Re: [PATCH 04/10] arm64: mops: document boot requirements for MOPS

[Catalin Marinas]

On Fri, Mar 24, 2023 at 01:00:43AM +0000, Kristina Martsenko wrote:
> On 17/03/2023 15:07, Catalin Marinas wrote:
> > On Thu, Feb 16, 2023 at 04:00:06PM +0000, Kristina Martsenko wrote:
> >> +  For CPUs with Memory Copy and Memory Set instructions (FEAT_MOPS):
> >> +
> >> +  - If the kernel is entered at EL1 and EL2 is present:
> >> +
> >> +    - HCRX_EL2.MSCEn (bit 11) must be initialised to 0b1.
> >> +
> >> +    - HCRX_EL2.MCE2 (bit 10) must be initialised to 0b0.
> > 
> > Regarding MCE2, does EL1 actually care if EL2 wants to handle all the
> > memcpy/memset exceptions? 
> 
> No, EL1 does not need to handle the exceptions itself, but I don't see any
> current use case for allowing EL2 to handle it.
> 
> If it was allowed, I think booting.txt would need to specify exactly what Linux
> expects EL2 to do if MCE2 is set (eg, that EL2 handles the exception by
> reformatting registers, modifying single step state, etc).

What I meant is that an EL1 kernel shouldn't care if MCE2 is 0 or 1. We
could clarify that if set to 1, it is expected that the hypervisor
handles the memory copy/set exceptions accordingly.

> > There may even be a valid case to do this at
> > EL2 if you run a guest that uses these instructions but has no clue on
> > how to deal with the specific exception like WrongOption.
> 
> Not sure I follow - this series adds the exception handling, so how can a Linux
> guest not know how to handle the exception?

The guest may not always be Linux (e.g. some microkernel) and may not
expect the hardware implementation to change underneath.

> Or do you mean that there may be times when EL1 can't take the exception but
> EL2 may move it to another CPU, and so EL2 would need to handle the exception?

I haven't thought of this but it's actually a good point. Are there any
cases where Linux can't handle a memcpy exception? I guess we need to be
careful with taking such exception in an atomic context (e.g. no
rescheduling on the return path).

> I'm not sure if Linux ever uses mops instructions at times like that.

The compiler may generate a memcpy() call by simply assigning a
structure to another. So we can't control where those instructions are
used.

> Note that this series does not add support for mops in guests yet. 

You mean there's no KVM support. But Linux may be run under a different
hypervisor (e.g. Hyper-V) as a guest.

> I think booting.txt can be updated when that support is added.

In booting.txt, when you say the kernel entered at EL1, it implies that
it may be run as a guest under a random hypervisor.

So maybe we should detail the MCE2 requirement a bit, saying that it can
be either 0 or 1 but, for the latter, the hypervisor must handle the
corresponding exceptions.

-- 
Catalin