[PING2!] RFR: 8230305: Cgroups v2: Container awareness

[Bob Vandette]

Cgroup V2 is about to go mainstream this year for popular distros such as Oracle Linux 8, Redhat Linux 8 and Fedora
so this fix it’s important to get into JDK 15 so we can start shaking out this support.

Please take a look and help get this change reviewed.

Thanks,
Bob Vandette



> On Nov 29, 2019, at 4:04 AM, Severin Gehwolf <sgehwolf at redhat.com> wrote:
> 
> On Fri, 2019-11-15 at 17:56 +0100, Severin Gehwolf wrote:
>> On Fri, 2019-11-08 at 15:21 +0100, Severin Gehwolf wrote:
>>> Hi Bob,
>>> 
>>> On Wed, 2019-11-06 at 10:47 +0100, Severin Gehwolf wrote:
>>>> On Tue, 2019-11-05 at 16:54 -0500, Bob Vandette wrote:
>>>>> Severin,
>>>>> 
>>>>> Thanks for taking on this cgroup v2 improvement.  
>>>>> 
>>>>> In general I like the implementation and the refactoring.  The CachedMetric class is nice.
>>>>> We can add any metric we want to cache in a more general way.
>>>>> 
>>>>> Is this the latest version of the webrev?
>>>>> 
>>>>> http://cr.openjdk.java.net/~sgehwolf/webrevs/cgroupsv2-hotspot/03/webrev/src/hotspot/os/linux/cgroupV2Subsystem_linux.cpp.html
>>>>> 
>>>>> It looks like you need to add the caching support for active_processor_count (JDK-8227006).
>>> [...]
>>>> I'll do a proper rebase ASAP.
>>> 
>>> Latest webrev:
>>> http://cr.openjdk.java.net/~sgehwolf/webrevs/cgroupsv2-hotspot/05/webrev/
>>> 
>>>>> I’m not sure it’s worth providing different strings for Unlimited versus Max or Scaled shares.
>>>>> I’d just try to be compatible with the cgroupv2 output so you don’t have to change the test.
>>>> 
>>>> OK. Will do.
>>> 
>>> Unfortunately, there is no way of NOT changing TestCPUAwareness.java as
>>> it expects CPU Shares to be written to the cgroup filesystem verbatim.
>>> That's no longer the case for cgroups v2 (at least for crun). Either
>>> way, most test changes are gone now.
>>> 
>>>>> I wonder if it’s worth trying to synthesize memory_max_usage_in_bytes() by keeping the highest
>>>>> value ever returned by the API.
>>>> 
>>>> Interesting idea. I'll ponder this a bit and get back to you.
>>> 
>>> This has been implemented. I'm not sure this is correct, though. It
>>> merely piggy-backs on calls to memory_usage_in_bytes() and keeps the
>>> high watermark value of that.
>>> 
>>> Testing passed on F31 with cgroups v2 controllers properly configured
>>> (podman) and hybrid (legacy hierarchy) with docker/podman.
>>> 
>>> Thoughts?
>> 
>> Ping?
> 
> Anyone willing to review this? It would be nice to make some progress.
> 
> Thanks,
> Severin
> 
>> Metrics work proposed for RFR here:
>> http://mail.openjdk.java.net/pipermail/core-libs-dev/2019-November/063464.html
>> 
>> Thanks,
>> Severin
>