[PATCH v2 03/10] riscv: Implement cmpxchg8/16() using Zabha

[Alexandre Ghiti]

Hi Andrea,

On Wed, Jul 10, 2024 at 1:51 AM Andrea Parri <parri.andrea at gmail.com> wrote:
>
> > > I admit that I found this all quite difficult to read; IIUC, this is
> > > missing an IS_ENABLED(CONFIG_RISCV_ISA_ZACAS) check.
> >
> > I'm not sure we need the zacas check here, since we could use a toolchain
> > that supports zabha but not zacas, run this on a zabha/zacas platform and it
> > would work.
>
> One specific set-up I was concerned about is as follows:
>
>   1) hardware implements both zabha and zacas
>   2) toolchain supports both zabha and zacas
>   3) CONFIG_RISCV_ISA_ZABHA=y and CONFIG_RISCV_ISA_ZACAS=n
>
> Since CONFIG_RISCV_ISA_ZABHA=y, the first asm goto will get executed
> and, since the hardware implements zacas, that will result in a nop.
> Then the second asm goto will get executed and, since the hardware
> implements zabha, it will result in the j zabha.  In conclusion, the
> amocas instruction following the zabha: label will get executed, thus
> violating (the semantics of) CONFIG_RISCV_ISA_ZACAS=n.  IIUC, the diff
> I've posted previously in this thread shared a similar limitation/bug.

So you mean that when disabling Zacas, we should actually disable
*all* the CAS instructions, even the Zabha ones. It makes sense and
allows for a single way to disable the CAS instructions but keeping
the other atomic operations.

I'll fix that and add a comment.

Thanks,

Alex

>
>   Andrea