Re: [PATCH] dma-buf: system_heap: avoid reclaim for order 4

[Jaewon Kim <jaewon31.kim@samsung.com>]

> On Tue, Jan 17, 2023 at 10:54 PM John Stultz <jstultz@google.com> wrote:
> >
> > On Tue, Jan 17, 2023 at 12:31 AM Jaewon Kim <jaewon31.kim@samsung.com> wrote:
> > > > Using order 4 pages would be helpful for many IOMMUs, but it could spend
> > > > quite much time in page allocation perspective.
> > > >
> > > > The order 4 allocation with __GFP_RECLAIM may spend much time in
> > > > reclaim and compation logic. __GFP_NORETRY also may affect. These cause
> > > > unpredictable delay.
> > > >
> > > > To get reasonable allocation speed from dma-buf system heap, use
> > > > HIGH_ORDER_GFP for order 4 to avoid reclaim.
> >
> > Thanks for sharing this!
> > The case where the allocation gets stuck behind reclaim under pressure
> > does sound undesirable, but I'd be a bit hesitant to tweak numbers
> > that have been used for a long while (going back to ion) without a bit
> > more data.
> >
> > It might be good to also better understand the tradeoff of potential
> > on-going impact to performance from using low order pages when the
> > buffer is used.  Do you have any details like or tests that you could
> > share to help ensure this won't impact other users?
> >
> > TJ: Do you have any additional thoughts on this?
> >
> I don't have any data on how often we hit reclaim for mid order
> allocations. That would be interesting to know. However the 70th
> percentile of system-wide buffer sizes while running the camera on my
> phone is still only 1 page, so it looks like this change would affect
> a subset of use-cases.
> 
> Wouldn't this change make it less likely to get an order 4 allocation
> (under memory pressure)? The commit message makes me think the goal of
> the change is to get more of them.

Hello John Stultz

I've been waiting for your next reply.

With my commit, we may gather less number of order 4 pages and fill the 
requested size with more number of order 0 pages. I think, howerver, stable
allocation speed is quite important so that corresponding user space
context can move on within a specific time.

Not only compaction but reclaim also, I think, would be invoked more if the 
__GFP_RECLAIM is added on order 4. I expect the reclaim could be decreased
if we move to order 0.

Thank you
Jaewon Kim

> 
> Actually with the low order being 0, I don't think __GFP_COMP makes
> sense in LOW_ORDER_GFP. But I guess that flag isn't harmful there.
> 
> > thanks
> > -john