On Fri, Mar 05, 2021 at 05:06:17PM +0100, Michal Hocko wrote:
> On Wed 03-03-21 12:23:22, Minchan Kim wrote:
> > On Wed, Mar 03, 2021 at 01:49:36PM +0100, Michal Hocko wrote:
> > > On Tue 02-03-21 13:09:48, Minchan Kim wrote:
> > > > LRU pagevec holds refcount of pages until the pagevec are drained.
> > > > It could prevent migration since the refcount of the page is greater
> > > > than the expection in migration logic. To mitigate the issue,
> > > > callers of migrate_pages drains LRU pagevec via migrate_prep or
> > > > lru_add_drain_all before migrate_pages call.
> > > >
> > > > However, it's not enough because pages coming into pagevec after the
> > > > draining call still could stay at the pagevec so it could keep
> > > > preventing page migration. Since some callers of migrate_pages have
> > > > retrial logic with LRU draining, the page would migrate at next trail
> > > > but it is still fragile in that it doesn't close the fundamental race
> > > > between upcoming LRU pages into pagvec and migration so the migration
> > > > failure could cause contiguous memory allocation failure in the end.
> > > >
> > > > To close the race, this patch disables lru caches(i.e, pagevec)
> > > > during ongoing migration until migrate is done.
> > > >
> > > > Since it's really hard to reproduce, I measured how many times
> > > > migrate_pages retried with force mode below debug code.
> > > >
> > > > int migrate_pages(struct list_head *from, new_page_t get_new_page,
> > > > ..
> > > > ..
> > > >
> > > > if (rc && reason == MR_CONTIG_RANGE && pass > 2) {
> > > > printk(KERN_ERR, "pfn 0x%lx reason %d\n", page_to_pfn(page), rc);
> > > > dump_page(page, "fail to migrate");
> > > > }
> > > >
> > > > The test was repeating android apps launching with cma allocation
> > > > in background every five seconds. Total cma allocation count was
> > > > about 500 during the testing. With this patch, the dump_page count
> > > > was reduced from 400 to 30.
> > >
> > > Have you seen any improvement on the CMA allocation success rate?
> >
> > Unfortunately, the cma alloc failure rate with reasonable margin
> > of error is really hard to reproduce under real workload.
> > That's why I measured the soft metric instead of direct cma fail
> > under real workload(I don't want to make some adhoc artificial
> > benchmark and keep tunes system knobs until it could show
> > extremly exaggerated result to convice patch effect).
> >
> > Please say if you belive this work is pointless unless there is
> > stable data under reproducible scenario. I am happy to drop it.
>
> Well, I am not saying that this is pointless. In the end the resulting
> change is relatively small and it provides a useful functionality for
> other users (e.g. hotplug). That should be a sufficient justification.
Yub, that was my impression to worth upstreaming rather than keeping
downstream tree so made divergent.
>
> I was asking about CMA allocation success rate because that is a much
> more reasonable metric than how many times something has retried because
> retries can help to increase success rate and the patch doesn't really
> remove those. If you want to use number of retries as a metric then the
> average allocation latency would be more meaningful.
I believe the allocation latency would be pretty big and retrial part
would be marginal so doubt it's meaningful.
Let me send next revision with as-is descripion once I fix places
you pointed out.
Thanks.
