[PATCH v2] fs/dax: Fix pmd vs pte conflict detection

[Dan Williams]

Users reported a v5.3 performance regression and inability to establish
huge page mappings. A revised version of the ndctl "dax.sh" huge page
unit test identifies commit 23c84eb78375 "dax: Fix missed wakeup with
PMD faults" as the source.

Update get_unlocked_entry() to check for NULL entries before checking
the entry order, otherwise NULL is misinterpreted as a present pte
conflict. The 'order' check needs to happen before the locked check as
an unlocked entry at the wrong order must fallback to lookup the correct
order.

Reported-by: Jeff Smits <jeff.smits@xxxxxxxxx>
Reported-by: Doug Nelson <doug.nelson@xxxxxxxxx>
Cc: <stable@xxxxxxxxxxxxxxx>
Fixes: 23c84eb78375 ("dax: Fix missed wakeup with PMD faults")
Reviewed-by: Jan Kara <jack@xxxxxxx>
Cc: Jeff Moyer <jmoyer@xxxxxxxxxx>
Cc: Matthew Wilcox (Oracle) <willy@xxxxxxxxxxxxx>
Signed-off-by: Dan Williams <dan.j.williams@xxxxxxxxx>
---
Changes in v2:
- Update the changelog to reflect the user visible effects of the bug
  (Jeff)

 fs/dax.c |    5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)

diff --git a/fs/dax.c b/fs/dax.c
index 6bf81f931de3..2cc43cd914eb 100644
--- a/fs/dax.c
+++ b/fs/dax.c
@@ -220,10 +220,11 @@ static void *get_unlocked_entry(struct xa_state *xas, unsigned int order)
 
 	for (;;) {
 		entry = xas_find_conflict(xas);
+		if (!entry || WARN_ON_ONCE(!xa_is_value(entry)))
+			return entry;
 		if (dax_entry_order(entry) < order)
 			return XA_RETRY_ENTRY;
-		if (!entry || WARN_ON_ONCE(!xa_is_value(entry)) ||
-				!dax_is_locked(entry))
+		if (!dax_is_locked(entry))
 			return entry;
 
 		wq = dax_entry_waitqueue(xas, entry, &ewait.key);