[libvirt] [PATCH] hotplug: Fix libvirtd crash on qemu-attached guest

[John Ferlan]

On 09/22/2014 05:23 PM, Eric Blake wrote:
> On 09/22/2014 02:49 PM, John Ferlan wrote:
>> https://bugzilla.redhat.com/show_bug.cgi?id=1141621
>>
>> Using qemu-attach to attach to a qemu created process and then
>> attempting to virsh detach-interface caused a libvirtd crash
>> since the assumption was that device aliases were in place and
>> that assumption is not necessarily true for the qemu-attach environment.
> 
> Hmm. I'm wondering whether it is better to patch qemu-attach to populate
> alias information at attach time rather than having to chase all the
> places that will later fail when the assumption is broken.  While your
> patch is certainly more appealing for being smaller, it makes me wonder
> if it is the only such place bitten by the assumptions.
> 

I took this approach mainly because it seemed that generating aliases in
the qemu-attach path was explicitly not done for a reason while it was
done for the qemuConnectDomainXMLToNative and qemuDomainQemuAttach
paths. For each of the other attach paths, there are the separate calls
to the specific qemuAssignDevice*Alias functions.  So it does seem that
this qemu-attach path is the only current one without the aliases set.

If there is some path in the future that doesn't set the alias properly,
then at the very least we don't have a libvirtd crash at least for this
path.

FWIW: If I modify the patch as follows:

-    if (virQEMUCapsGet(priv->qemuCaps, QEMU_CAPS_DEVICE) &&
-        detach->info.alias) {
+    if (virQEMUCapsGet(priv->qemuCaps, QEMU_CAPS_DEVICE)) {
+        if (!detach->info.alias) {
+            virReportError(VIR_ERR_OPERATION_FAILED, "%s",
+                           ("device alias not found: cannot detach net "
+                            "device not properly attached"));
+            qemuDomainObjExitMonitor(driver, vm);
+            goto cleanup;
+        }

then rerun the sequence of steps, I get:

error: Failed to detach interface
error: operation failed: device alias not found: cannot detach net
device not properly attached


If I then add qemuAssignDeviceAliases() as described below and rerun,
the error changes to:

error: Failed to detach interface
error: operation failed: detaching net0 device failed: Device 'net0' not
found


Since the original start didn't use the "-netdev/-device" syntax that
would be expected if the DRIVE capability was set.  So no worse than we
were before.

Should I just post the additional patch calling qemuAssignDeviceAliases
and go from there?


>>
>> Add some more verbiage to the virsh man page.
>>
>> Signed-off-by: John Ferlan <jferlan at redhat.com>
>> ---
>> Notes:
>>  * The bz also mentions a failure for the hotplug/virsh attach-interface,
>>    but that has been resolved by commit id 'ba7468db'
>>  * For the bz, the attempt to attach-interface/hotplug still is not
>>    successful, since device aliases are not set for the PCI controller.
>>    * I did try adding a call to qemuAssignDeviceAliases() during the qemu-
>>      attach code; however, that failed as well (perhaps for a different
>>      reason) as the attach to PCI slot 3 function 0 was not available
>>      for rtl8139 since it was being used by e1000
> 
> That is, I'm wondering if this approach should be used after all, and we
> have yet another problem to solve first.
> 

Using virsh/libvirt, I can define/start a guest without a network and
add one... So perhaps the qemu-attach path is "unique" with respect to
how qemu defines things and perhaps there's something "hidden"/unknown
causing e1000 to grab PCI slot 3.  I didn't dig into that.

John

>>    * If it's desired to add the alias call in that's fine, then that could
>>      be added in qemuDomainQemuAttach prior to the qemuDomainAssignAddresses 
>>      call similar to qemuConnectDomainXMLToNative.
>>
>>  src/qemu/qemu_hotplug.c | 3 ++-
>>  tools/virsh.pod         | 5 +++--
>>  2 files changed, 5 insertions(+), 3 deletions(-)
>>
>> diff --git a/src/qemu/qemu_hotplug.c b/src/qemu/qemu_hotplug.c
>> index 7bc19cd..b7514ce 100644
>> --- a/src/qemu/qemu_hotplug.c
>> +++ b/src/qemu/qemu_hotplug.c
>> @@ -3520,7 +3520,8 @@ qemuDomainDetachNetDevice(virConnectPtr conn,
>>      qemuDomainMarkDeviceForRemoval(vm, &detach->info);
>>  
>>      qemuDomainObjEnterMonitor(driver, vm);
>> -    if (virQEMUCapsGet(priv->qemuCaps, QEMU_CAPS_DEVICE)) {
>> +    if (virQEMUCapsGet(priv->qemuCaps, QEMU_CAPS_DEVICE) &&
>> +        detach->info.alias) {
>>          if (qemuMonitorDelDevice(priv->mon, detach->info.alias) < 0) {
>>              qemuDomainObjExitMonitor(driver, vm);
>>              virDomainAuditNet(vm, detach, NULL, "detach", false);
>> diff --git a/tools/virsh.pod b/tools/virsh.pod
>> index 9919f92..947adaf 100644
>> --- a/tools/virsh.pod
>> +++ b/tools/virsh.pod
>> @@ -3686,8 +3686,9 @@ using the UNIX driver. Ideally the process will also have had the
>>  
>>  Not all functions of libvirt are expected to work reliably after
>>  attaching to an externally launched QEMU process. There may be
>> -issues with the guest ABI changing upon migration, and hotunplug
>> -may not work.
>> +issues with the guest ABI changing upon migration and device hotplug
>> +or hotunplug may not work. The attached environment should be considered
>> +primarily read-only.
> 
> This doc hunk is fine to push now, whether or not we have consensus on
> the code side.
>