure I follow the rationale. If speed is important the
>>> application can cache the value the first time it reads it with a trap.
>> 
>> It is also about compatibility also. Exposing the register is not
>> backwards compatible but using the aux vector is. 
> 
> So long as we have HWCAP_CPUID describing the availability of register
> access [2], then userspace can test for that before attempting to access
> the MIDR.

So two checks always. Bad choice. 

> 
> Other than that, I don't see a backwards or forwards compatibility
> issue.
> 
>>>> +u32 get_arm64_midr(void)
>>>> +{
>>>> +    int i;
>>>> +    u32 midr = 0;
>>>> +
>>>> +    for_each_online_cpu(i) {
>>>> +        struct cpuinfo_arm64 *cpuinfo = &per_cpu(cpu_data, i);
>>>> +        u32 oldmidr = midr;
>>>> +
>>>> +        midr = cpuinfo->reg_midr;
>>>> +        /*
>>>> +         * If there are cpus which have a different
>>>> +         * midr just return 0.
>>>> +         */
>>>> +        if (oldmidr && oldmidr != midr)
>>>> +            return 0;
>>>> +    }
>>>> +
>>>> +    return midr;
>>>> +}
>>> 
>>> If I have a big.LITTLE system where all the big CPUs are currently
>>> offline, this will leave the MIDR the little CPUs in the auxvec.
>>> However, at any point after this has run, I could hotplug the big CPUs
>>> on and the little CPUs off, leaving this reporting a MIDR that
>>> represents none of the online CPUs.
>>> 
>>> Given big.LITTLE and the potential for physical/dynamic hotplug (where
>>> we won't know all the MIDRs in advance), I don't think that we can
>>> generally expose a common MIDR in this fashion, and I don't think that
>>> we should give the impression that we can.
>> 
>> This is standard issue with hot plug and big.little. Really big.little
>> is a design flaw but I am not going into that here. 
> 
> Regardless of our personal feelings on big.LITTLE, it's something we
> have to deal with.

You did not respond to the caching in userspace issue. You raised that as a speed optimization but then considered my patch as a non starter. 

> 
> Hopefully it's a non-issue anyway; a MIDR provided by this interface can
> really only be used to derive optimisation criteria rather than
> non-architected properties required for correctness.

Like hardware workarounds? Yes that is going to be used for that and is already right now by reading /proc/cpuinfo . Also it would useful to use what ever interface for gcc's -mcpu=native option. 


> 
>>> I think that the only things we can do are expose the MIDR for CPU the
>>> code is currently executing on (as Suzuki's patches do), and/or expose
>>> all the MIDRs for currently online CPUs (as Steve's [1] patch does).
>>> Anything else leaves us trying to provide semantics that we cannot
>>> guarantee.
>> 
>> Except they are not backwards compatible which means nobody in their
>> right mind would use the register to get the midr that way.
> 
> I assume you missed the discussion of HWCAP_CPUID, which prevents the
> compatibility issue I believe you're considering here.

And you suggested to cache midr while not considering big.little. You can't have it both ways. I read those and I still think they were wrong in rejecting this.  Also there are two markets arm is in and things like this is causing one of those markets to suffer. Big.little is not going into servers. 

> 
>> I am sorry but having a newer version of glibc working on a year old
>> kernel is not going to fly. 
> 
> I'm not sure I follow this, unless you meant _not_ working.

Because of the double checks, it will be slower and the trap. And gives a bad interface to userland really.  Aux vector is a much cleaner interface to userland than a trapped instruction. 

Thanks,
Andrew

> 
> Thanks,
> Mark.
> 
> [1] http://lists.infradead.org/pipermail/linux-arm-kernel/2015-July/359127.html
> [2] http://lists.infradead.org/pipermail/linux-arm-kernel/2015-August/363559.html