On Thu, May 05, 2011 at 12:55:46AM +0200, Samuel Thibault wrote: > I've noticed a port leak on buildds when starting a translator. The > backtrace of the port allocation is this: > > _ports_create_port_internal > ports_create_port_noinstall > diskfs_start_protid > diskfs_create_protid > diskfs_S_dir_lookup (libdiskfs/dir-lookup.c:260) > _Xdir_lookup > > What I understand from the diskfs_S_dir_lookup code with my poor > libports knowledge: > - diskfs_make_peropen (dnp, 0, dircred->po, &newpo) creates a peropen > structure "newpo". Its reference counter is initialized to 0. > - diskfs_create_portid(newpo, user, &newpi) creates a libport port with > receive right. Its > reference count is initialized to 1. newpo is made to reference it, > its reference counter gets increased to 1. The port right is added to > the diskfs_port_bucket portset. > - dirport = ports_get_send_right(newpi) adds a port send right to newpi, whose > reference counter gets increased to 2. > - ports_port_deref(newpi) drops a reference, the counter thus gets back > to 1. As I understand it, it's supposed to be the send one. > - newpi=0 forgets about the receive part of newpi, only the send part > dirport remains. > - dirport is given to fshelp_fetch_root(), which copies its send right > only. > - mach_port_deallocate (mach_task_self(), dirport); release a send > right. > > I'm unsure how the receive part of the kernel port in newpi port is > supposed to get deallocated in that scenario, I guess the send right > release at the end is supposed to trigger NOTIFY_NO_SENDERS, and > libports handles deallocation from that? The thing is that it doesn't > seem to happen, thus leaving a port leak here. > > Any idea?
Does this leak appear often ? It seems to be triggered by starting a new translator only. I only took a quick look but are you sure the dirport (which is copied as dotdot in fshelp_fetch_root()) isn't kept around in the "child" translator as long as it runs, as a reference to the translator managing .. (that is, ports[INIT_PORT_CWDIR]) ? In addition, is this a mach port leak or a libports leak ? -- Richard Braun
