Re: [PATCH] filelock: don't do security checks on nfsd setlease calls

[Christian Brauner]

On Tue, Feb 06, 2024 at 06:59:49AM +1100, NeilBrown wrote:
> On Mon, 05 Feb 2024, Jeff Layton wrote:
> > Zdenek reported seeing some AVC denials due to nfsd trying to set
> > delegations:
> > 
> >     type=AVC msg=audit(09.11.2023 09:03:46.411:496) : avc:  denied  { lease } for  pid=5127 comm=rpc.nfsd capability=lease  scontext=system_u:system_r:nfsd_t:s0 tcontext=system_u:system_r:nfsd_t:s0 tclass=capability permissive=0
> > 
> > When setting delegations on behalf of nfsd, we don't want to do all of
> > the normal capabilty and LSM checks. nfsd is a kernel thread and runs
> > with CAP_LEASE set, so the uid checks end up being a no-op in most cases
> > anyway.
> > 
> > Some nfsd functions can end up running in normal process context when
> > tearing down the server. At that point, the CAP_LEASE check can fail and
> > cause the client to not tear down delegations when expected.
> > 
> > Also, the way the per-fs ->setlease handlers work today is a little
> > convoluted. The non-trivial ones are wrappers around generic_setlease,
> > so when they fail due to permission problems they usually they end up
> > doing a little extra work only to determine that they can't set the
> > lease anyway. It would be more efficient to do those checks earlier.
> > 
> > Transplant the permission checking from generic_setlease to
> > vfs_setlease, which will make the permission checking happen earlier on
> > filesystems that have a ->setlease operation. Add a new kernel_setlease
> > function that bypasses these checks, and switch nfsd to use that instead
> > of vfs_setlease.
> > 
> > There is one behavioral change here: prior this patch the
> > setlease_notifier would fire even if the lease attempt was going to fail
> > the security checks later. With this change, it doesn't fire until the
> > caller has passed them. I think this is a desirable change overall. nfsd
> > is the only user of the setlease_notifier and it doesn't benefit from
> > being notified about failed attempts.
> > 
> > Cc: Ondrej Mosnáček <omosnacek@xxxxxxxxx>
> > Reported-by: Zdenek Pytela <zpytela@xxxxxxxxxx>
> > Closes: https://bugzilla.redhat.com/show_bug.cgi?id=2248830
> > Signed-off-by: Jeff Layton <jlayton@xxxxxxxxxx>
> 
> Reviewed-by: NeilBrown <neilb@xxxxxxx>
> 
> It definitely nice to move all the security and sanity check early.
> This patch allows a minor clean-up in cifs which could possibly be
> included:
> diff --git a/fs/smb/client/cifsfs.c b/fs/smb/client/cifsfs.c
> index 2a4a4e3a8751..0f142d1ec64f 100644
> 
> --- a/fs/smb/client/cifsfs.c
> +++ b/fs/smb/client/cifsfs.c
> @@ -1094,9 +1094,6 @@ cifs_setlease(struct file *file, int arg, struct file_lock **lease, void **priv)
>  	struct inode *inode = file_inode(file);
>  	struct cifsFileInfo *cfile = file->private_data;
>  
> -	if (!(S_ISREG(inode->i_mode)))
> -		return -EINVAL;
> -
>  	/* Check if file is oplocked if this is request for new lease */
>  	if (arg == F_UNLCK ||
>  	    ((arg == F_RDLCK) && CIFS_CACHE_READ(CIFS_I(inode))) ||
> 
> 
> as ->setlease() is now never called for non-ISREG files.

I've added the following on top. I've made you author and added your
SoB. Please tell me if you have any problems with this: