On Thu, Jan 24, 2013 at 08:47:13AM -0500, Daniel J Walsh wrote:
-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1
(Resend to the correct list)
lxc-enter-namespace allows a process from outside a container to start a
process inside a container. One problem with the current code is the process
running within the container would run with the label of the process that
created it.
For example if the admin process is running as unconfined_t and executes the
following command
# virsh -c lxc:/// lxc-enter-namespace --nolabel dan -- /bin/ps -eZ
LABEL PID TTY TIME CMD
system_u:system_r:svirt_lxc_net_t:s0:c0.c1023 1 pts/0 00:00:00 systemd
system_u:system_r:svirt_lxc_net_t:s0:c0.c1023 3 pts/1 00:00:00 sh
system_u:system_r:svirt_lxc_net_t:s0:c0.c1023 24 ? 00:00:00 systemd-journal
system_u:system_r:svirt_lxc_net_t:s0:c0.c1023 29 ? 00:00:00 dhclient
staff_u:unconfined_r:unconfined_t:s0-s0:c0.c1023 47 ? 00:00:00 ps
Note the ps command is running as unconfined_t, After this patch,
virsh -c lxc:/// lxc-enter-namespace dan -- /bin/ps -eZ
LABEL PID TTY TIME CMD
system_u:system_r:svirt_lxc_net_t:s0:c0.c1023 1 pts/0 00:00:00 systemd
system_u:system_r:svirt_lxc_net_t:s0:c0.c1023 3 pts/1 00:00:00 sh
system_u:system_r:svirt_lxc_net_t:s0:c0.c1023 24 ? 00:00:00 systemd-journal
system_u:system_r:svirt_lxc_net_t:s0:c0.c1023 32 ? 00:00:00 dhclient
system_u:system_r:svirt_lxc_net_t:s0:c0.c1023 38 ? 00:00:00 ps
I also add a --nolabel command to virsh, which can go back to the original
behaviour.
virsh -c lxc:/// lxc-enter-namespace --nolabel dan -- /bin/ps -eZ
LABEL PID TTY TIME CMD
system_u:system_r:svirt_lxc_net_t:s0:c0.c1023 1 pts/0 00:00:00 systemd
system_u:system_r:svirt_lxc_net_t:s0:c0.c1023 3 pts/1 00:00:00 sh
system_u:system_r:svirt_lxc_net_t:s0:c0.c1023 24 ? 00:00:00 systemd-journal
system_u:system_r:svirt_lxc_net_t:s0:c0.c1023 32 ? 00:00:00 dhclient
staff_u:unconfined_r:unconfined_t:s0-s0:c0.c1023 37 ? 00:00:00 ps
One problem I had when I originally did the patch is
lxcDomainGetSecurityLabel was returning the incorrect label, I needed the
label of the initpid within the container not its parent process, so I
changed this function to match OpenNamespaces function.
One last strangeness, about half the time I run this, virsh hangs and never
returns.
Seems like
if (conn->driver->domainGetSecurityLabel(domain,
seclabel) == 0) {
Gets hung up. I have attached the strace in out1.gz
virDomainLxcEnterNamespace is run in the child process context - it is
forbidden to use the RPC connection once you've forked, becuase it
belongs to the parent process. The reason for the random hang is that
sometimes your child process reads the incoming I/O, sometimes the
parent process reads the incoming I/O.
This has to be implemented entirely in virsh, rather than in the
virDomainLxcEnterNamespace API. virsh should call virDomainGetSecurityLabel
before forking, and then call setexeccon inside the child process with
the data it obtained. That way no libvirt RPC calls take place in the
child.
@@ -7719,13 +7721,17 @@ cmdLxcEnterNamespace(vshControl *ctl, const
vshCmd *cmd)
*/
if (virFork(&pid) < 0)
goto cleanup;
+
+ if (!vshCommandOptBool(cmd, "nolabel"))
+ flags |= SECURITY_LABEL;
+
if (pid == 0) {
...at this point you're in the child process...
if (virDomainLxcEnterNamespace(dom,
nfdlist,
fdlist,
NULL,
NULL,
- 0) < 0)
+ flags) < 0)
_exit(255);
/* Fork a second time because entering the
Daniel
--
|:
http://berrange.com -o-
http://www.flickr.com/photos/dberrange/ :|
|:
http://libvirt.org -o-
http://virt-manager.org :|
|:
http://autobuild.org -o-
http://search.cpan.org/~danberr/ :|
|:
http://entangle-photo.org -o-
http://live.gnome.org/gtk-vnc :|