Re: [PATCH v4] target/s390x: Implement the MVPG condition-code-option bit

[David Hildenbrand]

On 11.03.21 16:58, Richard Henderson wrote:
> On 3/11/21 8:03 AM, David Hildenbrand wrote:
> > As talked with Thomas off-list, there is no trusting on host==NULL
> > as well (see comment in struct S390Access). host==NULL simply
> > means we have to do individual ld/st.
>
> I think that comment is stale with the use of probe_access instead of
> tlb_vaddr_to_host -- TLB_DIRTY is in fact handled now.

Yes, it might be worth exploring in which cases we will still have 
issues and updating the comment. LAP is certainly one.

> > +    env->tlb_fill_exc = 0;
> >        flags = probe_access_flags(env, vaddr1, access_type, mmu_idx,
> >                                   nofault, &haddr1, ra);
> > +    if (env->tlb_fill_exc) {
> > +        /* We cannot rely on TLB_INVALID_MASK or haddr being NULL. */
> > +        return env->tlb_fill_exc;
> > +    }
> >        if (unlikely(size2)) {
> >            /* The access crosses page boundaries. */
> >            vaddr2 = wrap_address(env, vaddr1 + size1);
> >            flags |= probe_access_flags(env, vaddr2, access_type, mmu_idx,
> >                                        nofault, &haddr2, ra);
> > +        if (env->tlb_fill_exc) {
> > +            /* We cannot rely on TLB_INVALID_MASK or haddr being NULL. */
> > +            return env->tlb_fill_exc;
>
> But this is pretty clean, and definitely works.

Except that I need to special case CONFIG_USER_ONLY .... I'll send a v5 
later, then we can discuss when looking at the full diff (including an 
addon patch to handle r1/r2).

--
Thanks,

David / dhildenb