Fix nfs snapdir automount

The current implementation for allowing nfs to access snapdir is very buggy.
It uses a special fh for snapdirs, such that the next time nfsd does
fh_to_dentry, it actually returns the root inode inside the snapshot. So nfsd
never knows it cross a mountpoint.

The problem is that nfsd will not hold a reference on the vfsmount of the
snapshot. This cause auto unmounter to unmount the snapshot even though nfs is
still holding dentries in it.

To fix this, we return the inode for the snapdirs themselves. However, we also
trigger automount upon fh_to_dentry, and return ESTALE so nfsd will revalidate
and see the mountpoint and do crossmnt.

Because nfsd will now be aware that these are different filesystems users
must add crossmnt to their export options to access snapshot directories.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Chunwei Chen <david.chen@osnexus.com>
Closes #3794
Closes #4716
Closes #5810 
Closes #5833

man/man8/zfs.8

@@ -1340,7 +1340,7 @@ Controls whether the file system is shared via \fBNFS\fR, and what options are u
 .sp
 .in +4
 .nf
-/usr/sbin/exportfs -i -o sec=sys,rw,no_subtree_check,no_root_squash,mountpoint *:<mountpoint of dataset>
+/usr/sbin/exportfs -i -o sec=sys,rw,crossmnt,no_subtree_check,no_root_squash,mountpoint *:<mountpoint of dataset>
 .fi
 .in -4
 .sp
@@ -3746,6 +3746,10 @@ The following commands show how to set \fBsharenfs\fR property options to enable
 .LP
 If you are using \fBDNS\fR for host name resolution, specify the fully qualified hostname.
 
+.sp
+.LP
+If you want to access snapdir through NFS, be sure to add \fBcrossmnt\fR to the options.
+
 .LP
 \fBExample 17 \fRDelegating ZFS Administration Permissions on a ZFS Dataset
 .sp