Re: [Qemu-devel] [PATCH 2/4] qapi: output visitor crashes qemu if it encounters a NULL value

[Michael Roth <mdroth@linux.vnet.ibm.com>]

Quoting Markus Armbruster (2014-05-15 12:19:08)
> Michael Roth <mdroth@linux.vnet.ibm.com> writes:
> 
> > Quoting Markus Armbruster (2014-05-15 11:13:09)
> >> Marcel Apfelbaum <marcel.a@redhat.com> writes:
> >> 
> >> > A NULL value is not added to visitor's stack, but there
> >> > is no check for that when the visitor tries to return
> >> > that value, leading to Qemu crash.
> >> >
> >> > Reviewed-by: Eric Blake <eblake@redhat.com>
> >> > Signed-off-by: Marcel Apfelbaum <marcel.a@redhat.com>
> >> > ---
> >> >  qapi/qmp-output-visitor.c | 5 +++++
> >> >  1 file changed, 5 insertions(+)
> >> >
> >> > diff --git a/qapi/qmp-output-visitor.c b/qapi/qmp-output-visitor.c
> >> > index 74a5684..0562f49 100644
> >> > --- a/qapi/qmp-output-visitor.c
> >> > +++ b/qapi/qmp-output-visitor.c
> >> > @@ -66,6 +66,11 @@ static QObject *qmp_output_pop(QmpOutputVisitor *qov)
> >> >  static QObject *qmp_output_first(QmpOutputVisitor *qov)
> >> >  {
> >> >      QStackEntry *e = QTAILQ_LAST(&qov->stack, QStack);
> >> > +
> >> > +    if (!e) {
> >> > +        return NULL;
> >> > +    }
> >> > +
> >> >      return e->value;
> >> >  }
> >> 
> >> Let's see how this thing works.
> >> 
> >> The visitor's mutable state is a QStack, which is stack of (QObject,
> >> bool).  We can ignore the bool; it's just for qmp_output_next_list().
> >> 
> >> Visits start with an empty stack.  See qmp_output_visitor_new().
> >> 
> >> qmp_output_first() returns the object on the bottom of the stack.
> >> qmp_output_last() returns the object on the top of the stack.
> >> 
> >> <rant>
> >> When you implement a stack with a double-ended queue, you're totally
> >> free to pick either end of the queue for top of stack.  You're also free
> >> to name your functions accessing top and the bottom of the stack however
> >> you like.  "Of course" the author picked queue end and function names
> >> for maximum confusion:
> >> 
> >>     static QObject *qmp_output_first(QmpOutputVisitor *qov)
> >>     {
> >>         QStackEntry *e = QTAILQ_LAST(&qov->stack, QStack);
> >>         return e->value;
> >>     }
> >> 
> >>     static QObject *qmp_output_last(QmpOutputVisitor *qov)
> >>     {
> >>         QStackEntry *e = QTAILQ_FIRST(&qov->stack);
> >>         return e->value;
> >>     }
> >> 
> >> I hate you.
> >> </rant>
> >> 
> >> The result of the visit sits at the bottom of the stack.  Empty stack,
> >> null result.  See qmp_output_get_qobject().
> >> 
> >> Visiting a scalar type creates the appropriate scalar QObject, and
> >> "adds" it.  We'll find out what "adding" means shortly.  See
> >> qmp_output_type_{int,bool,str,number}().
> >> 
> >> Special case: null strings get converted to empty strings.  See
> >> qmp_output_type_str().
> >> 
> >> Starting a struct visit creates a QDict, adds it, and pushes it onto the
> >> stack.  Ending it pops it from the stack.  See
> >> qmp_output_{start,end}_struct().
> >> 
> >> Starting a list visit creates a QList, adds it, and pushes it onto the
> >> stack.  Ending it pops it from the stack.  See
> >> qmp_output_{start,end}_list().
> >> 
> >> Visiting a list member does nothing interesting; see
> >> qmp_output_next_list().  Aside: I suspect the GenericList traversal
> >> stuff now done in every next_list() method should be done in the visitor
> >> core instead.
> >> 
> >> Now let's figure out what it means to "add" an object.  This is
> >> qmp_output_add_obj().
> >> 
> >> If the stack is still empty, the object is the root object, and it gets
> >> pushed.
> >> 
> >> Else, if the object on top of the stack is a QDict, we're visiting a
> >> struct.  Enter the object into the QDict.
> >> 
> >> Else, if the object on top of the stack is a QList, we're visiting a
> >> list.  Append the object to the QList.
> >> 
> >> Else, the object on top of the stack must be scalar, and I think it must
> >> be the root object.  We replace it by the object being added.  WTF?
> >> 
> >> This feels more complicated than it could be.  Anyway, how could a null
> >> object end up at the bottom of the stack, so that qmp_output_first()
> >> chokes on it?  I can't see that.
> >> 
> >> If it can get added, then why can it be seen only by qmp_output_first(),
> >> but not by qmp_output_last() and qmp_output_pop()?
> >
> > See my note above, the corner case we're hitting seems to be when there's
> > nothing in the stack at all: generating a QObject from an empty
> > QmpOutputVisitor.
> 
> The other user of qmp_output_first() calls it like this:
> 
>     QObject *root = QTAILQ_EMPTY(&v->stack) ? NULL : qmp_output_first(v);
> 
> Patching qmp_output_first() makes this check redundant.
> 
> I suspect we should change both callers to test QTAILQ_EMPTY() instead.

Or remove the redundant check, don't have a strong preference either way,
though personally I think qmp_output_first should handle it internally
and just give us the (possibly NULL) QObject, since it exposes less
internals to callers.

> 
> > This occurs with object_property_get_str skips visit_type_str if the
> > property-specific accessor returns NULL, but we still covert the
> > visitor to a QObject to pull the string out later.
> 
> Can't see visit_type_str() being called from object_property_get_str().
> Do you mean property_get_str()?

Yah, I think it's something like:
 object_property_get_str->object_property_get->prop.get->property_get_str

> 
> static void property_get_str(Object *obj, Visitor *v, void *opaque,
>                              const char *name, Error **errp)
> {
>     StringProperty *prop = opaque;
>     char *value;
> 
>     value = prop->get(obj, errp);
>     if (value) {
>         visit_type_str(v, &value, name, errp);
>         g_free(value);
>     }
> }
> 
> Why do we skip visit_type_str() when value is null?

Not sure..., seems like an explicit fall through to the
QERR_INVALID_PARAMETER_TYPE error in object_property_get_str that
wasn't reachable prior to Marcel's patch (due to segfault), so
I'm not sure this code path was in play until now.