On Tue, Dec 13, 2011 at 10:18:10AM -0500, Stephen Gallagher wrote:
> On Tue, 2011-12-13 at 16:01 +0100, Jakub Hrozek wrote:
> >
https://fedorahosted.org/sssd/ticket/1090
> >
> > Unfortunately there is not enough info in rhbz #754114 where the bug
> > originated, so just band-aid the crash for now.
>
>
> Nack.
>
> Looking at this closely, I think what probably happened is that the
> child died while waiting for the ping_check to time out, so that when it
> finally returned, the memory for the mt_svc object had been freed.
> Because we then use talloc_get_type() to retrieve that memory, it
> detects that it's invalid and simply returns NULL (which we then
> dereference and crash on).
>
> The real solution here would be to ensure that we cancel the
> DBusPendingCall with dbus_pending_call_cancel() if the mt_svc is freed.
>
> So we need to modify the mt_svc structure to contain a DBusPendingCall *
> attribute that we assign to it when service_send_ping() calls
> sbus_conn_send() (currently, we just pass NULL for the 'pending' return
> to ignore it).
>
> Then, we need to make sure that the mt_svc destructor will call
> dbus_pending_call_cancel() on mt_svc->pending, which will prevent us
> from reaching the ping_check() function.
>
> (Obviously, we need to NULLify the mt_svc->pending attribute once we
> reach the ping_check() function correctly).
>
> There's certainly no reason we can't ALSO include the band-aid of this
> patch, but I'd rather fix the underlying problem.
Thank you for the detailed explanation. A new patch is attached.