Bug 749181

Description of problem:

We run RHEL 6.1 NFS servers and clients and mount home directories via NFS3.
This setup generally works but we sometimes get a 'permission denied' error,
for instance when running an 'ls' in ssh sessions that were left idle for
hours:

$ cd ~/a/b/c
$ ls 
1 2 3
... (wait a few hours)
$ ls
ls: cannot open directory .: Permission denied

The export is not the issue because the error immediately goes away
when the directory is accessed using its absolute path:

$ ls
ls: cannot open directory .: Permission denied
$ ls ~/a/b/c
1 2 3
$ ls
1 2 3

Turing off iptables and SELinux on server and clients does not help.
Mounting with all kinds of options like lookupcache=none or sync
does also have no impact.

The error can more or less forced to show up immediately by modifying the 
file caching behaviour of the server:

# echo 20000 > /proc/sys/vm/vfs_cache_pressure
# echo 3 > /proc/sys/vm/drop_caches
# sync 

Over the network I see this with wireshark when the 'ls' fails:

1	0.000000	x	y	NFS	V3 ACCESS Call, FH:0xd01522c9
2	0.000115	x	y	NFS	V3 ACCESS Reply (Call In 1) Error:NFS3ERR_ACCES

I did some initial bug hunting with a kernel built from 
kernel-2.6.32-131.12.1.el6.src.rpm that has some more dprintk()s spread
over the nfsd code, starting with nfsd_access() in fs/nfsd/nfs3proc.c.
Here are my results so far:

The NFS3ERR_ACCES comes from an EACCES detected by the following code in the
function nfsd_set_fh_dentry() in fs/nfsd/nfsfh.c:

   242          if (fileid_type == FILEID_ROOT)
   243                  dentry = dget(exp->ex_path.dentry);
   244          else {
   245                  dentry = exportfs_decode_fh(exp->ex_path.mnt, fid,
   246                                  data_left, fileid_type,
   247                                  nfsd_acceptable, exp);
   248          }
   249          if (dentry == NULL)
   250                  goto out;
   251          if (IS_ERR(dentry)) {
   252                  if (PTR_ERR(dentry) != -EINVAL)
   253                          error = nfserrno(PTR_ERR(dentry));
   254                  goto out;
   255          }

fileid_type is FILEID_ROOT when the error occurs and therefore the
entry exp->ex_path.dentry is used and seems to be an -EACCES error
code instead of a pointer to a valid dentry. I have not found out
what code sets exp->ex_path.dentry to -EACCES but i suspect that should
actually never be the case otherwise the dget() in line 243 is unsafe
because it assumes to operate on a valid pointer.

If someone of the nfs-utils or kernel nfs wizards could help to
debug this further and fix the issue would be greatly appreciated.

Version-Release number of selected component (if applicable):

kernel-2.6.32-131.12.1.el6.x86_64
nfs-utils-1.2.3-7.el6.x86_64

How reproducible:

As described above. We failed so far to reproduce it on a simple test server.
On our productive server with tens of exports and many clients it always shows
up.

Steps to Reproduce:
1. Configure a fairly large NFS3 server 
2. Mount a home directory via NFS3
3. Checne the current directory to some subdirectory
4. try to do an 'ls'
  
Actual results:

As described above.

Expected results:

There should be no errors.