1 files changed, 16 insertions, 2 deletions

diff --git a/fs/xfs/libxfs/xfs_btree.c b/fs/xfs/libxfs/xfs_btree.c
index 35f574421670..c4649cc624e1 100644
--- a/fs/xfs/libxfs/xfs_btree.c
+++ b/fs/xfs/libxfs/xfs_btree.c
@@ -2913,9 +2913,22 @@ xfs_btree_split_worker(
 }
 
 /*
- * BMBT split requests often come in with little stack to work on. Push
+ * BMBT split requests often come in with little stack to work on so we push
  * them off to a worker thread so there is lots of stack to use. For the other
  * btree types, just call directly to avoid the context switch overhead here.
+ *
+ * Care must be taken here - the work queue rescuer thread introduces potential
+ * AGF <> worker queue deadlocks if the BMBT block allocation has to lock new
+ * AGFs to allocate blocks. A task being run by the rescuer could attempt to
+ * lock an AGF that is already locked by a task queued to run by the rescuer,
+ * resulting in an ABBA deadlock as the rescuer cannot run the lock holder to
+ * release it until the current thread it is running gains the lock.
+ *
+ * To avoid this issue, we only ever queue BMBT splits that don't have an AGF
+ * already locked to allocate from. The only place that doesn't hold an AGF
+ * locked is unwritten extent conversion at IO completion, but that has already
+ * been offloaded to a worker thread and hence has no stack consumption issues
+ * we have to worry about.
  */
 STATIC int					/* error */
 xfs_btree_split(
@@ -2929,7 +2942,8 @@ xfs_btree_split(
 	struct xfs_btree_split_args	args;
 	DECLARE_COMPLETION_ONSTACK(done);
 
-	if (cur->bc_btnum != XFS_BTNUM_BMAP)
+	if (cur->bc_btnum != XFS_BTNUM_BMAP ||
+	    cur->bc_tp->t_highest_agno == NULLAGNUMBER)
 		return __xfs_btree_split(cur, level, ptrp, key, curp, stat);
 
 	args.cur = cur;