On Wed, Feb 13, 2013 at 01:29:35PM -0800, Eric W. Biederman wrote: > "J. Bruce Fields" <bfields@xxxxxxxxxxxx> writes: > > > On Wed, Feb 13, 2013 at 09:51:41AM -0800, Eric W. Biederman wrote: > >> From: "Eric W. Biederman" <ebiederm@xxxxxxxxxxxx> > >> > >> When a new rpc connection is established with an in-kernel server, the > >> traffic passes through svc_process_common, and svc_set_client and down > >> into svcauth_unix_set_client if it is of type RPC_AUTH_NULL or > >> RPC_AUTH_UNIX. > >> > >> svcauth_unix_set_client then looks at the uid of the credential we > >> have assigned to the incomming client and if we don't have the groups > >> already cached makes an upcall to get a list of groups that the client > >> can use. > >> > >> The upcall encodes send a rpc message to user space encoding the uid > >> of the user whose groups we want to know. Encode the kuid of the user > >> in the initial user namespace as nfs mounts can only happen today in > >> the initial user namespace. > > > > OK, I didn't know that. > > > > (Though I'm unclear how it should matter to the server what user > > namespace the client is in?) > > Perhaps I have the description a little scrambled. The short version > is that to start I only support the initial network namespace. > > If I haven't succeeded it is my intent to initially limit the servers > to the initial user namespace as well. I should see if I can figure > that out. > > >> When a reply to an upcall comes in convert interpret the uid and gid values > >> from the rpc pipe as uids and gids in the initial user namespace and convert > >> them into kuids and kgids before processing them further. > >> > >> When reading proc files listing the uid to gid list cache convert the > >> kuids and kgids from into uids and gids the initial user namespace. As we are > >> displaying server internal details it makes sense to display these values > >> from the servers perspective. > > > > All of these caches are already per-network-namespace. Ideally wouldn't > > we also like to associate a user namespace with each cache somehow? > > Ideally yes. I read through the caches enough to figure out where there > user space interfaces were, and to make certain we had conversions > to/from kuids and kgids. > > I haven't looked at what user namespace makes sense for these > caches. For this cache my first guess is that net->user_ns > is what we want as it will be shared by all users in network namespace I > presume. Oh, I didn't know about net->user_ns--so each network namespace is associated with a single user namespace, great, that simplifies life. Yes, that sounds exactly right. --b. _______________________________________________ Containers mailing list Containers@xxxxxxxxxxxxxxxxxxxxxxxxxx https://lists.linuxfoundation.org/mailman/listinfo/containers
