[libvirt] [RFC PATCH 2/2] conf: Add VM Generation ID device
John Ferlan
jferlan at redhat.com
Wed Mar 21 23:49:15 UTC 2018
On 03/21/2018 03:19 AM, Peter Krempa wrote:
> On Tue, Mar 20, 2018 at 18:55:44 -0400, John Ferlan wrote:
>> Add VM Generation ID device XML schema, parse, format, and documentation.
>>
>> Signed-off-by: John Ferlan <jferlan at redhat.com>
>> ---
>> docs/formatdomain.html.in | 54 ++++++++++++++
>> docs/schemas/domaincommon.rng | 21 ++++++
>> src/conf/domain_conf.c | 112 +++++++++++++++++++++++++++++-
>> src/conf/domain_conf.h | 11 +++
>> tests/qemuxml2argvdata/vmgenid-auto.xml | 32 +++++++++
>> tests/qemuxml2argvdata/vmgenid.xml | 32 +++++++++
>> tests/qemuxml2xmloutdata/vmgenid-auto.xml | 1 +
>> tests/qemuxml2xmloutdata/vmgenid.xml | 1 +
>> tests/qemuxml2xmltest.c | 3 +
>> 9 files changed, 266 insertions(+), 1 deletion(-)
>> create mode 100644 tests/qemuxml2argvdata/vmgenid-auto.xml
>> create mode 100644 tests/qemuxml2argvdata/vmgenid.xml
>> create mode 120000 tests/qemuxml2xmloutdata/vmgenid-auto.xml
>> create mode 120000 tests/qemuxml2xmloutdata/vmgenid.xml
>>
>> diff --git a/docs/formatdomain.html.in b/docs/formatdomain.html.in
>> index 6fd2189cd2..895e51b343 100644
>> --- a/docs/formatdomain.html.in
>> +++ b/docs/formatdomain.html.in
>> @@ -8027,6 +8027,60 @@ qemu-kvm -net nic,model=? /dev/null
>> </dd>
>> </dl>
>>
>> + <h4><a id="elementsVmgenid">VM Generation ID device</a></h4>
>> +
>> + <p>
>> + <span class="since">Since 4.2.0</span>, the <code>vmgenid</code>
>> + element can be used to add a Virtual Machine Generation ID device.
>> + A <code>vmgenid</code> device is an emulated device which exposes
>> + a 128-bit, cryptographically random, integer value identifier,
>> + referred to as a Globally Unique Identifier, or GUID. The value is
>> + stored within the virtual machine's BIOS so that programs running in
>
> According to the qemu docs, it's stored in the ACPI table, so it does
> not depend on BIOS being used. EFI can be used as well.
>
>> + the virtual machine can protect themselves from potential corruption
>> + by checking that the Generation ID has not changed immediately prior
>> + to committing a transaction.
>
> This last sentence is weird. I'd refrain from giving examples and rather
> stick to describe which changes can be tracked by this.
>
>
>> +
>> + The <code>vmgenid</code> device will update the BIOS entry each time
>
> BIOS ...
>
>> + the virtual machine executes from a different configuration file such
>> + as executing from a recovered snapshot or executing after restoring
>
> Okay. From this it seems that the GUID should change where the VM state
> could be reverted to a point in history of the execution of the guest
> software. This basically means that it should be always left to
> autogenerating except in cases of migration where the new qemu process
> needs to continue to use the old GUID.
>
> This means that you'll need to query it if 'auto' is specified and qemu
> does not guarantee that it's properly transported in the migration
> stream. On the other hand, if it is transported in the migration stream,
> then save/restore and reversion of external snapshots with memory will
> be fun, since that uses the migration code.
>
> You'll need to figure out where qemu stores this and how it's handled on
> migration, since I did not see it in the doc.
I read something along the way during the review of the patch series
where the migration part of this was handled properly. The v8 is here:
http://lists.nongnu.org/archive/html/qemu-devel/2017-02/msg03774.html
and within the pile of responses for patch 4 this was discussed. I'll
certainly go through that again once the XML pieces are agreed upon.
>
> Also there is one other problem and that is reversion of internal
> snapshots, which in some cases does not require restarting the emulator
> in case of qemu. In such case you'll need to make sure that the
> condition is modified so that a VM with this device will always restart
> qemu.
>
Perhaps this is what danpb was alluding to in his comment in bz 1149445.
I hadn't yet begun the process of thinking through the various
transitions - I do know I can add the vmgenid device to a "new" guest
(either either a static guid or an 'auto' guid). But that just POC for
me that it's possible to do the simple thing.
One "issue" is that there's currently no way to update/provide the guid
value for a transaction that QEMU isn't restarted. However, there is
code upstream for qemu to add that, see:
http://lists.nongnu.org/archive/html/qemu-devel/2018-03/msg00551.html
>> + from backup. Programs running in a virtual machine can protect themselves
>> + from potential corruption by checking that the generation ID has not
>> + changed immediately prior to committing a transaction, they can also
>> + use the data provided in the 128-bit identifier as a high entropy
>> + random data source.
>
> Whoah. The qemu docs don't speak about entropy source. Is it described
> in the Microsoft documentation? In any case, if the GUID is going to be
> used as an entropy source, we need to approach it way more carefully.
> Especially don't report it to unauthorized eyes. I'd be very happy if
> will not have to keep it secret. This will get especially tricky since
> there will be cases where we'll need to put it on the command line
> (either on migration or on save/restore).
>
Yes, that's all from the MS docs. I toyed with keeping the QEMU wording
only, but since this is "currently" essentially a Windows guest feature,
I went with the MS wording even though yes, it's a bit more onerous.
And based on the Aug 1, 2012 date of the MS document and the target OS
(Windows Server 2012 or Windows 8), I'd say they weren't thinking
anything beyond BIOS. In any case, I can reword things to be a bit more
vague as this moves forward.
If we avoid the capability to "supply" a guid and only allowed an 'auto'
guid, then being able to fetch the guid after the creation of the device
and save it in the qemu private data would be possible. Although saving
that does open the question again about being able to 'see' it. It is
possible to fetch the value and display by a new virsh command too, but
based on this perhaps it's best that code never sees the light of day.
>> + </p>
>> +
>> + <p>
>> + Example:
>> + </p>
>> +<pre>
>> +...
>> +<devices>
>> + <vmgenid guid='3e3fce45-4f53-4fa7-bb32-11f34168b82b'/>
>
> So I think we should use a more generic element with 'vmgenid' as a
> model. It looks like this is one of possible implementations of such a
> thing and it's fully possible to emulate a PCI device or anything else
> with similar semantics.
>
Oh interesting implementation idea. Wish you'd provided generic example
XML as well.
<devices>
<generic model='vmgenid'>
<params guid='auto'/>
</generic>
</devices>
Could make for some awful RNG code that I'll dislike writing. I prefer
to keep things a bit more simple and not try to over engineer this, but
I'm willing to work through an example.
>> +</devices>
>> +...
>> +</pre>
>> +
>> +<pre>
>> +...
>> +<devices>
>> + <vmgenid guid='auto'/>
>> +</devices>
>> +...
>> +</pre>
>> +
>> + <dl>
>> + <dt><code>guid</code></dt>
>> + <dd>
>> + <p>
>> + The required <code>guid</code> attribute can be either a provided
>> + (<a href="#elementsMetadata"><code>uuid</code></a>) formatted value
>> + or the string 'auto' if the underlying hypervisor supports creating
>
> So 'auto' in place of the guid is a qemuism according to the docs. You
> should go with a separate element for it in cases where you need to
> configure that the setting is automatic, but you also need to transfer
> the guid itself.
>
> Also the necessity to allow configuration of the GUID really depends on
> the questions I've asked above. It might be even unnecessary to allow
> configuration of the GUID itself depending on that. In cases where we'll
> need to preserve it accross migration, we'll really need a way to store
> it in the XML.
>
Now that you've brought it up, I can certainly see the downsides of
providing a static guid and enforcing that it's always autogenerated.
>> + its own value.
>> + </p>
>> + </dd>
>> + </dl>
>> +
>> <h3><a id="seclabel">Security label</a></h3>
>>
>> <p>
>> <interleave>
>
> [...]
>
>> diff --git a/src/conf/domain_conf.c b/src/conf/domain_conf.c
>> index 2f07180faa..aaba2a47f7 100644
>> --- a/src/conf/domain_conf.c
>> +++ b/src/conf/domain_conf.c
>
> [...]
>
>> @@ -15690,6 +15692,45 @@ virDomainIOMMUDefParseXML(xmlNodePtr node,
>> }
>>
>>
>> +static virDomainVMGenIDDefPtr
>> +virDomainVMGenIDDefParseXML(xmlNodePtr node,
>> + xmlXPathContextPtr ctxt)
>> +{
>> + virDomainVMGenIDDefPtr vmgenid = NULL;
>> + virDomainVMGenIDDefPtr ret = NULL;
>> + xmlNodePtr save = ctxt->node;
>> + char *guidxml = NULL;
>> +
>> + ctxt->node = node;
>> +
>> + if (VIR_ALLOC(vmgenid) < 0)
>> + goto cleanup;
>> +
>> + if (!(guidxml = virXMLPropString(node, "guid"))) {
>> + virReportError(VIR_ERR_XML_ERROR, "%s",
>> + _("missing required 'guid' attribute"));
>> + goto cleanup;
>> + }
>> +
>> + if (STREQ(guidxml, "auto")) {
>> + vmgenid->autogenerate = true;
>> + } else {
>> + if (virUUIDParse(guidxml, vmgenid->guidstr) < 0) {
>> + virReportError(VIR_ERR_XML_ERROR,
>> + _("malformed guid='%s' provided"), guidxml);
>> + goto cleanup;
>> + }
>> + }
>> +
>> + VIR_STEAL_PTR(ret, vmgenid);
>> +
>> + cleanup:
>
> 'guidxml' is leaked
>
oh right...
>> + VIR_FREE(vmgenid);
>> + ctxt->node = save;
>> + return ret;
>> +}
>> +
>> +
>> virDomainDeviceDefPtr
>> virDomainDeviceDefParse(const char *xmlStr,
>> const virDomainDef *def,
>
> [...]
>
>> @@ -21812,6 +21871,25 @@ virDomainIOMMUDefCheckABIStability(virDomainIOMMUDefPtr src,
>>
>>
>> static bool
>> +virDomainVMGenIDDefCheckABIStability(virDomainVMGenIDDefPtr src,
>> + virDomainVMGenIDDefPtr dst)
>> +{
>> + if (memcmp(src->guidstr, dst->guidstr, VIR_UUID_BUFLEN) != 0) {
>
> This does not check 'auto' separately. It will not be able to
> differentiate between 'auto' and an UUID of '00000-000...'. Also it will
> report weird error message in that case.
>
> Also given that the GUID is to be changed during lifetime of the VM I'm
> not sure whether checking it is necessary at all.
>
Hmmm... true, ironically my first pass at this was to compare src/dst
autogenerated values.
I hedged my bets in case there was a "desire" to fetch the QEMU
generated GUID value and save it while the domain was running - although
I suppose that would go in the domain private ...
Of course then I wondered why even bother with this other than some
reviewer asking - is there an ABI concerns.
John
>> + char guidsrc[VIR_UUID_STRING_BUFLEN];
>> + char guiddst[VIR_UUID_STRING_BUFLEN];
>> + virUUIDFormat(src->guidstr, guidsrc);
>> + virUUIDFormat(dst->guidstr, guiddst);
>> + virReportError(VIR_ERR_CONFIG_UNSUPPORTED,
>> + _("Target domain vmgenid guid '%s' does not match "
>> + "source '%s'"),
>> + guiddst, guidsrc);
More information about the libvir-list
mailing list