From Bugzilla Helper:
User-Agent: Mozilla/4.77 [en] (X11; U; Linux 2.2.16-3 i686; Nav)

Description of problem:
The NFS client in the linux kernel (both 2.2.18 and 2.4.6) does the wrong
thing when an NFS server returns NFS3ERR_BAD_COOKIE on a readdir operation.

What it does:  At line 399 in kernel-2.4.6/linux/fs/nfs/dir.c
... 
		if (res == -EBADCOOKIE) {
			/* This means either end of directory */
			if (desc->entry->cookie != desc->target) {
				/* Or that the server has 'lost' a cookie */
				res = uncached_readdir(desc, dirent, filldir);
				if (res >= 0)
					continue;
			}
			res = 0;
			break;
		} else if (res < 0)
...

What this appears to do is to start the readdir() from the beginning again.

What it should do is to return an error code to the user who called
readdir().

See RFC1813 for a discussion of the behavior.

I saw the problem when using a network appliance NFS server.  I have one
nfs client creating and deleting files in a directory, and another nfs
client is trying to do readdir() to go through all of the files in the
directory.  I expected that the readdir() operation would see every file
that was neither created nor deleted, and it might see some of the created
files or some of the deleted files.   Instead, the client sees some files
twice, and some files get skipped.  The readdir should not be starting
over.  The client should have been notified that there was a problem, and
then the client could have started the readdir() operation all over,
throwing away its partial results.

How reproducible:
Always

Steps to Reproduce:
1. Use an network appliance NFS server and two linux clients.
2. Client A is doing a readdir() to get the entries in a directory
containing files named A*.
3. Client B is concurrently adding and removing subdirectories or files
from the directory.  The files being added and deleted are named B*.


	

Actual Results:  Client A will see the same A* file twice, and will see
some A* files skipped.  (It is OK if Client A sees B* files, or doesn't see
them.

Expected Results:  Client A should have seen each A* file exactly once.

Additional info: