NFS: Fix security flavor negotiation with legacy binary mounts

Darrick J. Wong <darrick.wong@oracle.com> reports:
> I have a kvm-based testing setup that netboots VMs over NFS, the
> client end of which seems to have broken somehow in 3.10-rc1.  The
> server's exports file looks like this:
>
> /storage/mtr/x64	192.168.122.0/24(ro,sync,no_root_squash,no_subtree_check)
>
> On the client end (inside the VM), the initrd runs the following
> command to try to mount the rootfs over NFS:
>
> # mount -o nolock -o ro -o retrans=10 192.168.122.1:/storage/mtr/x64/ /root
>
> (Note: This is the busybox mount command.)
>
> The mount fails with -EINVAL.

Commit 4580a92 "NFS: Use server-recommended security flavor by
default (NFSv3)" introduced a behavior regression for NFS mounts
done via a legacy binary mount(2) call.

Ensure that a default security flavor is specified for legacy binary
mount requests, since they do not invoke nfs_select_flavor() in the
kernel.

Busybox uses klibc's nfsmount command, which performs NFS mounts
using the legacy binary mount data format.  /sbin/mount.nfs is not
affected by this regression.

Reported-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Tested-by: Darrick J. Wong <darrick.wong@oracle.com>
Acked-by: Weston Andros Adamson <dros@netapp.com>
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>

fs/nfs/super.c

@@ -1942,6 +1942,7 @@ static int nfs23_validate_mount_data(void *options,
 		args->namlen		= data->namlen;
 		args->bsize		= data->bsize;
 
+		args->auth_flavors[0] = RPC_AUTH_UNIX;
 		if (data->flags & NFS_MOUNT_SECFLAVOUR)
 			args->auth_flavors[0] = data->pseudoflavor;
 		if (!args->nfs_server.hostname)
@@ -2637,6 +2638,7 @@ static int nfs4_validate_mount_data(void *options,
 			goto out_no_address;
 		args->nfs_server.port = ntohs(((struct sockaddr_in *)sap)->sin_port);
 
+		args->auth_flavors[0] = RPC_AUTH_UNIX;
 		if (data->auth_flavourlen) {
 			if (data->auth_flavourlen > 1)
 				goto out_inval_auth;