[libvirt] [PATCH] virnwfilterbindingobj: Introduce and use virNWFilterBindingObjStealDef

[Michal Privoznik]

On 3/20/19 12:17 PM, John Ferlan wrote:
> 
> 
> On 3/20/19 6:31 AM, Michal Privoznik wrote:
>> https://bugzilla.redhat.com/show_bug.cgi?id=1686927
>>
>> When trying to create a nwfilter binding via
>> nwfilterBindingCreateXML() we may encounter a crash. The sequence
>> of functions called is as follows:
>>
>> 1) nwfilterBindingCreateXML() parses the XML and calls
>> virNWFilterBindingObjListAdd() which calls
>> virNWFilterBindingObjListAddLocked()
>>
>> 2) Here, @binding is not found because binding->remove is set.
>>
>> 3) Therefore, controls continue with creating new @binding,
>> setting its def to the one from 1) and adding it to the hash
>> table.
>>
>> 4) This fails, because the binding is still in the hash table
>> (duplicate key is detected).
>>
>> 5) The control jumps to 'error' label where
>> virNWFilterBindingObjEndAPI() is called which frees the binding
>> definition passed.
>>
>> 6) Error is propagated to the caller, which calls
>> virNWFilterBindingDefFree() over the definition again.
>>
>> The solution is to unset binding->def in case of failure so it's
>> not freed in step 5).
>>
>> Signed-off-by: Michal Privoznik <mprivozn at redhat.com>
>> ---
>>
>> Technically, this is a v2 of:
>>
>> https://www.redhat.com/archives/libvir-list/2019-March/msg01209.html
>>
>> But since this one implements different approach than v1 I'm not marking
>> it as such.
>>
>>   src/conf/virnwfilterbindingobj.c     | 10 ++++++++++
>>   src/conf/virnwfilterbindingobj.h     |  3 +++
>>   src/conf/virnwfilterbindingobjlist.c |  4 ++++
>>   src/libvirt_private.syms             |  1 +
>>   4 files changed, 18 insertions(+)
>>
> 
> Why wouldn't your other series take care of this in a more efficient
> way?, e.g. patch 3/3:
> 
> https://www.redhat.com/archives/libvir-list/2019-March/msg01321.html
> 

These are separate issues really. virHashAddEntry() can fail in more 
cases than just 'Duplicate key'. For instance, on OOM. The patches you 
reference help us prevent tickling this bug in case of 'Duplicate key' 
but not really for OOM.

> Isn't the problem that we can get to the virNWFilterBindingObjSetDef and
> virNWFilterBindingObjListAddObjLocked because of the same portdevname
> key name which causes the virHashAddEntry to fail.
> 
> If we fail after virNWFilterBindingObjListFindByPortDevLocked because
> the other patch recognized that the @binding was in the process of being
> removed, then we'd never create a new @obj with our new @def that had
> the duplicated portdevname key.

Again, patch you refer to will only prevent from tickling this bug.

> 
> 
>> diff --git a/src/conf/virnwfilterbindingobj.c b/src/conf/virnwfilterbindingobj.c
>> index 23978d4207..68afb9c434 100644
>> --- a/src/conf/virnwfilterbindingobj.c
>> +++ b/src/conf/virnwfilterbindingobj.c
>> @@ -88,6 +88,16 @@ virNWFilterBindingObjSetDef(virNWFilterBindingObjPtr obj,
>>   }
>>   
>>   
>> +virNWFilterBindingDefPtr
>> +virNWFilterBindingObjStealDef(virNWFilterBindingObjPtr obj)
>> +{
>> +    virNWFilterBindingDefPtr def;
>> +
>> +    VIR_STEAL_PTR(def, obj->def);
>> +    return def;
>> +}
>> +
>> +
>>   bool
>>   virNWFilterBindingObjGetRemoving(virNWFilterBindingObjPtr obj)
>>   {
>> diff --git a/src/conf/virnwfilterbindingobj.h b/src/conf/virnwfilterbindingobj.h
>> index 8e5fbee35f..b26bb3c8ec 100644
>> --- a/src/conf/virnwfilterbindingobj.h
>> +++ b/src/conf/virnwfilterbindingobj.h
>> @@ -39,6 +39,9 @@ void
>>   virNWFilterBindingObjSetDef(virNWFilterBindingObjPtr obj,
>>                               virNWFilterBindingDefPtr def);
>>   
>> +virNWFilterBindingDefPtr
>> +virNWFilterBindingObjStealDef(virNWFilterBindingObjPtr obj);
>> +
>>   bool
>>   virNWFilterBindingObjGetRemoving(virNWFilterBindingObjPtr obj);
>>   
>> diff --git a/src/conf/virnwfilterbindingobjlist.c b/src/conf/virnwfilterbindingobjlist.c
>> index 06ccbf53af..4ee2c1b194 100644
>> --- a/src/conf/virnwfilterbindingobjlist.c
>> +++ b/src/conf/virnwfilterbindingobjlist.c
>> @@ -167,6 +167,7 @@ virNWFilterBindingObjListAddLocked(virNWFilterBindingObjListPtr bindings,
>>                                      virNWFilterBindingDefPtr def)
>>   {
>>       virNWFilterBindingObjPtr binding;
>> +    bool stealDef = false;
>>   
>>       /* See if a binding with matching portdev already exists */
>>       if ((binding = virNWFilterBindingObjListFindByPortDevLocked(
>> @@ -181,6 +182,7 @@ virNWFilterBindingObjListAddLocked(virNWFilterBindingObjListPtr bindings,
>>           goto error;
>>   
>>       virNWFilterBindingObjSetDef(binding, def);
>> +    stealDef = true;
>>   
>>       if (virNWFilterBindingObjListAddObjLocked(bindings, binding) < 0)
>>           goto error;
>> @@ -188,6 +190,8 @@ virNWFilterBindingObjListAddLocked(virNWFilterBindingObjListPtr bindings,
>>       return binding;
>>   
>>    error:
>> +    if (stealDef)
>> +        virNWFilterBindingObjStealDef(binding);
> 
> I think this is no different than modifying the failure path for
> virNWFilterBindingObjListAddObjLocked to clear out @def acknowledging
> our caller would free @def upon failure return.
> 
>     if !AddObj() {
>        ObjSet(binding, NULL)
>        goto error
>     }

As said in review for v1, virNWFilterBindingObjSetDef() calls 
virNWFilterBindingDefFree(binding->def) and only after that it sets 
binding->def to desired value. This approach won't help.

> 
> Alternatively, reorder the code to do the ObjSetDef after successful
> return from virNWFilterBindingObjListAddLocked which would take
> def->portdevname as a "const char *key" type parameter; however, there
> could be some affect with virNWFilterBindingObjListLoadStatus
> processing. In particular how virNWFilterBindingObjParseXML sets obj->def...

That's an alternative approach. There are two (or more) ways to fix this 
bug. I've chosen one of them. What advantages does the other have that 
it should be done that way?

> 
> BTW: When I read the create a StealDef I was thinking more in the terms
> of "if binding is in the process of being removed", then use the
> StealDef to return the current @binding->def so it could be Free'd and
> clear the removing bit. There may need to be some obj ref counting magic
> to make sure some other thread in the process of deleting the object
> doesn't do something bad or logic in that thread to recognize the @obj
> is no longer being deleted. I didn't dig into the removing code and
> there's not enough coffee in me yet to think more clearly about that.

In general, I don't think such logic is desired. Just look at my 
reproducer (1/3 from the linked series). One thread is doing 'destroy' 
the other is doing 'create'. There can't be any logic that would make 
both APIs happy, can it? One of the pair has to fail.

Michal