MAC: improve handling of listening sockets

[tuexen]

Summary

so_peerlabel can only be used when the socket is not listening.

Test Plan

kldload mac_test
nc -l 1234

and then terminate nc using CTRL-C.

[markj]

Hum, this bug is quite old. :(

sys/kern/uipc_socket.c

4178

The right place to check this is in mac_getsockopt_peerlabel(). Without the socket lock (which is acquired in that function), it is possible for listen() and getsockopt(SO_PEERLABEL) to race with each other.

[tuexen]

Address Mark's comment.

[tuexen]

sys/kern/uipc_socket.c

4178

I totally agree. Fixed. But doesn't this comment apply also to the others usages of SOLISTENING() in sogetopt()?

[markj]

sys/kern/uipc_socket.c

4178

Most likely, yes. There are many SOLISTENING races, unfortunately. Some of the ones below are mostly harmless in that they'll just return a garbage value if the race is lost.

sys/security/mac/mac_socket.c

584

Shouldn't this be in mac_getsockopt_peerlabel()?

[tuexen]

Fix mac_getsockopt_label(), not mac_setsockopt_label() as pointed out by Mark.

[tuexen]

sys/kern/uipc_socket.c

4178

OK, I'll fix them. I decided to not call SOCK_LOCK(), because it wan't called in that places...

sys/security/mac/mac_socket.c

584

Yes, of course. Thanks for catching that.

[markj]

I think this is ok. The bug suggests that we should optionally run the regression test suite with mac_test loaded. I tried it locally and easily reproduced the original panic. Perhaps @olivier would want to try this in his test environment?