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

[Amir Goldstein]

On Tue, Feb 20, 2018 at 2:43 PM, Jan Kara <jack@xxxxxxx> wrote:
> On Mon 19-02-18 21:07:28, Amir Goldstein wrote:
[...]
>>
>> 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.
>
> So if you don't pass FAN_Q_UNLIMITED, you will get queue with a fixed size
> and will get Q_OVERFLOW if that is exceeded. So is your concern that you'd
> like some other fixed limit? Larger one or smaller one and for what
> reason?
>

My use case was that with the default queue size, I would get Q_OVERFLOW
on bursty fs workloads, but using  FAN_Q_UNLIMITED and allowing to
consume entire system memory with events was not a desired alternative.
The actual queue size was not important, only allowing admin to tune the
system to bursty workloads without overflowing the event queue.

Something like FAN_Q_BESTEFFORT (i.e. Q_OVERFLOW on ENOMEM)
+ allowing to restrict event allocation to memcg, would allow admin to tune
the system to bursty workloads.

>> 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.
>
> So if there is still some uncovered use case for finer tuning of event
> queue length than setting or not setting FAN_Q_UNLIMITED (+ possibly
> putting the task to memcg to limit memory usage), we can talk about how to
> address that but at this point I don't see a strong reason to bind this to
> whether / how events are accounted to memcg...

Agreed.

>
> And we still need to make sure we properly do ENOMEM -> Q_OVERFLOW
> translation and use GFP_NOFAIL for FAN_Q_UNLIMITED groups before merging

Good. it wasn't clear to me from your summary if were going to require
ENOEM -> Q_OVERFLOW before merging this work. If you put it this way,
I think it makes sense to let user to choose between GFP_NOFAIL and
Q_OVERFLOW behavior when queue is not limited, for example by using new
fanotify_init flag FAN_Q_BESTEFFORT (or better name), but I have no problem
with postponing that for later.

Thanks,
Amir.