Bug 135688

Description of problem:
If a file is modified on an NFS server which is exporting that
directory to NFS clients, the NFS clients may receive numerous ESTALE
errors when attempting to access that file using cached file handles.
 The number of ESTALES recevied and reported to user space appears
related to how many directories above that file may have been removed

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

How reproducible:
always

Steps to Reproduce:
1) NFS mount a share from a client to a server.  Lets say we mount
server:/tmp on the client at /mnt/tmp. You may want to mount it with a
large timeo value to ensure that you get ESTALE on open from the cache
file handles

2) on the server in /tmp, create the following file:
a/b/c

3) tar the contents of directory a recusively, preserving timestamps
on the file.  I use the command:
 tar -c --file ./test.tar a
run from the /tmp directory on the server

4) on the client run:
cd /mnt/tmp
cat a/b/c
This should get the file handle for c in the NFS cache.

5) on the server run:
rm -rf a
tar -x -v --file ./test.tar
This will recreate the file tree on the server, and make the cached
file handles on the client stale

6) on the client run:
cat a/b/c
If you are running without the patch that I posted you will of course
get an ESTALE error.  If you tcpdumped the connection, you will find 3
ESTALE errors returned, in two lookup responses, and in 1 getattr
response.


Actual results:
ESTALES are returned to the calling userspace application

Expected results:
arguably, since ESTALE is largely a transient error on network file
systems, the operation should be silently retried.

Additional info: