[libvirt] [PATCH] virDomainObjListAddLocked: fix double free

[Michal Privoznik]

On 08/27/2018 06:06 PM, Marc Hartmayer wrote:
> On Mon, Aug 27, 2018 at 04:03 PM +0200, Michal Prívozník <mprivozn at redhat.com> wrote:
>> On 08/27/2018 03:20 PM, Marc Hartmayer wrote:
>>> If @vm has flagged as "to be removed" virDomainObjListFindByNameLocked
>>> returns NULL (although the definition actually exists). Therefore, the
>>> possibility exits that "virHashAddEntry" will raise the error
>>> "Duplicate key" => virDomainObjListAddObjLocked fails =>
>>> virDomainObjEndAPI(&vm) is called and this leads to a freeing of @def
>>> since @def is already assigned to vm->def. But actually this leads to
>>> a double free since the common usage pattern is that the caller of
>>> virDomainObjListAdd(Locked) is responsible for freeing @def in case of
>>> an error.
>>>
>>> Let's fix this by setting vm->def to NULL in case of an error.
>>>
>>> Backtrace:
>>>
>>>    ➤  bt
>>>    #0  virFree (ptrptr=0x7575757575757575)
>>>    #1  0x000003ffb5b25b3e in virDomainResourceDefFree
>>>    #2  0x000003ffb5b37c34 in virDomainDefFree
>>>    #3  0x000003ff9123f734 in qemuDomainDefineXMLFlags
>>>    #4  0x000003ff9123f7f4 in qemuDomainDefineXML
>>>    #5  0x000003ffb5cd2c84 in virDomainDefineXML
>>>    #6  0x000000011745aa82 in remoteDispatchDomainDefineXML
>>>    ...
>>>
>>> Reviewed-by: Bjoern Walk <bwalk at linux.ibm.com>
>>> Signed-off-by: Marc Hartmayer <mhartmay at linux.ibm.com>
>>> ---
>>>  src/conf/virdomainobjlist.c | 4 +++-
>>>  1 file changed, 3 insertions(+), 1 deletion(-)
>>>
>>> diff --git a/src/conf/virdomainobjlist.c b/src/conf/virdomainobjlist.c
>>> index 52171594f34f..805fe9440afe 100644
>>> --- a/src/conf/virdomainobjlist.c
>>> +++ b/src/conf/virdomainobjlist.c
>>> @@ -329,8 +329,10 @@ virDomainObjListAddLocked(virDomainObjListPtr doms,
>>>              goto cleanup;
>>>          vm->def = def;
>>>
>>> -        if (virDomainObjListAddObjLocked(doms, vm) < 0)
>>> +        if (virDomainObjListAddObjLocked(doms, vm) < 0) {
>>> +            vm->def = NULL;
>>>              goto error;
>>> +        }
>>>      }
>>>   cleanup:
>>>      return vm;
>>>
>>
>>
>> How about setting vm->def only after virDomainObjListAddObjLocked()
>> succeded?
> 
> Unfortunately, this doesn’t work as vm->def->name is used by
> virDomainObjListAddObjLocked().

Ah right.

> 
> Another solution that came to my mind is to “hand over” the
> responsibility for @def to virDomainObjListAdd(Locked) (this would
> require that a 'virDomainDefPtr *def' instead of 'virDomainDefPtr def'
> is passed). This would also eliminates the “def = NULL;” lines after a
> successful virDomainObjListAddObj call and it would solve this bug.

Yep, this looks like a good solution. Because we are close to the
release I am going to push this version and what you are suggesting can
then be saved to a follow up patch.

> 
>>
>> However, this points to a more sever bug. If a domain is being removed,
>> and some other thread is trying to define it back, I guess the whole
>> operation should succeed. Definitely not fail with some (for user)
>> cryptic message like "double key found in a hash table".
>>
>> The problem is that:
>>
>> a) both virDomainObjListFindByUUIDLocked() and
>> virDomainObjListFindByNameLocked() return NULL even if domain exists.
>> They should return the found domain and only the caller should decide if
>> vm->removing is relevant or not.
> 
> Yep, but this is another problem since virDomainObjListAddObjLocked can
> also fail under other circumstances.
> 
>>
>> b) nothing is clearing out the vm->removing flag. The
>> virDomainObjListAddObjLocked() looks like a good candidate to do so.
> 
> Haven’t looked into much detail for that part.

Okay, I'll look into it. By skimming the code quickly I can see a
potential problem there. But maybe I'm wrong.

ACKed and pushed.

Michal