Re: [patch] mm, oom: stop reclaiming if GFP_ATOMIC will start failing soon

[Vlastimil Babka <vbabka@suse.cz>]

On 4/28/20 11:48 PM, David Rientjes wrote:
> On Tue, 28 Apr 2020, Vlastimil Babka wrote:
> 
> Yes, order-0 reclaim capture is interesting since the issue being reported 
> here is userspace going out to lunch because it loops for an unbounded 
> amount of time trying to get above a watermark where it's allowed to 
> allocate and other consumers are depleting that resource.
> 
> We actually prefer to oom kill earlier rather than being put in a 
> perpetual state of aggressive reclaim that affects all allocators and the 
> unbounded nature of those allocations leads to very poor results for 
> everybody.

Sure. My vague impression is that your (and similar cloud companies) kind of
workloads are designed to maximize machine utilization, and overshooting and
killing something as a result is no big deal. Then you perhaps have more
probability of hitting this state, and on the other hand, even an occasional
premature oom kill is not a big deal?

My concers are workloads not designed in such a way, where premature oom kill
due to temporary higher reclaim activity together with burst of incoming network
packets will result in e.g. killing an important database. There, the tradeoff
looks different.

> I'm happy to scope this solely to an order-0 reclaim capture.  I'm not 
> sure if I'm clear on whether this has been worked on before and patches 
> existed in the past?

Andrew mentioned some. I don't recall any, so it might have been before my time.

> Somewhat related to what I described in the changelog: we lost the "page 
> allocation stalls" artifacts in the kernel log for 4.15.  The commit 
> description references an asynchronous mechanism for getting this 
> information; I don't know where this mechanism currently lives.
>