bcachefs: Fix race between journal_seq_copy() and journal_seq_drop()

In bch2_btree_interior_update_will_free_node, we copy the journal pins
from outstanding writes on the btree node we're about to free. But, this
can race with the writes completing, and dropping their journal pins.

To guard against this, just use READ_ONCE() in bch2_journal_pin_copy().

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

1 files changed, 12 insertions, 2 deletions

diff --git a/fs/bcachefs/journal_reclaim.c b/fs/bcachefs/journal_reclaim.c
index e8fd11abe4c3..1141b7d3a060 100644
--- a/fs/bcachefs/journal_reclaim.c
+++ b/fs/bcachefs/journal_reclaim.c
@@ -384,12 +384,22 @@ void bch2_journal_pin_set(struct journal *j, u64 seq,
 	struct journal_entry_pin_list *pin_list;
 
 	spin_lock(&j->lock);
+
+	if (seq < journal_last_seq(j)) {
+		/*
+		 * bch2_journal_pin_copy() raced with bch2_journal_pin_drop() on
+		 * the src pin - with the pin dropped, the entry to pin might no
+		 * longer to exist, but that means there's no longer anything to
+		 * copy and we can bail out here:
+		 */
+		spin_unlock(&j->lock);
+		return;
+	}
+
 	pin_list = journal_seq_pin(j, seq);
 
 	__journal_pin_drop(j, pin);
 
-	BUG_ON(!atomic_read(&pin_list->count) && seq == journal_last_seq(j));
-
 	atomic_inc(&pin_list->count);
 	pin->seq	= seq;
 	pin->flush	= flush_fn;