Re: [Qemu-devel] "x-disable-pcie" virtio-pci property in compat_props (HW_COMPAT_2_4)

[Marcel Apfelbaum]

On 02/09/2016 07:38 PM, Laurent Vivier wrote:
> Hi,
>
> I'm playing with a qemu-2.5.0 and pc-i440fx-2.4 machine type, and
> perhaps I don't understand correctly the compat_props machinery but
> there is something strange for me:
>
> in qemu-2.5.0, hw/virtio/virtio-pci.c:
>
>     1880     DEFINE_PROP_BIT("x-disable-pcie", VirtIOPCIProxy, flags,
>     1881                     VIRTIO_PCI_FLAG_DISABLE_PCIE_BIT, false),
>
> So I guess a new default machine (pc-i440fx-2.5) will be defined by
> default with "x-disable-pcie" set to "false".

Correct, but we are talking about both pc and q35.
By default, virtio devices *can be* PCIe for all new machines.

> and in qemu-2.5, include/hw/compat.h:
>
>        3
>        4 #define HW_COMPAT_2_4 \
>        5         {\
> ..
>       14             .driver   = "virtio-pci",\
>       15             .property = "x-disable-pcie",\
>       16             .value    = "on",\
>       17         },{\
>
> So I guess a new machine pc-i440fx-2.4 created with qemu-2.5.0 will have
> "x-disable-pcie" set to true (as with qemu-2.4.0 this feature is not
> available).

Exactly, machines before 2.5 will not have virtio devices PCIe
Old q35 machines are also included.

> But in this case "x-disable-pcie" is always set to "false".

For older machines yes, for newer no.

> So the questions are:
> - do I understand correctly?

Does the above respond to your question?

> - I check "x-disable-pcie" for "virtio-*-pci" devices (virtio-net-pci,
> virtio-balloon-pci, ..."), should they inherit the behavior from their
> parent "virtio-pci"? (they should, as virtio-pci is an abstract device type)

Yes, they do inherit this property.

> My test:
>
> - take qemu-2.5.0 (a8c40fa Update version for v2.5.0 release)
> - start it with "-M pc-i440fx-2.5 -device virtio-balloon-pci -S",
>    "info qtree":
>
>      bus: pci.0
>        type PCI
>        dev: virtio-balloon-pci, id ""
> ...
>          x-disable-pcie = false

This is correct again.

> - start it with "-M pc-i440fx-2.4 -device virtio-balloon-pci -S",
>    "info qtree":
>
>      bus: pci.0
>        type PCI
>        dev: virtio-balloon-pci, id ""
> ...
>          x-disable-pcie = false

This is a bug! Thanks for finding it.
It took a little but I found the root cause.

We have 2 virtio features mapped into the same bit:
- #define VIRTIO_PCI_FLAG_MIGRATE_EXTRA_BIT 4
- #define VIRTIO_PCI_FLAG_DISABLE_PCIE_BIT 4

Since both properties appear in HW_COMPAT_2_4, but
migrate-extra is the last one used, it wins the race.


To the obvious question of "how did that happen?" I can say we had an unlucky 
break.
Both Jason and me worked on a new different virtio feature in the same time,
and they were both merged in the same pull request.
We both saw BIT 3 as the last used :)

https://lists.gnu.org/archive/html/qemu-devel/2015-11/msg03041.html

I'll send a patch to map virtio pcie feature to a different  bit.
Thanks,
Marcel


> Any explanation is very welcome.
>
> Thanks,
> Laurent