On Thu 25-08-16 15:34:54, David Rientjes wrote:

On Thu, 25 Aug 2016, Michal Hocko wrote:

quoted

quoted

I don't believe it has been an issue in the past for any archs that
don't use thp.

Well, fragmentation is a real problem and order-0 reclaim will be never
anywhere close to reliably provide higher order pages. Well, reclaiming
a lot of memory can increase the probability of a success but that
can quite often lead to over reclaim and long stalls. There are other
sources of high order requests than THP so this is not about THP at all
IMHO.

Would it be possible to list the high-order allocations you are concerned 
about other than thp that doesn't have fallback behavior like skbuff and 
slub allocations?  struct task_struct is an order-1 allocation and there 
may be order-1 slab bucket usage, but what is higher order or requires 
aggressive compaction to allocate?

kernel stacks (order-2 on many arches), some arches need higher order
pages for page table allocations (at least the upper level AFAIR).

Surely you're not suggesting that order-0 reclaim cannot form order-1
memory.

I haven't seen fragmentation that bad that order-1 would be completely
depleted so I wouldn't be all that worried about this. But order-2 can
get depleted as our last oom reports show.

I am concerned about kernels that require a small memory footprint and
cannot enable all of CONFIG_COMPACTION and CONFIG_MIGRATION.  Embedded
devices are not a negligible minority of kernels.

Fair enough. And nobody discourages them from disabling the
compaction. I would expect that kernels for those machines are
configured by people who know what they are doing. They have to be
careful about disabling many other things already and carefully weight
the missing functionality vs. code size savings. I also expect that
workloads on those machines are also careful to not require large
physically contiguous memory blocks very much. Otherwise they would have
problems described by the help text.

So I am not really sure what you are objecting to. I am not making
COMPACTION on unconditionally. I just want to make sure that regular
users do not think this is just a THP thing which is not true since the
lumpy reclaim is gone. On my laptop I have more than 40 slab caches
which have pagesperslab > 2.
-- 
Michal Hocko
SUSE Labs

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>