[CRIU] Avoid dumping specific VMA ranges

[Pavel Emelyanov]

On 12/04/2016 08:20 PM, Raghavendra Prabhu wrote:
> Hi,
> 
> I have seen similar problem manifesting in the form such as
> https://bugs.launchpad.net/percona-server/+bug/1092645 wherein people
> don't want large memory pools in their coredumps. In this case, they
> mark it as MADV_DONTDUMP.  Is it good to  assume that whatever is not
> required in the coredump is not good for checkpoint images either.  If
> not MADV_DONTDUMP, /proc/$pid/coredump_filter is used.
> 
> So, wondering if something like
> https://gist.github.com/ronin13/59c6ffab786db7ef9e54db4fb29b4992 would
> be useful here.

I tend to agree, but ... from my perspective, pages that are not required
in a core-dump might still be required for an application to work. So your
patch is OK, but it required a CLI option to be turned on. Always silently 
dropping the MADV_DONTDUMP pages may cause problems.

-- Pavel

> 
> On Wed, Nov 30, 2016 at 11:17 AM, Pavel Emelyanov <xemul at virtuozzo.com> wrote:
>> On 11/30/2016 12:47 PM, Rodrigo Bruno wrote:
>>> Hi all,
>>>
>>> I am using CRIU to checkpoint JVMs with heaps > 10GB.
>>>
>>> I want to reduce the size of the checkpoint and I know that there are memory ranges that
>>> although being allocated and touched by the JVM, represent garbage and could be discarted
>>
>> Heh... This is something I thought of long ago. The idea by that time was
>> that we could ship a library that applications can link with, and, when
>> criu dumps them, it can detect the library is there and call some code from
>> victim task. So that application can do some preparatory work before being
>> checkpointed. But this would require patching JVM :\
>>
>>> when dumping the JVM (don't worry, I can ensure that the JVM is stopped while running this operation).
>>>
>>> Q1. Does CRIU support this?
>>
>> Nope :)
>>
>>> Q2. Without changing CRIU, is there any way to achieve this? Could I change a page table
>>> entry to force a page with garbage to be mapped to the zero page?
>>
>> Yes and no. There are no standard knobs for this, but (!) we have announced
>> compel recently. With it you can cook a parasite code that would call
>> MADVISE_DONTNEED on the desired ranges thus punching the holes in memory.
>>
>> -- Pavel
>>
>> _______________________________________________
>> CRIU mailing list
>> CRIU at openvz.org
>> https://lists.openvz.org/mailman/listinfo/criu
> .
>