On Wed, Jul 15, 2020 at 3:46 PM Stephen Smalley <stephen.smalley.work@xxxxxxxxx> wrote: > > On Wed, Jul 15, 2020 at 2:52 PM Mike Palmiotto > <mike.palmiotto@xxxxxxxxxxxxxxx> wrote: > > > > On Wed, Jul 15, 2020 at 1:10 PM Mike Palmiotto > > <mike.palmiotto@xxxxxxxxxxxxxxx> wrote: > > > > > > On Wed, Jul 15, 2020 at 12:28 PM Stephen Smalley > > > <stephen.smalley.work@xxxxxxxxx> wrote: > > > > > > > > On Wed, Jul 15, 2020 at 12:04 PM Mike Palmiotto > > > > <mike.palmiotto@xxxxxxxxxxxxxxx> wrote: > > > > > > > > > > On Tue, Jul 14, 2020 at 6:42 PM Mike Palmiotto > > > > > <mike.palmiotto@xxxxxxxxxxxxxxx> wrote: > > > > > > > > > > > > On Tue, Jul 14, 2020 at 5:35 PM Stephen Smalley > > > > > > <stephen.smalley.work@xxxxxxxxx> wrote: > > > > > > > > > > > > > > On Tue, Jul 14, 2020 at 5:20 PM Mike Palmiotto > > > > > > > <mike.palmiotto@xxxxxxxxxxxxxxx> wrote: > > > > > <snip> > > > > > > > > Do you think we should go ahead and completely swap in sestatus? I was > > > > > > > > just worried about breaking userspace object managers that are > > > > > > > > currently using netlink threads by default, for instance > > > > > > > > systemd-dbusd. I can spend some more time getting those to work with > > > > > > > > the status page if you think that's worthwhile. > > > > > > > > > > > > > > I'd be interested in understanding the impact of such a change on > > > > > > > existing userspace object managers. If we can switch the default > > > > > > > behavior for applications that are not explicitly using > > > > > > > avc_netlink_*() interfaces themselves (e.g. they are only using > > > > > > > selinux_check_access or avc_has_perm), then that would be beneficial > > > > > > > since I think it fully removes the need for a system call on the AVC > > > > > > > cache-hit code path. > > > > > > > > > > > > I'll have to do a bit of digging to see how this will affect dbus, et > > > > > > al. On first blush, it looks like they're just doing > > > > > > avc_netlink_check_nb() in their watch thread. Presumably other object > > > > > > managers are doing something similar so we would just need to make > > > > > > sure there is a netlink fd available. > > > > > > > > > > So it looks like dbus (at least) is directly checking for a netlink > > > > > socket[1], so just doing away with the avc_netlink_open call wouldn't > > > > > work out. My thinking is we have two options: > > > > > > > > > > 1) Add a new seopt to use sestatus and let userspace object managers opt-in > > > > > > > > > > 2) Call both selinux_status_open and avc_netlink_open in > > > > > avc_init_internal. This would satisfy the hard requirement for a > > > > > netlink socket. Then we can default to using sestatus in all of the > > > > > netlink processing paths, as you suggested in your last reply. We > > > > > could > > > > > > > > > > Option 2 seems better from the standpoint of using sestatus by > > > > > default, but it looks like recvfrom will never be called and the > > > > > messages will just sit in kernel memory. > > > > > > > > > > I'm inclined to go with option 1 at this point. > > > > > > > > > > [1] https://gitlab.freedesktop.org/dbus/dbus/-/blob/master/bus/selinux.c#L269 > > > > > > > > Couldn't we change avc_netlink_acquire_fd() to test whether fd hasn't > > > > yet been set (i.e. == -1) and call avc_netlink_open() in that case? > > > > Then dbus would still gets its netlink fd as expected but we wouldn't > > > > need to open it inside of avc_init? > > > > > > Much better. I'll send a new patch up shortly. > > > > Okay, maybe not that shortly. I tried your suggestion and dbus doesn't > > appear to receive the netlink messages. > > > > Initially I figured it had something to do with > > avc_create_thread(avc_netlink_loop) call still being in > > avc_init_internal, so I exposed the thread pointer in avc_internal.h > > and moved the call into avc_netlink_open, which seemed more > > appropriate. Still no dice. > > > > I'm probably doing something wrong -- I'll figure it out, but it's > > going to take me longer than I thought to track this down. > > Recommend writing a little test program to exercise it first without > bringing in all of the dbus baggage. Don't think Fedora is even using > dbusd anymore, just dbus-broker. > If you get stuck, feel free to send your patch and I can take a look too. Interestingly, the test program is working fine: https://github.com/mpalmi/selinux/tree/sestatus https://github.com/mpalmi/sestatus-test On a test run, I'm seeing both the status page and netlink socket notifications for load_polcy (twice for each case): ``` ./test opened avc successfully got netlink socket: 4 watching netlink socket for events avc: received policyload notice (seqno=3) policy reload notice received avc: received policyload notice (seqno=4) policy reload notice received ^C watching sestatus page for events avc: received policyload notice (seqno=5) policy reload notice received avc: received policyload notice (seqno=6) policy reload notice received ^Cclosing netlink socket: 4 destroying avc goodbye ``` Still seeing the MAC_POLICY_LOAD audit message, but none of the usual USER_AVC policyload notices. I'm not certain if I've covered enough, but this is where things currently stand. I can submit another patch to the list if that's preferable. I just figured I'd keep the noise down as much as possible until we figure out what's going on with dbus. Thanks for all the help -- Mike Palmiotto https://crunchydata.com
