On Thu, 01 Apr 2010 09:17:49 -0400 Trond Myklebust <trond.myklebust@xxxxxxxxxx> wrote: > On Thu, 2010-04-01 at 08:46 -0400, Jeff Layton wrote: > > I'm working on an issue in an older kernel where we see occasional > > panics when trying to refresh credentials. Here's the bug in case > > anyone is interested: > > > > https://bugzilla.redhat.com/show_bug.cgi?id=572870 > > > > ...I think I understand the problem well enough now. The problem is > > pretty complex, but the issue is that some operations are done using > > credentials from a stateowner associated with a nfs_client, but using > > the rpc_clnt in nfs_server->client. The two can have different > > authtypes if there are a mix of mounts with different authtypes to the > > same server. This problem seems to have been fixed in mainline with the > > introduction of the auth_generic code. > > > > It leaves me wondering though...what exactly is the reason for having > > two rpc_clients per NFS mount? To clarify, I'm talking about these two, > > which seem to be somewhat redundant: > > > > nfs_server->client > > nfs_server->nfs_client->cl_rpcclient > > > > On mount, the nfs4_set_client calls nfs_get_client to search the list > > of nfs_client structs until it finds one that matches the address, port, > > etc of the NFS server. If one isn't found, the kernel creates one using > > whatever authtype was requested for the mount. > > > > Later, nfs_init_server_rpcclient looks at the rpc_clnt in the > > nfs_client and copies it. If the auth pseudoflavor doesn't match > > however, it creates a new rpc_auth for it. > > > > What exactly is the point of having two rpc_clnt's? Why not just get > > always use nfs_client->cl_rpcclient instead of nfs_server->client and > > simply have nfs_get_client filter by authtype? > > > > Look again at nfs_init_server_rpcclient(). The pseudoflavour is not the > only thing that is changed. We also change the soft flag and the timeout > properties of the server->client. > > The point is that users sometimes want to specify per-mountpoint > transport properties, and so we try to give them that possibility, while > at the same time sharing sockets/rdma connections. > > Cheers > Trond > Ok, I think I sort of understand now. The nfs_client holds some global state info that's common to all nfs mounts on the server so that info has to be shared between NFS mounts. The nfs_server struct holds stuff that's specific to a single superblock. That said, I still am not certain I see why we want to have an rpc_clnt that's shared between the mounts though clearly there's benefit to sharing the xprt between them. -- Jeff Layton <jlayton@xxxxxxxxxx> -- To unsubscribe from this list: send the line "unsubscribe linux-nfs" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
