Re: [PATCH RFC v1 1/1] hw/s390x/ccw: Disable legacy virtio-pci by default (v11.1+)

[Cornelia Huck <cohuck@redhat.com>]

On Mon, Apr 20 2026, Matthew Rosato <mjrosato@linux.ibm.com> wrote:

> On 4/20/26 6:12 AM, Cornelia Huck wrote:
>> On Fri, Apr 17 2026, Jaehoon Kim <jhkim@linux.ibm.com> wrote:
>> 
>>> On the s390 Linux kernel, IO_SPACE_LIMIT has been 0 since the initial
>>> zPCI implementation (commit cd24834130ac "s390/pci: base support"),
>>> making I/O BARs unusable.
>>>
>>> However, when virtio-pci devices operate in transitional mode, QEMU
>>> unconditionally exposes the legacy interface via BAR0. This results in
>>> firmware warnings during PCI enumeration, such as:
>>>
>>>   pci 0005:00:00.0: [Firmware Bug]: BAR 0: invalid; can't size
>>>
>>> even though BAR0 is never usable on the s390 kernel.
>>>
>>> Close this gap by disabling legacy virtio-pci support starting from
>>> machine version 11.1. This effectively makes virtio-pci devices
>>> non-transitional and prevents the creation of the unusable legacy I/O
>>> BAR.
>>>
>>> This introduces s390x-specific global compatibility properties that
>>> set disable-legacy=on as the default for virtio-pci devices. Machine
>>> versions v11.0 and earlier set disable-legacy=off to maintain their
>>> original default behavior (legacy support enabled), ensuring VMs
>>> created with those versions continue to work identically.
>>>
>>> Users can override the default on the command line if needed:
>>>   - On v11.1+: -global virtio-pci.disable-legacy=off (to enable legacy)
>>>   - On v11.0-: -global virtio-pci.disable-legacy=on (to disable legacy)
>> 
>> Makes sense to disable legacy virtio-pci devices, if they cannot work
>> anyway. I'm wondering if we want to have a generic "no legacy" switch as
>> well. I remember a patch from some time ago, but that was concerned with
>> endianness IIRC.
>
> Thanks Connie, I was hoping to get your input on this idea.
>
> By generic, do you mean disabling legacy virtio-pci (and legacy
> virtio-ccw?) across QEMU for all platforms vs only disabling it for s390x?

If we are moving towards multi-arch binaries, it would need to be the
latter; another benefit would be removing legacy bits from the virtio
core. But I think we'd still need quite some time before we get there
(virtio-pci switched to modern by default later than virtio-ccw.)

>
>> 
>> Anyway, we cannot turn off pre-1.0 for virtio-ccw at the moment, because
>> the bios still uses legacy virtio. Would be a bit of an effort to change
>> that, but still sounds like a good idea to me.
>> 
>
> Good point, hadn't thought about that and I guess it would be a pre-req
> if you really wanted to disable legacy virtio everywhere; we can
> certainly put this on our todo list for later, but I'd still like a
> patch like this one to fix the awkward s390x virtio-pci situation now.

Yep, this patch is fine for the current situation, I guess legacy
virtio-ccw will stay around for a bit longer (and there's no urgent need
to remove support for it as long as the virtio core still supports
legacy easily.)