Re: [PATCH v2] piix: fix regression during unplug in Xen HVM domUs

[John Snow]

On Fri, May 12, 2023 at 5:14 PM Stefano Stabellini
<sstabellini@kernel.org> wrote:
>
> On Wed, 10 May 2023, Olaf Hering wrote:
> > Wed, 10 May 2023 00:58:27 +0200 Olaf Hering <olaf@aepfle.de>:
> >
> > > In my debugging (with v8.0.0) it turned out the three pci_set_word
> > > causes the domU to hang. In fact, it is just the last one:
> > >
> > >    pci_set_byte(pci_conf + 0x20, 0x01);  /* BMIBA: 20-23h */
> > >
> > > It changes the value from 0xc121 to 0x1.
> >
> > If I disable just "pci_set_word(pci_conf + PCI_COMMAND, 0x0000);" it works 
> > as well.
> > It changes the value from 0x5 to 0.
> >
> > In general I feel it is wrong to fiddle with PCI from the host side.
> > This is most likely not the intention of the Xen unplug protocol.
> > I'm sure the guest does not expect such changes under the hood.
> > It happens to work by luck with pvops kernels because their PCI discovery
> > is done after the unplug.
> >
> > So, what do we do here to get this off the table?
>
> I don't have a concrete suggestion because I don't understand the root
> cause of the issue. Looking back at Paolo's reply from 2021
>
> https://marc.info/?l=xen-devel&m=161669099305992&w=2
>
> I think he was right. We can either fix the root cause of the issue or
> avoid calling qdev_reset_all on unplug. I am OK with either one.

I haven't touched IDE or block code in quite a long while now -- I
don't think I can help land this fix, but I won't get in anyone's way,
either. Maybe just re-submit the patches with an improved commit
message / cover letter that helps collect the info from the previous
thread, the core issue, etc.

--js