RFR: Container Fixes (8219652, 8217766, 8212528)

David Holmes david.holmes at oracle.com
Fri Mar 15 00:44:17 UTC 2019

Hi Bob,

Sorry I took a look but can't really review it in detail as I don't know 
any of the cgroup details. Not sure who may ... perhaps Misha (cc'd).

On 15/03/2019 12:15 am, Bob Vandette wrote:
> Ping ...
> Please review these three fixes for Linux Docker/cgroup container support.
> http://cr.openjdk.java.net/~bobv/8217766/webrev.0


The copyright update should have been from:

* Copyright (c) 2018, Oracle


* Copyright (c) 2018, 2019, Oracle


> https://bugs.openjdk.java.net/browse/JDK-8219562 - Line of code in osContainer_linux.cpp#L102 appears unreachable
> This change corrects a rarely used hotspot code path making it compatible with the Java based Metrics.
> https://bugs.openjdk.java.net/browse/JDK-8212528 - Wrong cgroup subsystem being used for some CPU Container Metrics
> Most Linux distros provide symbolic links for cpuacct and cpu controller directories.  Docker on the Mac does not.
> This causes some of the cpu statistics to be unreported since the directory name was incorrect.
> https://bugs.openjdk.java.net/browse/JDK-8217766 - Container Support doesn't work for some Join Controllers combinations
> The cgroup identification -implemented by parsing /proc/self/mountinfo
> and /proc/self/cgroup- assumed each cgroup controller was mounted
> disjoint from the others (except for "cpu" and "cpuacct" controllers).
> Which means, we expected one single controller per mountinfo line.
> This matches the way most Linux distributions currently configure
> cgroupsv1 by default. Yet controllers can be grouped arbitrarily.
> For instance, using the JoinControllers systemd directive.
> One use case for that is to let Kubernetes' kubelet discover his own
> dedicated and reserved cgroup hierarchy. In that situation, the JVM
> fails to discover the expected cgroup controllers set, and, when running
> containerized, default to a suboptimal understanding of available resources.
> Supporting arbitrarily controllers groups per mountpoint actually
> allows for simpler and easier to follow code, as we don't need nested
> if/else for every controller.
> This fix also updates the Containers Metrics, to support joint controllers.
> Bob.

More information about the core-libs-dev mailing list