Howdy,
I just wanted to take a moment to thank several people, working quite a few hours today putting together this advisory:
Avati - for committing the fixes to the source tree in the first place
Vikas - for helping me run down the github commits and issue tracker notes
AB - for explaining to me in some detail how the race conditions happened
...and also for correcting my initial mistakes :)
These types of advisories are new for us, and we're open to your suggestions on the best way to communicate this information going forward.
Please do not hesitate to contact me if you should need assistance.
Thanks,
John Mark Walker
Gluster Community Guy
________________________________
From: gluster-users-bounces at gluster.org [gluster-users-bounces at gluster.org] on behalf of Craig Carl [craig at gluster.com]
Sent: Friday, June 24, 2011 3:50 PM
To: Gluster Users
Subject: Re: UPDATED: Enomaly users, please read -
All -
A much more detailed update from engineering is below. The range of affected versions is bigger that I thought, I apologize for not getting that right earlier. Please let me know if you have any questions.
VERSIONS AFFECTED: All Current GlusterFS releases up to 3.1.5 and 3.2.1
SEVERITY: For Enomaly users, a conflict results in denial of service but no loss of data. We have not observed the same race conditions in non-Enomaly environments.
CAUSE: Gluster and Enomaly incompatibility
There is an incompatibility between GlusterFS and Enomaly and how they perform directory operations. Enomaly's agents monitor node failures and migrate VMs automatically. These distributed agents communicate with each other by performing directory operations, such as mkdir, for the purpose of inter-node communication. These directory operations triggers a race condition GlusterFS, locking up a storage node. Enomaly agents get confused and propagate the error across the site even for a single node failure.
There was a race condition related to changing GFID's that was fixed in 3.1.5 - http://blog.gluster.com/2011/06/glusterfs-3-1-5-now-available/ - this is a partial fix for the behavior described. Race conditions can occur again.
After fixing the initial outage, if any node fails, you will see the issue again. Upgrading GlusterFS to 3.1.5 and restarting GlusterFS and Enomaly is a temporary fix. A permanent solution requires the Gluster 3.1.6 or 3.2.2 release (coming soon, see "Solution" below).
Other possible race conditions are fixed in the current source tree, subject to further testing.
SOLUTION:
This issue has been fixed in our source repository (https://github.com/gluster) and will be released soon with 3.1.6 and 3.2.2. If you'd like to help test the current fixes, please contact us before you do anything foolish (read: use in production). Users who test these patches in their non-critical, development environments and send us feedback will each get a Gluster t-shirt, maybe even a hat!!
We will send out another alert as soon as both releases are GA.
--
Thanks,
Craig Carl
Senior Systems Engineer | Gluster
408-829-9953<callto:+1408-829-9953> | San Francisco, CA
http://www.gluster.com/gluster-for-aws/
________________________________
Craig Carl<mailto:craig at gluster.com>
June 24, 2011 1:57 PM
All -
Gluster has identified a serious issue that affects anyone hosting VM images for the Enomaly Elastic Cloud Platform with Gluster. This issue is limited to Gluster versions =<3.1.4. We strongly encourage anyone using Enomaly and Gluster to upgrade to Gluster version 3.1.5 or higher.
What causes the failure -
Use a distribute-replicate volume.
Mount using either NFS or the GlusterFS.
Fail one of replica nodes.
** Production is unaffected at this point.
Restart the failed node.
** All the virtual machines fail.
** The ecpagent service on each hypervisor will constantly restart.
Root cause -
Enomaly uses a locking mechanism in addition to and above the standard POSIX locking to make sure that a VM never starts on two servers at the same time. When a Enomaly sever starts a VM it writes a file (<randomstuff.tmp> to the directory. The combination of self-heal and a race between the ecpagents on the hypervisors results in the VMs failing to start. No data is lost or damaged.
Again, this issue is specific to Enomaly. Enomaly users should immediatly upgrade to a version of Gluster =>3.1.5.
http://download.gluster.com/pub/gluster/glusterfs/
_______________________________________________
Gluster-users mailing list
Gluster-users at gluster.org<mailto:Gluster-users at gluster.org>
http://gluster.org/cgi-bin/mailman/listinfo/gluster-users
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://gluster.org/pipermail/gluster-users/attachments/20110625/7f527779/attachment-0001.htm>
