Re: [PATCH 06/10] mm: vmscan: demote anon DRAM pages to PMEM node

[Yang Shi]

On 3/26/19 5:35 PM, Keith Busch wrote:
> On Mon, Mar 25, 2019 at 12:49:21PM -0700, Yang Shi wrote:
> > On 3/24/19 3:20 PM, Keith Busch wrote:
> > > How do these pages eventually get to swap when migration fails? Looks
> > > like that's skipped.
> > Yes, they will be just put back to LRU. Actually, I don't expect it would be
> > very often to have migration fail at this stage (but I have no test data to
> > support this hypothesis) since the pages have been isolated from LRU, so
> > other reclaim path should not find them anymore.
> >
> > If it is locked by someone else right before migration, it is likely
> > referenced again, so putting back to LRU sounds not bad.
> >
> > A potential improvement is to have sync migration for kswapd.
> Well, it's not that migration fails only if the page is recently
> referenced. Migration would fail if there isn't available memory in
> the migration node, so this implementation carries an expectation that
> migration nodes have higher free capacity than source nodes. And since
> your attempting THP's without ever splitting them, that also requires
> lower fragmentation for a successful migration.

Yes, it is possible. However, migrate_pages() already has logic to 
handle such case. If the target node has not enough space for migrating 
THP in a whole, it would split THP then retry with base pages.

Swapping THP has been optimized to swap in a whole too. It would try to 
add THP into swap cache in a whole, split THP if the attempt fails, then 
add base pages into swap cache.

So, I think we can leave this to migrate_pages() without splitting in 
advance all the time.

Thanks,
Yang

> Applications, however, may allocate and pin pages directly out of that
> migration node to the point it does not have so much free capacity or
> physical continuity, so we probably shouldn't assume it's the only way
> to reclaim pages.