[Ocfs2-announce] Heads up regarding using nfs with ocfs2

Sunil Mushran Sunil.Mushran at oracle.com
Fri Jun 20 13:57:47 PDT 2008


All,

This is a heads up only for users exporting OCFS2 volumes as NFS mounts.
If not, please disregard this email.

Recently there was a bugzilla filed that mentioned observing file system
lockups when accessing OCFS2 exported volumes with FreeBSD NFS clients.
The lockups were not observed by him with Linux NFS clients.
http://oss.oracle.com/bugzilla/show_bug.cgi?id=977

Analysis showed that the lockup was being caused by a deadlock triggered
during the READDIRPLUS RPC call available in NFS3. The bugzilla has the
kernel stack of the deadlocked process.

Disabling READDIRPLUS in the NFS client stopped the deadlocks. The user
was then able to successfully run his regression on a mixture of Linux 2.6,
FreeBSD6 and FreeBSD7 NFS clients.

Now, even though the user originally reported that the problem did not
reproduce with Linux NFS clients, it was probably only so because the
Linux NFS client appears to be more judicious in its use of that RPC call
than FreeBSD. For e.g., it does not use it when accessing large
directories as was the case here.

What this means is that Linux NFS clients can also experience the same
process deadlock though probably not as frequently as with other NFS
clients.

If you are using OCFS2 is such an environment and have experienced
such lockups, look into disabling READDIRPLUS in the NFS clients.
Refer to the NFS man page in your distribution for instructions on
the same. On Linux, it typically is done by mounting with the
mount option "nordirplus".

Also, it would be helpful if all users using OCFS2 with NFS can let
us know whether they have or not experienced the issue. Please include
the number of nodes in the cluster, kernel version, OCFS2 version,
the distribution running the NFS client and whether it supports
disabling READDIRPLUS or not.

While the actual problem will eventually likely be addressed in NFS,
we are trying to decide whether we should workaround the issue internally
in OCFS2. The above information will help us decide the same.

Lastly, we would like to thank Sérgio Surkamp for helping us to
resolve this issue.

Thank you
OCFS2 Team



More information about the Ocfs2-announce mailing list