Re: [PATCH 3/4] mm: Reclaim small amounts of memory when an external fragmentation event occurs

[Mel Gorman]

On Thu, Nov 22, 2018 at 04:35:58PM +0100, Vlastimil Babka wrote:
> >> I think I don't understand this comment :( Do you want to avoid waking
> >> up kswapd from steal_suitable_fallback() (introduced above) for
> >> allocations without __GFP_KSWAPD_RECLAIM? But returning 0 here means
> >> actually allowing the allocation go through steal_suitable_fallback()?
> >> So should it return ALLOC_NOFRAGMENT below, or was the intent different?
> >>
> > 
> > I want to avoid waking kswapd in steal_suitable_fallback if waking
> > kswapd is not allowed.
> 
> OK, but then this 'if' should return ALLOC_NOFRAGMENT, not 0?
> But that will still not prevent waking kswapd for nodes where there's no
> ZONE_DMA32, or any node when get_page_from_freelist() retries without
> fallback.
> 
> > If the calling context does not allow it, it does
> > mean that fragmentation will be allowed to occur. I'm banking on it
> > being a relatively rare case but potentially it'll be problematic. The
> > main source of allocation requests that I expect to hit this are THP and
> > as they are already at pageblock_order, it has limited impact from a
> > fragmentation perspective -- particularly as pageblock_order stealing is
> > allowed even with ALLOC_NOFRAGMENT.
> 
> Yep, THP will skip the wakeup in steal_suitable_fallback() via 'goto
> single_page' above it. For other users of ~__GFP_KSWAPD_RECLAIM (are
> there any?) we could maybe just ignore and wakeup kswapd anyway, since
> avoiding fragmentation is more important? Or if we wanted to avoid
> wakeup reliably, then steal_suitable_fallback() would have to know and
> check gfp_flags I'm afraid, and that doesn't seem worth the trouble.

Indeed. While it works in some cases, it'll be full of holes and while
I could close them, it just turns into a subtle mess. I've prepared a
preparation path that encodes __GFP_KSWAPD_RECLAIM in alloc_flags and checks
based on that.  It's a lot cleaner overall, it's less of a mess than passing
gfp_flags all the way through for one test and there are fewer side-effects.

Thanks!

-- 
Mel Gorman
SUSE Labs