Re: [Qemu-devel] [PATCH 1/2] pcie: drop version_id field for live migration

[Juan Quintela]

Jason Baron <address@hidden> wrote:
> While testing q35 live migration, I found that the migration would abort with
> the following error: "Unknown savevm section type 76".

Before we start, migration of PCI is ugly due to backwards compatibility.

> The error is due to this check failing in 'vmstate_load_state()':
>
>     while(field->name) {
>         if ((field->field_exists &&
>              field->field_exists(opaque, version_id)) ||
>             (!field->field_exists &&
>              field->version_id <= version_id)) {

This is a long history, but to make things short, when we started with
vmstate, there were no cose where this were not true.  It could
potentially be true, but there were no examples.

> The VMSTATE_PCIE_DEVICE() currently has a 'version_id' set to 2. However,
> 'version_id' in the above check is 1. And thus we fail to load the pcie device
> field. Further the code returns to 'qemu_loadvm_state()' which produces the
> error that I saw.

Agreed about removing the set of the value.

> I'm proposing to fix this by simply dropping the 'version_id' field from
> VMSTATE_PCIE_DEVICE(). VMSTATE_PCI_DEVICE() defines no such field and further
> the vmstate_pcie_device that VMSTATE_PCI_DEVICE() refers to is already
> versioned. Thus, any versioning issues could be detected at the vmsd level.
>
> Taking a step back, I think that the 'field->version_id' should be compared
> against a saved version number for the field not the 'version_id'. Futhermore,
> once vmstate_load_state() is called recursively on another vmsd, the check of:
>
>     if (version_id > vmsd->version_id) {
>         return -EINVAL;
>     }

This is a can of worms you don't want to open.  Basically it is
imposible to get it _sane_ and backward compatible.  We choose on its
time to maintain it backward compatible, and that is why it is not sane.
I would have to dig the details, but the problems where when we were
loading things that were not backward compatible.  The problem was that
PCI is the other "section inside a section", where we store the
version_id.  This is wrong for a number of reasons, between them that if
we "increase" the version_id of pci, the "outermost" section id, has
changed but it would not receive a new version number.

Remember the part about _sane_ and _backward compatible_ :-()

> Will never fail since version_id is always equal to vmsd->version_id. So I'm
> wondering why we aren't storing the vmsd version id of the source in the
> migration stream?
>
> This patch also renames the 'name' field of vmstate_pcie_device from:
> PCIDevice -> PCIEDevice to differentiate it from vmstate_pci_device.
>
> Signed-off-by: Jason Baron <address@hidden>

Acked-by: Juan Quintela <address@hidden>

I agree with the migration bits of this series.  I am not ack'ing the
other patch because I don't understand pci well enough to know what the
code does (before or after).

Later, Juan.