Re: [PATCH -v2 0/9] migrate_pages(): batch TLB flushing

["Huang, Ying" <ying.huang@intel.com>]

Mike Kravetz <mike.kravetz@oracle.com> writes:

> On 01/12/23 15:17, Huang, Ying wrote:
>> Mike Kravetz <mike.kravetz@oracle.com> writes:
>> 
>> > On 01/12/23 08:09, Huang, Ying wrote:
>> >> Hi, Mike,
>> >> 
>> >> Mike Kravetz <mike.kravetz@oracle.com> writes:
>> >> 
>> >> > On 01/10/23 17:53, Mike Kravetz wrote:
>> >> >> Just saw the following easily reproducible issue on next-20230110.  Have not
>> >> >> verified it is related to/caused by this series, but it looks suspicious.
>> >> >
>> >> > Verified this is caused by the series,
>> >> >
>> >> > 734cbddcfe72 migrate_pages: organize stats with struct migrate_pages_stats
>> >> > to
>> >> > 323b933ba062 migrate_pages: batch flushing TLB
>> >> >
>> >> > in linux-next.
>> >> 
>> >> Thanks for reporting.
>> >> 
>> >> I tried this yesterday (next-20230111), but failed to reproduce it.  Can
>> >> you share your kernel config?  Is there any other setup needed?
>> >
>> > Config file is attached.
>> 
>> Thanks!
>> 
>> > Are you writing a REALLY big value to nr_hugepages?  By REALLY big I
>> > mean a value that is impossible to fulfill.  This will result in
>> > successful hugetlb allocations until __alloc_pages starts to fail.  At
>> > this point we will be stressing compaction/migration trying to find more
>> > contiguous pages.
>> >
>> > Not sure if it matters, but I am running on a 2 node VM.  The 2 nodes
>> > may be important as the hugetlb allocation code will try a little harder
>> > alternating between nodes that may perhaps stress compaction/migration
>> > more.
>> 
>> Tried again on a 2-node machine.  Still cannot reproduce it.
>> 
>> >> BTW: can you bisect to one specific commit which causes the bug in the
>> >> series?
>> >
>> > I should have some time to isolate in the next day or so.
>
> Isolated to patch,
> [PATCH -v2 4/9] migrate_pages: split unmap_and_move() to _unmap() and _move()
>
> Actually, recreated/isolated by just applying this series to v6.2-rc3 in an
> effort to eliminate any possible noise in linux-next.
>
> Spent a little time looking at modifications made there, but nothing stood out.
> Will investigate more as time allows.

Thank you very much!  That's really helpful.

Checked that patch again, found that there's an issue about non-lru
pages.  Do you enable zram in your test system?  Can you try the
below debug patch on top of

[PATCH -v2 4/9] migrate_pages: split unmap_and_move() to _unmap() and _move()

The following patches in series need to be rebased for the below
change.  I will test with zram enabled too.

Best Regards,
Huang, Ying

---------------------------8<------------------------------------------------------
diff --git a/mm/migrate.c b/mm/migrate.c
index 4c35c2a49574..7153d954b8a2 100644
--- a/mm/migrate.c
+++ b/mm/migrate.c
@@ -1187,10 +1187,13 @@ static int __migrate_folio_move(struct folio *src, struct folio *dst,
 	int rc;
 	int page_was_mapped = 0;
 	struct anon_vma *anon_vma = NULL;
+	bool is_lru = !__PageMovable(&src->page);
 
 	__migrate_folio_extract(dst, &page_was_mapped, &anon_vma);
 
 	rc = move_to_new_folio(dst, src, mode);
+	if (!unlikely(is_lru))
+		goto out_unlock_both;
 
 	/*
 	 * When successful, push dst to LRU immediately: so that if it
@@ -1211,6 +1214,7 @@ static int __migrate_folio_move(struct folio *src, struct folio *dst,
 		remove_migration_ptes(src,
 			rc == MIGRATEPAGE_SUCCESS ? dst : src, false);
 
+out_unlock_both:
 	folio_unlock(dst);
 	/* Drop an anon_vma reference if we took one */
 	if (anon_vma)