Red Hat Bugzilla – Bug 977250
running dbench results in leaked fds leading to OOM killer killing glusterfsd.
Last modified: 2013-07-17 07:23:58 EDT
+++ This bug was initially created as a clone of Bug #976800 +++
Description of problem:
Running dbench on a distributed replicate volume causes leaked fds on the server thereby causing the OOM killer to kill the brick process.
Output of dmesg on server:
VFS: file-max limit 188568 reached
Out of memory: Kill process 12235 (glusterfsd) score 215 or sacrifice child
Killed process 12235, UID 0, (glusterfsd) total-vm:3138856kB, anon-rss:466728kB, file-rss:1028kB
glusterfsd invoked oom-killer: gfp_mask=0x201da, order=0, oom_adj=0, oom_score_adj=0
glusterfsd cpuset=/ mems_allowed=0
Pid: 12333, comm: glusterfsd Not tainted 2.6.32-358.6.2.el6.x86_64 #1
Steps to Reproduce:
1. create a 2x2 distributed replicate volume and FUSE mount it
2. On the mount point,run "dbench -s -F -S -x --one-byte-write-fix --stat-check 10"
3.Kill dbench after running for about 3 minutes.
4. On the server, do
ls -l /proc/pid_of_brick(s)/fd|grep deleted
We can still see open/unlinked fds even though dbench was killed. Also if dbench is run till completion, we can observe some of the bricks getting killed by the oom killer (ps aux |grep glusterfsd)
Once dbench is killed, the brick processes must not have open/unlinked fds.
Bisected the leak to the following commit ID on upstream:
* 8909c28 - cluster/afr: fsync() guarantees POST-OP completion
--- Additional comment from Anand Avati on 2013-06-24 03:03:34 EDT ---
REVIEW: http://review.gluster.org/5248 (cluster/afr: Fix fd/memory leak on fsync) posted (#1) for review on master by Pranith Kumar Karampuri (email@example.com)
Tested with the above scenario from multiple runs. No fd leak seen.
Since the problem described in this bug report should be
resolved in a recent advisory, it has been closed with a
resolution of ERRATA.
For information on the advisory, and where to find the updated
files, follow the link below.
If the solution does not work for you, open a new bug report.