Re: [PATCH v2] fs: fsnotify: account fsnotify metadata to kmemcg

[Amir Goldstein]

On Mon, Feb 19, 2018 at 3:50 PM, Jan Kara <jack@xxxxxxx> wrote:
[...]
> For fanotify without FAN_UNLIMITED_QUEUE the situation is similar as for
> inotify - IMO low practical impact, apps should generally handle queue
> overflow so I don't see a need for any opt in (more accurate memcg charging
> takes precedense over possibly broken apps).
>
> For fanotify with FAN_UNLIMITED_QUEUE the situation is somewhat different -
> firstly there is a practical impact (memory consumption is not limited by
> anything else) and secondly there are higher chances of the application
> breaking (no queue overflow expected) and also that this breakage won't be
> completely harmless (e.g., the application participates in securing the
> system). I've been thinking about this "conflict of interests" for some
> time and currently I think that the best handling of this is that by
> default events for FAN_UNLIMITED_QUEUE groups will get allocated with
> GFP_NOFAIL - such groups can be created only by global CAP_SYS_ADMIN anyway
> so it is reasonably safe against misuse (and since the allocations are
> small it is in fact equivalent to current status quo, just more explicit).
> That way application won't see unexpected queue overflow. The process
> generating event may be looping in the allocator but that is the case
> currently as well. Also the memcg with the consumer of events will have
> higher chances of triggering oom-kill if events consume too much memory but
> I don't see how this is not a good thing by default - and if such reaction
> is not desirable, there's memcg's oom_control to tune the OOM behavior
> which has capabilities far beyond of what we could invent for fanotify...
>
> What do you think Amir?
>

If I followed all your reasoning correctly, you propose to change behavior to
always account events to group memcg and never fail event allocation,
without any change of API and without opting-in for new behavior?
I think it makes sense. I can't point at any expected breakage,
so overall, this would be a good change.

I just feel sorry about passing an opportunity to improve functionality.
The fact that fanotify does not have a way for defining the events queue
size is a deficiency IMO, one which I had to work around in the past.
I find that assigning group to memgc and configure memcg to desired
memory limit and getting Q_OVERFLOW on failure to allocate event
is going to be a proper way of addressing this deficiency.

But if you don't think we should bind these 2 things together,
I'll let Shakeel decide if he want to pursue the Q_OVERFLOW change
or not.

Thanks,
Amir.