bcachefs: fix another deadlock

4 files changed, 77 insertions, 18 deletions

diff --git a/fs/bcachefs/btree_update_interior.c b/fs/bcachefs/btree_update_interior.c
index 4a75b2e8705c..2ce45bcecbbb 100644
--- a/fs/bcachefs/btree_update_interior.c
+++ b/fs/bcachefs/btree_update_interior.c
@@ -601,8 +601,7 @@ static void bch2_btree_update_free(struct btree_update *as)
 
 static void btree_update_nodes_reachable(struct closure *cl)
 {
-	struct btree_update *as =
-		container_of(cl, struct btree_update, cl);
+	struct btree_update *as = container_of(cl, struct btree_update, cl);
 	struct bch_fs *c = as->c;
 
 	bch2_journal_pin_drop(&c->journal, &as->journal);
@@ -636,10 +635,26 @@ static void btree_update_nodes_reachable(struct closure *cl)
 	bch2_btree_update_free(as);
 }
 
+static void btree_update_wait_on_journal(struct closure *cl)
+{
+	struct btree_update *as = container_of(cl, struct btree_update, cl);
+	struct bch_fs *c = as->c;
+	int ret;
+
+	ret = bch2_journal_open_seq_async(&c->journal, as->journal_seq, cl);
+	if (ret < 0)
+		goto err;
+	if (!ret)
+		continue_at(cl, btree_update_wait_on_journal, system_wq);
+
+	bch2_journal_flush_seq_async(&c->journal, as->journal_seq, cl);
+err:
+	continue_at(cl, btree_update_nodes_reachable, system_wq);
+}
+
 static void btree_update_nodes_written(struct closure *cl)
 {
-	struct btree_update *as =
-		container_of(cl, struct btree_update, cl);
+	struct btree_update *as = container_of(cl, struct btree_update, cl);
 	struct bch_fs *c = as->c;
 	struct btree *b;
 
@@ -736,13 +751,10 @@ retry:
 		 */
 		bch2_journal_pin_drop(&c->journal, &as->journal);
 
-		/*
-		 * And, do a journal write to write the pointer to the new root,
-		 * then wait for it to complete before freeing the nodes we
-		 * replaced:
-		 */
-		bch2_journal_meta_async(&c->journal, cl);
-		break;
+		as->journal_seq = bch2_journal_last_unwritten_seq(&c->journal);
+
+		btree_update_wait_on_journal(cl);
+		return;
 	}
 
 	continue_at(cl, btree_update_nodes_reachable, system_wq);
diff --git a/fs/bcachefs/journal.c b/fs/bcachefs/journal.c
index 5bfa7cd0ca4b..1870534df6d0 100644
--- a/fs/bcachefs/journal.c
+++ b/fs/bcachefs/journal.c
@@ -31,6 +31,12 @@ static void journal_pin_add_entry(struct journal *,
 				  struct journal_entry_pin *,
 				  journal_pin_flush_fn);
 
+static inline void journal_wake(struct journal *j)
+{
+	wake_up(&j->wait);
+	closure_wake_up(&j->async_wait);
+}
+
 static inline struct journal_buf *journal_cur_buf(struct journal *j)
 {
 	return j->buf + j->reservations.idx;
@@ -1280,7 +1286,7 @@ void bch2_journal_halt(struct journal *j)
 	} while ((v = atomic64_cmpxchg(&j->reservations.counter,
 				       old.v, new.v)) != old.v);
 
-	wake_up(&j->wait);
+	journal_wake(j);
 	closure_wake_up(&journal_cur_buf(j)->wait);
 	closure_wake_up(&journal_prev_buf(j)->wait);
 }
@@ -1449,7 +1455,7 @@ static int journal_entry_open(struct journal *j)
 	mod_delayed_work(system_freezable_wq,
 			 &j->write_work,
 			 msecs_to_jiffies(j->write_delay_ms));
-	wake_up(&j->wait);
+	journal_wake(j);
 	return 1;
 }
 
@@ -1545,7 +1551,7 @@ int bch2_journal_replay(struct bch_fs *c, struct list_head *list)
 		}
 
 		if (atomic_dec_and_test(&j->replay_pin_list->count))
-			wake_up(&j->wait);
+			journal_wake(j);
 	}
 
 	j->replay_pin_list = NULL;
@@ -1712,7 +1718,7 @@ static void journal_reclaim_fast(struct journal *j)
 	}
 
 	if (popped)
-		wake_up(&j->wait);
+		journal_wake(j);
 }
 
 /*
@@ -1741,7 +1747,7 @@ static inline void __journal_pin_add(struct journal *j,
 	 * If the journal is currently full,  we might want to call flush_fn
 	 * immediately:
 	 */
-	wake_up(&j->wait);
+	journal_wake(j);
 }
 
 static void journal_pin_add_entry(struct journal *j,
@@ -1991,7 +1997,7 @@ static void journal_reclaim_work(struct work_struct *work)
 			ja->last_idx = (ja->last_idx + 1) % ja->nr;
 			spin_unlock(&j->lock);
 
-			wake_up(&j->wait);
+			journal_wake(j);
 		}
 
 		/*
@@ -2231,7 +2237,7 @@ out:
 		     &j->reservations.counter);
 
 	closure_wake_up(&w->wait);
-	wake_up(&j->wait);
+	journal_wake(j);
 
 	if (test_bit(JOURNAL_NEED_WRITE, &j->flags))
 		mod_delayed_work(system_freezable_wq, &j->write_work, 0);
@@ -2536,6 +2542,43 @@ int bch2_journal_res_get_slowpath(struct journal *j, struct journal_res *res,
 	return ret < 0 ? ret : 0;
 }
 
+u64 bch2_journal_last_unwritten_seq(struct journal *j)
+{
+	u64 seq;
+
+	spin_lock(&j->lock);
+	seq = atomic64_read(&j->seq);
+	if (j->reservations.prev_buf_unwritten)
+		seq--;
+	spin_unlock(&j->lock);
+
+	return seq;
+}
+
+int bch2_journal_open_seq_async(struct journal *j, u64 seq, struct closure *parent)
+{
+	int ret;
+
+	spin_lock(&j->lock);
+	BUG_ON(seq > atomic64_read(&j->seq));
+
+	if (seq < atomic64_read(&j->seq) ||
+	    journal_entry_is_open(j)) {
+		spin_unlock(&j->lock);
+		return 1;
+	}
+
+	ret = journal_entry_open(j);
+	if (!ret)
+		closure_wait(&j->async_wait, parent);
+	spin_unlock(&j->lock);
+
+	if (!ret)
+		journal_reclaim_work(&j->reclaim_work.work);
+
+	return ret;
+}
+
 void bch2_journal_wait_on_seq(struct journal *j, u64 seq, struct closure *parent)
 {
 	spin_lock(&j->lock);
diff --git a/fs/bcachefs/journal.h b/fs/bcachefs/journal.h
index ebb846438915..9461d2d8a23e 100644
--- a/fs/bcachefs/journal.h
+++ b/fs/bcachefs/journal.h
@@ -354,6 +354,9 @@ out:
 	return 0;
 }
 
+u64 bch2_journal_last_unwritten_seq(struct journal *);
+int bch2_journal_open_seq_async(struct journal *, u64, struct closure *);
+
 void bch2_journal_wait_on_seq(struct journal *, u64, struct closure *);
 void bch2_journal_flush_seq_async(struct journal *, u64, struct closure *);
 void bch2_journal_flush_async(struct journal *, struct closure *);
diff --git a/fs/bcachefs/journal_types.h b/fs/bcachefs/journal_types.h
index 52fc7a6ca27d..e39b18f27058 100644
--- a/fs/bcachefs/journal_types.h
+++ b/fs/bcachefs/journal_types.h
@@ -140,6 +140,7 @@ struct journal {
 
 	/* Used when waiting because the journal was full */
 	wait_queue_head_t	wait;
+	struct closure_waitlist	async_wait;
 
 	struct closure		io;
 	struct delayed_work	write_work;