Re: [PATCH v4 1/7] target/arm: Improve masking of HCR RES0 bits

[Richard Henderson]

On 2/28/20 11:03 AM, Peter Maydell wrote:
> It occurs to me that we should check what the required
> semantics are for the opposite half of the register
> if the guest writes to one half of it via hcr_writehigh()
> or hcr_writelow() -- is the un-accessed half supposed
> to stay exactly as it is, or is it ok for the
> RES0-for-aarch32 bits to get squashed in the process?
> That would seem at least a bit odd even if it's valid,
> so maybe better to do aarch32 RES0 masking in
> hcr_writehigh() and hcr_writelow()?

Hmm.  You're thinking of a situation in which

 1) EL3 invokes EL2 with aa64 which sets HCR_EL2,
 2) EL3 invokes EL2 in aa32 which sets HCR which
    as a side-effect clears some of the bits in HCR2
 3) EL3 invokes EL2 with aa64 again and HCR_EL2
    incorrectly has some high bits clear.

I can't find any language that explicitly says, but "architecturally mapped"
means that the bits backing the storage must be the same.  So while it isn't
legal for aa32 to set HCR2 bit 8 (HCR_EL2.APK), I imagine that it would still
read as written.

So I think you're correct that we shouldn't alter the half B when writing to
half A.

Perhaps I should do some masking for aa32 in arm_hcr_el2_eff.


r~