Re: [PATCH] arm64/sve: Document that __SVE_VQ_MAX is much larger than needed

From: Dave Martin
Date: Wed Feb 07 2024 - 07:01:59 EST


On Tue, Feb 06, 2024 at 04:27:01PM +0000, Mark Brown wrote:
> __SVE_VQ_MAX is defined without comment as 512 but the actual
> architectural maximum is 16, a substantial difference which might not
> be obvious to readers especially given the several different units used
> for specifying vector sizes in various contexts and the fact that it's
> often used via macros. In an effort to minimise surprises for users who
> might assume the value is the architectural maximum and use it to do
> things like size allocations add a comment noting the difference.

Well, the value 512 was semi-deliberately chosen to be surprising.

But the point about units is valid: to the casual reader, "512" does
look suspiciously like a bit count when it really really isn't...

>
> Signed-off-by: Mark Brown <broonie@xxxxxxxxxx>
> ---
> arch/arm64/include/uapi/asm/sve_context.h | 4 ++++
> 1 file changed, 4 insertions(+)
>
> diff --git a/arch/arm64/include/uapi/asm/sve_context.h b/arch/arm64/include/uapi/asm/sve_context.h
> index 754ab751b523..59f283f373a6 100644
> --- a/arch/arm64/include/uapi/asm/sve_context.h
> +++ b/arch/arm64/include/uapi/asm/sve_context.h
> @@ -13,6 +13,10 @@
>
> #define __SVE_VQ_BYTES 16 /* number of bytes per quadword */
>
> +/*
> + * Note that for future proofing __SVE_VQ_MAX is defined much larger
> + * than the actual architecture maximum of 16.
> + */

I think that putting shadow #defines in comments in UAPI headers is a
really bad idea... is this a normative statement about the user API,
or what?

My concern is that if we muddy the waters here different bits of
software will do different things and we will get a mess with no
advantages.

Portability issues may ensue if userspace software feels it can
substitute some other value for this constant, since we can't control
what userspace uses it for.

> #define __SVE_VQ_MIN 1

Would it be sufficient to say something like:

/*
* Yes, this is 512 QUADWORDS.
* Never allocate memory or size structures based on the value of this
* constant.
*/
> #define __SVE_VQ_MAX 512

Though comments might be better placed alongsize SVE_VQ_MAX at al., in
ptrace.h and sigcontext.h rather than here. The leading __ should at
least be a hint that __SVE_VQ_MAX shouldn't be used directly by
anyone...

Cheers
---Dave