Re: [PATCH] nfsd: enable UMH client tracker in container

["J. Bruce Fields" <bfields@fieldses.org>]

On Fri, Apr 19, 2013 at 07:01:50AM -0700, Jeff Layton wrote:
> On Tue, 16 Apr 2013 15:17:44 +0400
> Stanislav Kinsbursky <skinsbursky@parallels.com> wrote:
> 
> > This patch adds support for UserModeHelper tracker in a container.
> > The reason for this is that the only containerised tracker ("nfsdcld") is
> > going to be removed in 3.10 kernel, thus at least one more tracker have to be
> > containerised to replace the deprecated one.
> > UMH tracker looks more preferable comparing to legacy since it's the latest
> > one.
> > To make UMH tracker work in a container, we have to make sure, that it
> > executes right binary (i.e., this binary have to be taken from the container
> > environment).
> > But, UMH is a kernel thread, which works in global root environment by design
> > (kernel thread's root is inherited from kthreadd, which in turn inherited it's
> > root from global init). So, the root have to be swapped to the container's
> > one before binary execution.
> > 
> > This patch passes "init" callback and private "data" to UMH interface, which
> > are used to swap root for spawned kernel thread.
> > 
> > Note: container's root can be stored on stack, because UMH calls are
> > synchronous.
> > 
> > Signed-off-by: Stanislav Kinsbursky <skinsbursky@parallels.com>
> > ---
> >  fs/nfsd/nfs4recover.c |   29 ++++++++++++++++++++++-------
> >  1 files changed, 22 insertions(+), 7 deletions(-)
> > 
> > diff --git a/fs/nfsd/nfs4recover.c b/fs/nfsd/nfs4recover.c
> > index 899ca26..15f8de6 100644
> > --- a/fs/nfsd/nfs4recover.c
> > +++ b/fs/nfsd/nfs4recover.c
> > @@ -38,6 +38,7 @@
> >  #include <linux/crypto.h>
> >  #include <linux/sched.h>
> >  #include <linux/fs.h>
> > +#include <linux/fs_struct.h>
> >  #include <linux/module.h>
> >  #include <net/net_namespace.h>
> >  #include <linux/sunrpc/rpc_pipe_fs.h>
> > @@ -1122,12 +1123,28 @@ nfsd4_cltrack_legacy_recdir(const struct xdr_netobj *name)
> >  	return result;
> >  }
> >  
> > +static int nfsd_swap_root(struct subprocess_info *info, struct cred *new)
> > +{
> > +	struct path *root = info->data;
> > +	struct fs_struct *fs = current->fs;
> > +	struct path current_root;
> > +
> > +	spin_lock(&fs->lock);
> > +	current_root = fs->root;
> > +	fs->root = *root;
> > +	spin_unlock(&fs->lock);
> > +	if (current_root.dentry)
> > +		path_put(&current_root);
> > +	return 0;
> > +}
> > +
> >  static int
> >  nfsd4_umh_cltrack_upcall(char *cmd, char *arg, char *legacy)
> >  {
> >  	char *envp[2];
> >  	char *argv[4];
> >  	int ret;
> > +	struct path root;
> >  
> >  	if (unlikely(!cltrack_prog[0])) {
> >  		dprintk("%s: cltrack_prog is disabled\n", __func__);
> > @@ -1146,7 +1163,11 @@ nfsd4_umh_cltrack_upcall(char *cmd, char *arg, char *legacy)
> >  	argv[2] = arg;
> >  	argv[3] = NULL;
> >  
> > -	ret = call_usermodehelper(argv[0], argv, envp, UMH_WAIT_PROC);
> > +	get_fs_root(current->fs, &root);
> > +
> > +	ret = call_usermodehelper_fns(argv[0], argv, envp, UMH_WAIT_PROC,
> > +				       nfsd_swap_root, NULL, &root);
> > +
> >  	/*
> >  	 * Disable the upcall mechanism if we're getting an ENOENT or EACCES
> >  	 * error. The admin can re-enable it on the fly by using sysfs
> > @@ -1185,12 +1206,6 @@ bin_to_hex_dup(const unsigned char *src, int srclen)
> >  static int
> >  nfsd4_umh_cltrack_init(struct net __attribute__((unused)) *net)
> >  {
> > -	/* XXX: The usermode helper s not working in container yet. */
> > -	if (net != &init_net) {
> > -		WARN(1, KERN_ERR "NFSD: attempt to initialize umh client "
> > -			"tracking in a container!\n");
> > -		return -EINVAL;
> > -	}
> >  	return nfsd4_umh_cltrack_upcall("init", NULL, NULL);
> >  }
> >  
> > 
> 
> I think this looks correct and like it'll work, but I wonder whether
> we'd be best served by making this part of the UMH code itself?
> 
> IOW, add some fields to the struct subprocess_info to hold the new
> root, and then do what you're doing in nfsd_swap_root before calling
> the "init" function?
> 
> I imagine we'll eventually need do something similar for at least some
> of the callers of call_usermodehelper so it would make sense to me to
> not replicate copies of nfsd_swap_root all over the place.

Agreed.

--b.