Re: [RFC PATCH 2/5] system/memory: support unaligned access

[Peter Xu <peterx@redhat.com>]

On Wed, Dec 11, 2024 at 09:56:21AM +0000, Peter Maydell wrote:
> On Fri, 6 Dec 2024 at 16:43, Peter Xu <peterx@redhat.com> wrote:
> > I assume it's about xhci_cap_ops then.  If you agree we can also mention
> > xhci_cap_ops when dscribing it, so readers can easily reference the MR
> > attributes from the code alongside with understanding the use case.
> >
> > Does it mean that it could also work if xhci_cap_ops.impl.min_access_size
> > can be changed to 2 (together with additional xhci_cap_read/write support)?
> >
> > Note that I'm not saying it must do so even if it would work for xHCI, but
> > if the memory API change is only for one device, then it can still be
> > discussed about which option would be better on changing the device or the
> > core.
> 
> I think the memory system core has been broken in this area
> for a long time -- it purports to support impls which only
> do a subset of what the valid operations are, but it actually
> does buggy and wrong things in some cases. So far
> we have effectively worked around it by avoiding defining
> MemoryRegionOps that try to use the buggy areas, but I
> think it's much better to fix the code so it really does
> what it's theoretically intended to do.

Thanks, Peter.  I assume it means there're a lot of devices that can use
this model.  Then it makes perfect sense to do it in memory core.

Though I do have some confusion on why we needed impl.unaligned at all.  I
see that Tomoyuki raised similar question, even if not exactly the same
one.  I'll try to continue the discussion there.

-- 
Peter Xu