On Tue, Feb 04, 2020 at 10:52:42AM -0800, Tyler Sanderson wrote:

On Tue, Feb 4, 2020 at 12:29 AM David Hildenbrand [off-list ref] wrote:

    On 03.02.20 21:32, Tyler Sanderson wrote:
    > There were apparently good reasons for moving away from OOM notifier
    > callback:
    > https://lkml.org/lkml/2018/7/12/314
    > https://lkml.org/lkml/2018/8/2/322
    >
    > In particular the OOM notifier is worse than the shrinker because:

    The issue is that DEFLATE_ON_OOM is under-specified.

    >
    >  1. It is last-resort, which means the system has already gone through
    >     heroics to prevent OOM. Those heroic reclaim efforts are expensive
    >     and impact application performance.

    That's *exactly* what "deflate on OOM" suggests.


It seems there are some use cases where "deflate on OOM" is desired and others
where "deflate on pressure" is desired.
This suggests adding a new feature bit "DEFLATE_ON_PRESSURE" that registers the
shrinker, and reverting DEFLATE_ON_OOM to use the OOM notifier callback.

This lets users configure the balloon for their use case.

Right. Let's not repeat past mistakes and let's try to specify this
new one properly though :)

    Assume you are using virtio-balloon for some weird way of memory
    hotunplug (which is what some people do) and you want to minimize the
    footprint of your guest. Then you really only want to give the guest
    more memory (or rather, let it take back memory automatically in this
    case) in case it really needs more memory. It should try to reclaim first.

    Under-specified.


    >  2. It lacks understanding of NUMA or other OOM constraints.

    Ballooning in general lacks the understanding of NUMA.

    >  3. It has a higher potential for bugs due to the subtlety of the
    >     callback context.

    While that is a valid point, it doesn't explain why existing
    functionality is changed.

    Personally, I think DEFLATE_ON_OOM should never have been introduced (at
    least not in this form).

I'm actually not sure how you would safely do memory overcommit without
DEFLATE_ON_OOM. So I think it unlocks a huge use case.
 



    --
    Thanks,

    David / dhildenb