Re: arm64 mrs and system registers

> On 8 Apr 2024, at 22:31, Warner Losh <imp@bsdimp.com> wrote:
> 
> 
> 
> On Mon, Apr 8, 2024 at 9:03 PM Paul Floyd <pjfloyd@wanadoo.fr <mailto:pjfloyd@wanadoo.fr>> wrote:
>> Hi
>> 
>> I've been looking at this bugzilla item
>> 
>> https://bugs.kde.org/show_bug.cgi?id=392146
>> 
>> Is there any difference between Linux and FreeBSD when it comes to what 
>> registers and fields are exposed by the kernel (see comment 17 in the 
>> link above).
> 
> I don't think so. We've not seen issues with other drivers on aarch64 except
> when they were written on x86 and didn't have the synchronization needed
> for the weaker memory models in aarch64 systems.
>  
>> I did have a poke around the kernel code but it's a bit hard to tell 
>> exactly which of the access macros are being used, without exhaustively 
>> grepping for them one by one.
> 
> Yea, I think that there's missing atomics on the state transitions and/or
> some missing locking that "magically" provides barriers that make it work
> on x86.
> 

Hi

There aren’t any memory issues.

The problem is that the opcodes aren’t fully covered. There are 3 aspects to that
1. What the kernel exposes
2. What Valgrind implements (usually a subset of point 1 but it should claim things that the kernel doesn’t support).
3. Actually handling the opcode.

If Linux and FreeBSD expose the same things then I can go ahead with looking at a common solution.


A+
Paul