Question about issue #1576: Migration from v8.0.0-rc2 to v7.2.0 with pcie-root-port device fails

Question about issue #1576: Migration from v8.0.0-rc2 to v7.2.0 with pcie-root-port device fails

[Fiona Ebner]

Hi,
I'm trying to fix issue #1576 [0], but having a bit of a hard time.

The issue was introduced by commit 010746ae1d ("hw/pci/aer: Implement
PCI_ERR_UNCOR_MASK register") and the migration error is:
> qemu-system-x86_64: get_pci_config_device: Bad config data: i=0x10a read: 40 device: 0 cmask: ff wmask: 0 w1cmask:0

Since there is no dedicated subsection for the new register, but it's
just part of the existing buffer sent along with the device, the
approach described in docs/devel/migration.rst, "Connecting subsections
to properties" doesn't seem to work here.

AFAIU, it would be necessary to unset the new register in the buffer
before sending for older machine types (before the patch, its value was
0). Then older QEMU versions should accept the migration stream again.
Migration between >8.0 and >= 8.0 when using an older machine type would
be softly broken, because the register value is reset to 0, but not be a
hard error, because it's the same situation as forward migration from
7.2 (the error condition in get_pci_config_device() doesn't trigger,
because versions with commit 010746ae1d set the wmask for the new
register during device initialization).

Is there a good way to unset the register conditionally based on machine
type during save? Is there any fix for a similar issue in the past I can
look at?

Or will this be a WONTFIX, because it's just backwards migration and
QEMU 8.0 has already been released?

[0]: https://gitlab.com/qemu-project/qemu/-/issues/1576

Best Regards,
Fiona

Re: Question about issue #1576: Migration from v8.0.0-rc2 to v7.2.0 with pcie-root-port device fails

[Juan Quintela]

Fiona Ebner <f.ebner@proxmox.com> wrote:
> Hi,
> I'm trying to fix issue #1576 [0], but having a bit of a hard time.
>
> The issue was introduced by commit 010746ae1d ("hw/pci/aer: Implement
> PCI_ERR_UNCOR_MASK register") and the migration error is:
>> qemu-system-x86_64: get_pci_config_device: Bad config data: i=0x10a read: 40 device: 0 cmask: ff wmask: 0 w1cmask:0
>
> Since there is no dedicated subsection for the new register, but it's
> just part of the existing buffer sent along with the device, the
> approach described in docs/devel/migration.rst, "Connecting subsections
> to properties" doesn't seem to work here.
>
> AFAIU, it would be necessary to unset the new register in the buffer
> before sending for older machine types (before the patch, its value was
> 0). Then older QEMU versions should accept the migration stream again.
> Migration between >8.0 and >= 8.0 when using an older machine type would
> be softly broken, because the register value is reset to 0, but not be a
> hard error, because it's the same situation as forward migration from
> 7.2 (the error condition in get_pci_config_device() doesn't trigger,
> because versions with commit 010746ae1d set the wmask for the new
> register during device initialization).
>
> Is there a good way to unset the register conditionally based on machine
> type during save? Is there any fix for a similar issue in the past I can
> look at?
>
> Or will this be a WONTFIX, because it's just backwards migration and
> QEMU 8.0 has already been released?
>
> [0]: https://gitlab.com/qemu-project/qemu/-/issues/1576

There is a patch on list from Leonardo:

https://lists.gnu.org/archive/html/qemu-devel/2023-05/msg00350.html

Or I am missing something?

Later, Juan.