On Mon, Feb 28, 2022 at 07:16:56AM -0500, Michael S. Tsirkin wrote: > On Fri, Feb 25, 2022 at 02:35:36PM +0000, Daniel P. Berrangé wrote: > > On Fri, Feb 25, 2022 at 09:10:44AM -0500, Michael S. Tsirkin wrote: > > > QOM reference counting is not designed with an infinite amount of > > > references in mind, trying to take a reference in a loop will overflow > > > the integer. We will then eventually assert when dereferencing, but the > > > real problem is in object_ref so let's assert there to make such issues > > > cleaner and easier to debug. > > > > What is the actual bug / scenario that led you to hit this problem ? > > E.g. if during code development I call object_ref but not object_unref, > the counter eventually overflows. If this triggers in an error flow > and not a good path this kind of bug might thinkably make it through QE > into release code. > > > I'm surprised you saw an assert in object_unref, as that would > > imply you had exactly UINT32_MAX calls to object_ref and then > > one to object_unref. > > Any imbalance with # of unrefs > # refs > will trigger an existing assert in unref. > > However, an imbalance with # of refs > # unrefs does not trigger an > assert at the moment. >
A vsock patch Stefano just posted would be one example where this can happen. > > > Some micro-benchmarking shows using fetch and add this is essentially > > > free on x86. > > > > > > Signed-off-by: Michael S. Tsirkin <[email protected]> > > > --- > > > qom/object.c | 6 +++++- > > > 1 file changed, 5 insertions(+), 1 deletion(-) > > > > > > diff --git a/qom/object.c b/qom/object.c > > > index 4f0677cca9..5db3974f04 100644 > > > --- a/qom/object.c > > > +++ b/qom/object.c > > > @@ -1167,10 +1167,14 @@ GSList *object_class_get_list_sorted(const char > > > *implements_type, > > > Object *object_ref(void *objptr) > > > { > > > Object *obj = OBJECT(objptr); > > > + uint32_t ref; > > > + > > > if (!obj) { > > > return NULL; > > > } > > > - qatomic_inc(&obj->ref); > > > + ref = qatomic_fetch_inc(&obj->ref); > > > + /* Assert waaay before the integer overflows */ > > > + g_assert(ref < INT_MAX); > > > > Not that I expect this to hit, but why choose this lower > > bound instead of g_assert(ref > 0) which is the actual > > failure scenario, matching the existing object_unref > > assert. > > The earlier we catch it the better, if we overflowed to 0 some other > thread might already be confused. > > > > Regards, > > Daniel > > -- > > |: https://berrange.com -o- https://www.flickr.com/photos/dberrange > > :| > > |: https://libvirt.org -o- https://fstop138.berrange.com > > :| > > |: https://entangle-photo.org -o- https://www.instagram.com/dberrange > > :|
