From Bugzilla Helper:
User-Agent: Mozilla/5.0 (X11; U; Linux i686; en-US; rv:1.7.2)
Gecko/20040803

Description of problem:
The NFS direct I/O patch applied to RHEL 3.0 update 3 has a minor bug
that was found during testing with fsx-odirect.  The fix is in
function nfs_file_direct_IO in fs/nfs/direct.c:

        retval = filemap_fdatasync(mapping);
        if (retval == 0)
-               retval = fsync_inode_data_buffers(inode);
+               retval = nfs_wb_all(inode);
        if (retval == 0)
                retval = filemap_fdatawait(mapping); 

fsync_inode_data_buffers() comes from the normal direct I/O path, but
since NFS doesn't use inode data buffers, it is a no-op in the NFS
direct I/O path.  nfs_wb_all is the correct function to call here.

Note that there are no real applications I'm aware of that access a
file via direct I/O and writable map simultaneously.

Version-Release number of selected component (if applicable):
kernel-2.4.21-20.EL

How reproducible:
Always

Steps to Reproduce:
1. mmap a file in NFS, and open it with O_DIRECT
2. dirty the map with just a few bytes, then read from the O_DIRECT
file descriptor in quick succession

Actual Results:  I did this with fsx-odirect.  It dirtied 92KB of the
map, and the NFS client flushed only 64KB of the dirty pages before
the direct read operation. 

Expected Results:  All 92KB should have been written to the server
before the direct read operation.

Additional info:

The problem occurs because nfs_strategy only pushes writes out in
"wsize" chunks.  nfs_wb_all pushes everything no matter how many NFS
write operations it takes.