bcachefs: Fix deadlock on nocow locks in data move path

The recent nocow locking rework introduced a deadlock in the data move
path: the new nocow locking scheme uses a hash table with a fixed size
array for chaining, meaning on hash collision we may have to wait for
other locks to be released before we can lock a bucket.

And since the data move path needs to submit writes from the same thread
that's taking nocow locks and submitting reads, this introduces a
deadlock.

This shouldn't happen often in practice, but since the data move path
can keep large numbers of IOs in flight simultaneously, it's something
we have to handle.

This patch makes move_ctxt_wait_event() available to
bch2_data_update_init() and uses it when appropriate, which is our
normal solution to this kind of thing.

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

1 files changed, 13 insertions, 0 deletions

diff --git a/fs/bcachefs/move.h b/fs/bcachefs/move.h
index b14f679f6904..a2822d4a4afb 100644
--- a/fs/bcachefs/move.h
+++ b/fs/bcachefs/move.h
@@ -28,6 +28,16 @@ struct moving_context {
 	wait_queue_head_t	wait;
 };
 
+#define move_ctxt_wait_event(_ctxt, _trans, _cond)		\
+do {								\
+	bch2_moving_ctxt_do_pending_writes(_ctxt, _trans);	\
+								\
+	if (_cond)						\
+		break;						\
+	__wait_event((_ctxt)->wait,				\
+		     bch2_moving_ctxt_next_pending_write(_ctxt) || (_cond));\
+} while (1)
+
 typedef bool (*move_pred_fn)(struct bch_fs *, void *, struct bkey_s_c,
 			     struct bch_io_opts *, struct data_update_opts *);
 
@@ -35,6 +45,9 @@ void bch2_moving_ctxt_exit(struct moving_context *);
 void bch2_moving_ctxt_init(struct moving_context *, struct bch_fs *,
 			   struct bch_ratelimit *, struct bch_move_stats *,
 			   struct write_point_specifier, bool);
+struct moving_io *bch2_moving_ctxt_next_pending_write(struct moving_context *);
+void bch2_moving_ctxt_do_pending_writes(struct moving_context *,
+					struct btree_trans *);
 
 int bch2_scan_old_btree_nodes(struct bch_fs *, struct bch_move_stats *);