On Mon, Jan 26, 2009 at 04:19:14PM +0100, Konrad Eriksson1 wrote: There are lots of scenarios to consider - Direct local API usage. You have the PID, UID & GID of the process - Local usage via libvirtd + UNIX sockets. You can get the PID, UID & GID of the client end using the SCM_CREDENTIALS message. (see man 7 unix) - Remote usage via libvirtd + TCP sockets. Depending on the security & encryption settings you may have a SASL username, or a x509 certificate CNAME, or both, or neither. - Local usage via libvirtd + UNIX sockets + libvirt-qpid. The PID, UID & GID of the client end aren't particularly usage, since libvirt-qpid is just a demon running as root, which accepts calls on behalf of many remote apps. Does QPid provide any identifying info about the entity which put the message on the queue. - Remote usage with IPSec ? So there are multiple identifying credentials, in multiple formats, and need some way to associate this information with a connection. Applying RBAC to local (non-Daemon) API usage has clear limitations - if the user running virsh (or equiv) has direct access to the system, then they could trivially just replace the real virsh with their own virsh without RBAC. So RBAC usage in the non-Daemon context is only useful if the user does not have direct access to the ssytem (eg, virsh is being invoked on their behalf by another tool, or a constrained environment where its guarenteed they can't provide their own libraries/binaries. The key question is do we need to pass the client/callers identity at the time we create the connection object, or is it sufficient to provide it after the fact with a call like virConnectAddSecParam(VIR_SECPARAM_UNIX_UID, getuid()); virConnectAddSecParam(VIR_SECPARAM_UNIX_GID, getgid()); virConnectAddSecParam(VIR_SECPARAM_USERNAME, saslUsername); virConnectAddSecParam(VIR_SECPARAM_X509_CNAME, saslUsername); Or do we need to provide this info via some form of callback mechanism, perhaps via the exiting virConnectCredential struct, which is curently not used on the server end of remote connections. Daniel -- |: Red Hat, Engineering, London -o- http://people.redhat.com/berrange/ :| |: http://libvirt.org -o- http://virt-manager.org -o- http://ovirt.org :| |: http://autobuild.org -o- http://search.cpan.org/~danberr/ :| |: GnuPG: 7D3B9505 -o- F3C9 553F A1DA 4AC2 5648 23C1 B3DF F742 7D3B 9505 :|

Hi, it seems that we share the interest. Konrad Eriksson1 wrote: Sounds reasonable. Injecting AC without embedding hooks at API entries is pretty cool. Do you have any implementation? Yes, our AC takes effect only when it is invoked locally because uid is unavailable at a remote site. I agree in that some user-auth mechanisms are necessary, but now I concentrate on AC-module itself assuming that user-auth is finished. Daniel gave a concrete idea on user authentication and I find it fine. I have some coarse ideas where policies should be given as a SELinux's policy. It extends libvirt, like sVirt does, so that it calls SELinux API to make an authenticated user move to a qualified role (domain). Current sVirt is something different from AC itself though it employs SELinux. sVirt constrains qemu processes to use just allowed objects, while AC-module is aimed to constrain users to call just allowed sub commands on allowd VMs. libvirt is secure enough to me. So, I don't think it is mandatory that AC be realized by using SELinux. The main goal of AC in this context is to prevent differently authorized user from accidentally invoking unallowed operations. This seems a bit different from keeping the whole server untainted. So, proprietary implementations don't likely cause a serious security threat. Of cause, I understand that SELinux appears an attractive option. Regards, Hiroya