[PATCH] qemu: forbid updating any attributes of an interface <backend> with update-device

[Laine Stump]

On 2/21/23 4:20 AM, Martin Kletzander wrote:
> On Tue, Feb 21, 2023 at 09:11:11AM +0100, Martin Kletzander wrote:
>> On Wed, Feb 15, 2023 at 04:06:41PM -0500, Laine Stump wrote:
>>> Changing any of the attributes of an <interface>'s <backend> would
>>> require removing and re-adding the interface for the new setting to
>>> take effect, so fail any update-device that changes anything in
>>> <backend>
>>>
>>> Resolves: https://bugzilla.redhat.com/2169245
>>> Signed-off-by: Laine Stump <laine at redhat.com>
>>> ---
>>> src/conf/domain_conf.c   | 14 ++++++++++++++
>>> src/conf/domain_conf.h   |  2 ++
>>> src/libvirt_private.syms |  1 +
>>> src/qemu/qemu_hotplug.c  |  9 +++++++++
>>> 4 files changed, 26 insertions(+)
>>>
>>> diff --git a/src/conf/domain_conf.c b/src/conf/domain_conf.c
>>> index 3dfc5c87af..aa7bed7dc3 100644
>>> --- a/src/conf/domain_conf.c
>>> +++ b/src/conf/domain_conf.c
>>> @@ -19853,6 +19853,20 @@ 
>>> virDomainFsDefCheckABIStability(virDomainFSDef *src,
>>> }
>>>
>>>
>>> +bool
>>> +virDomainNetBackendIsEqual(virDomainNetBackend *src,
>>> +                           virDomainNetBackend *dst)
>>> +{
>>> +    if (src->type != dst->type ||
>>> +        STRNEQ_NULLABLE(src->tap, dst->tap) ||
>>> +        STRNEQ_NULLABLE(src->vhost, dst->vhost) ||
>>> +        STRNEQ_NULLABLE(src->logFile, dst->logFile)) {
>>
>> I thought you are missing the @upstream member comparison here, but it
>> looks like it is not used anywhere in the code.  So I guess
>>
>> Reviewed-by: Martin Kletzander <mkletzan at redhat.com>
>>
>> and I'll do some grave-digging about the long lost member.
>>
> 
> And so I did.  And I'm a bit baffled by the whole function now.
> 
> 1) There is no check that the sourceDev does not change. OK, I can add
>     that.

Well, yeah, I did miss that, but on the other hand, that isn't in 
<backend>, which is what the BZ was filed about, and the subject of the 
patch (and more specifically of this function). I know that's 
nitpicking, but that explains why the check isn't here (I mean that, 
along with the fact that I forgot :-))


> 
> 2) The other interface type using sourceDev (direct) is also not
>     checking the change.  OK, maybe we support that?

Believe me when I say there is *a lot* of stuff not checked in that 
function; at one time it may have been closer to complete, but things 
get added and people forget to add them to that function. I definitely 
should have included a check for ->sourceDev (and all the stuff in 
->backend / <backend>) in the original patches.


> 3) There is no code that changes that device, or copies
>     newdev->sourceDev from olddev->sourceDev (like with ->ifname, for
>     example)

It *shouldn't* be copied from olddev to newdev because it is never 
autofilled during parse or guest startup / device hotplug - the parts of 
the object that are copied from old to new (if new->blah is empty) are 
only those things that get filled in automatically with "something" if 
the attribute isn't specified in the original XML. If you don't specify 
<source dev='blah'> then sourceDev remains empty. (And in the case of 
the <source dev='blah'> for direct (macvtap) interfaces, it is a 
required field, so it can never be empty, and therefore will never be 
auto-filled with some value if it's missing)

olddev->ifname, on the other hand, almost always is auto-filled by 
libvirt as the guest is starting (a tap/macvtap device is created using 
a generated name, and that generated name is stored in ->ifname, which 
ends up being displayed in <target dev='blah'> when the netdef is 
formatted to XML). If the user has specified a <target dev> in the 
update XML, and that dev is different from what's in use, we do need to 
flag that as an error; but if they don't specify anything for target 
dev, or if what they specified happens to match what it is already set 
to, then that isn't an error; the cleanest way to handle those cases 
properly is to copy the attribute from olddev to newdev iff the 
attribute is empty in newdev.

> 
> 4) needReconnect is set in couple of places, together with the comment
>     "we'll need a full reconnect", giving the impression that the
>     function will handle that.  But the value is read in only one place,
>     giving an error.

needReconnect was a very nice idea, based on the (false) promise that it 
was going to be possible to restart the netdev (host) side of an 
emulated network device without completely unplugging it on the guest 
side. At one point back in 2011 or 2012 or something, *someone* on the 
QEMU team made a patch that madde this "kind of" work, and so I modified 
the update function to classify some previously disallowed changes as 
needing a full "reconnect" (basically remove the current -netdev and 
replace it with a new one, without the need of completely unplugging the 
device on the guest side). Until QEMU had that working, a single check 
for needReconnect at the end was left generating an error.

Unfortunately the QEMU side of that never happened (I have a vague 
memory that the person who at first suggested they could do it, later 
told me it simply wasn't possible, for [reasons that sounded plausible], 
so the bit that logs an error when needReconnect == true was never 
replaced with code to actually implement the reconnect.

> 
> Honestly, I feel uncertain about sending a patch to "fix" this because
> I could miss something and I don't feel like digging into this just for
> the sake of one member _maybe_ not being accounted for.

Well, you're correct that <source dev='blah'> should be checked for 
being changed, when the netdev type is 'user' (so it's in ->sourceDev), 
'direct' (so it's in ->data.direct.linkdev), and vdpa (so it's in 
->data.vdpa.devicepath). But again that's not in <backend>, and there 
are *so many* other things not being checked for, and it could be a real 
rabbit hole :-)

P.S. I forgot to add your Reviewed-by: before pushing (and also forgot 
to capitalize the first word of the comment). Sorry. That's what happens 
when I don't get enough sleep and then wake up early and push something 
before my brain is fully functioning :-/

> 
>> [...]
>>
>>> diff --git a/src/qemu/qemu_hotplug.c b/src/qemu/qemu_hotplug.c
>>> index c490e2b97a..b4cddef9f5 100644
>>> --- a/src/qemu/qemu_hotplug.c
>>> +++ b/src/qemu/qemu_hotplug.c
>>> @@ -3675,6 +3675,15 @@ qemuDomainChangeNet(virQEMUDriver *driver,
>>>         goto cleanup;
>>>     }
>>>
>>> +    /* nothing in <backend> can be modified in an existing interface -
>>> +     * the entire device will need to be removed and re-added.
>>
>> If I were a nit picker I could tell you to start the sentence with an
>> uppercase letter or remove the full stop.  O:-)
>>
>>> +     */
>>> +    if (!virDomainNetBackendIsEqual(&olddev->backend, 
>>> &newdev->backend)) {
>>> +        virReportError(VIR_ERR_OPERATION_UNSUPPORTED, "%s",
>>> +                       _("cannot modify network device backend 
>>> settings"));
>>> +        goto cleanup;
>>> +    }
>>> +
>>>     /* allocate new actual device to compare to old - we will need to
>>>      * free it if we fail for any reason
>>>      */
>>> -- 
>>> 2.39.1
>>>
> 
>