Re: [PATCH 03/14] accel/tcg: Suppress auto-invalidate in probe_access_internal

[Richard Henderson]

On 8/23/22 02:19, David Hildenbrand wrote:
> 1) s390_probe_access() documents to "With nonfault=1, return the PGM_
> exception that would have been injected into the guest; return 0 if no
> exception was detected."
>
> But in case of CONFIG_USER_ONLY, we return the flags returned by
> s390_probe_access(), not a PGM__* value. Maybe it doesn't matter,
> because we'll simply inject a SIGSEGV in any case ...

I would have said it would matter for MVPG, except that is incorrectly *not* marked as a 
privileged instruction.  There should be no CONFIG_USER_ONLY case to answer there.

> 2) s390_probe_access() documents that for "CONFIG_USER_ONLY, the
> faulting address is stored to env->__excp_addr.".
>
> However, that's only set in s390_cpu_record_sigsegv(). With nonfault=1
> that will never actually trigger, right?

Correct.

> I assume db9aab5783a2 ("target/s390x: Use probe_access_flags in
> s390_probe_access") might have introduced both. We had a flag conversion
> to PGM_ in there and stored env->__excp_addr:

Indeed, that commit is faulty in that it breaks the contract of 
s390_probe_access.
It's a shame, though, that we need to carry the extra code for the purpose, and that the 
generic interfaces are not sufficient.


r~