Re: [RFC 11/13] mm, compaction: add the ultimate direct compaction priority

[Vlastimil Babka]

On 05/13/2016 03:38 PM, Michal Hocko wrote:
> On Tue 10-05-16 09:36:01, Vlastimil Babka wrote:
> > During reclaim/compaction loop, it's desirable to get a final answer from
> > unsuccessful compaction so we can either fail the allocation or invoke the OOM
> > killer. However, heuristics such as deferred compaction or pageblock skip bits
> > can cause compaction to skip parts or whole zones and lead to premature OOM's,
> > failures or excessive reclaim/compaction retries.
> >
> > To remedy this, we introduce a new direct compaction priority called
> > COMPACT_PRIO_SYNC_FULL, which instructs direct compaction to:
> >
> > - ignore deferred compaction status for a zone
> > - ignore pageblock skip hints
> > - ignore cached scanner positions and scan the whole zone
> > - use MIGRATE_SYNC migration mode
>
> I do not think we can do MIGRATE_SYNC because fallback_migrate_page
> would trigger pageout and we are in the allocation path and so we
> could blow up the stack.

Ah, I thought it was just waiting for the writeout to complete, and you 
wanted to introduce another migrate mode to actually do the writeout. 
But looks like I misremembered.

> > The new priority should get eventually picked up by should_compact_retry() and
> > this should improve success rates for costly allocations using __GFP_RETRY,
>
> s@__GFP_RETRY@__GFP_REPEAT@

Ah thanks. Depending on the patch timing it might be __GFP_RETRY_HARD in 
the end, right :)

> > such as hugetlbfs allocations, and reduce some corner-case OOM's for non-costly
> > allocations.
>
> My testing has shown that even with the current implementation with
> deferring, skip hints and cached positions had (close to) 100% success
> rate even with close to OOM conditions.

Hmm, I thought you at one point said that ignoring skip hints was a 
large improvement, because the current resetting of them is just too fuzzy.

> I am wondering whether this strongest priority should be done only for
> !costly high order pages. But we probably want less special cases
> between costly and !costly orders.

Yeah, if somebody wants to retry hard, let him.

> > Signed-off-by: Vlastimil Babka <vbabka@xxxxxxx>
>
> Acked-by: Michal Hocko <mhocko@xxxxxxxx>
>
> > ---
> >   include/linux/compaction.h |  1 +
> >   mm/compaction.c            | 15 ++++++++++++---
> >   2 files changed, 13 insertions(+), 3 deletions(-)
> [...]
> > @@ -1631,7 +1639,8 @@ enum compact_result try_to_compact_pages(gfp_t gfp_mask, unsigned int order,
> >   								ac->nodemask) {
> >   		enum compact_result status;
> >
> > -		if (compaction_deferred(zone, order)) {
> > +		if (prio > COMPACT_PRIO_SYNC_FULL
> > +					&& compaction_deferred(zone, order)) {
> >   			rc = max_t(enum compact_result, COMPACT_DEFERRED, rc);
> >   			continue;
> >   		}
>
> Wouldn't it be better to pull the prio check into compaction_deferred
> directly? There are more callers and I am not really sure all of them
> would behave consistently.

I'll check, thanks.