On Thu 09-09-21 15:45:45, Vlastimil Babka wrote:

On 9/9/21 13:54, Michal Hocko wrote:

quoted

On Wed 08-09-21 14:00:19, Mike Kravetz wrote:

quoted

On 9/7/21 1:50 AM, Hillf Danton wrote:

quoted

On Mon, 6 Sep 2021 16:40:28 +0200 Vlastimil Babka wrote:

And/or clamp reclaim retries for costly orders

	reclaim retries = MAX_RECLAIM_RETRIES - order;

to pull down the chance for stall as low as possible.

Thanks, and sorry for not replying quickly.  I only get back to this as
time allows.

We could clamp the number of compaction and reclaim retries in
__alloc_pages_slowpath as suggested.  However, I noticed that a single
reclaim call could take a bunch of time.  As a result, I instrumented
shrink_node to see what might be happening.  Here is some information
from a long stall.  Note that I only dump stats when jiffies > 100000.

[ 8136.874706] shrink_node: 507654 total jiffies,  3557110 tries
[ 8136.881130]              130596341 reclaimed, 32 nr_to_reclaim
[ 8136.887643]              compaction_suitable results:
[ 8136.893276]     idx COMPACT_SKIPPED, 3557109

Can you get a more detailed break down of where the time is spent. Also
How come the number of reclaimed pages is so excessive comparing to the
reclaim target? There is something fishy going on here.

I would say it's simply should_continue_reclaim() behaving similarly to
should_compact_retry(). We'll get compaction_suitable() returning
COMPACT_SKIPPED because the reclaimed pages have been immediately stolen,
and compaction indicates there's not enough base pages to begin with to form
a high-order pages. Since the stolen pages will appear on inactive lru, it
seems to be worth continuing reclaim to make enough free base pages for
compaction to no longer be skipped, because "inactive_lru_pages >
pages_for_compaction" is true.

So, both should_continue_reclaim() and should_compact_retry() are unable to
recognize that reclaimed pages are being stolen and limit the retries in
that case. The scenario seems to be uncommon, otherwise we'd be getting more
reports of that.

Thanks this sounds like a viable explanation. For some reason I have
thought that a single reclaim round can take that much time but reading
again it seems like a cumulative thing.

I do agree that we should consider the above scenario when bailing out.
It is simply unreasonable to reclaim so much memory without any forward
progress.

A more robust way to address this problem (which is not really new)
would be to privatize reclaimed pages in the direct reclaim context and
reuse them in the compaction so that it doesn't have to just
pro-actively reclaim to keep watermarks good enough. A normal low order
requests would benefit from that as well because the reclaim couldn't be
stolen from them as well.

Another approach would be to detect the situation and treat reclaim
success which immediatelly leads to compaction deferral due to
watermarks as a failure.

-- 
Michal Hocko
SUSE Labs