bcachefs: Don't make foreground writes wait behind journal reclaim too long

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

fs/bcachefs/btree_update_leaf.c

@@ -586,6 +586,28 @@ static inline int do_bch2_trans_commit(struct btree_trans *trans,
 	return 0;
 }
 
+static int journal_reclaim_wait_done(struct bch_fs *c)
+{
+	int ret;
+
+	ret = bch2_journal_error(&c->journal);
+	if (ret)
+		return ret;
+
+	ret = !bch2_btree_key_cache_must_wait(c);
+	if (ret)
+		return ret;
+
+	if (mutex_trylock(&c->journal.reclaim_lock)) {
+		ret = bch2_journal_reclaim(&c->journal);
+		mutex_unlock(&c->journal.reclaim_lock);
+	}
+
+	if (!ret)
+		ret = !bch2_btree_key_cache_must_wait(c);
+	return ret;
+}
+
 static noinline
 int bch2_trans_commit_error(struct btree_trans *trans,
 			    struct btree_insert_entry *i,
@@ -668,13 +690,12 @@ int bch2_trans_commit_error(struct btree_trans *trans,
 	case BTREE_INSERT_NEED_JOURNAL_RECLAIM:
 		bch2_trans_unlock(trans);
 
-		do {
-			mutex_lock(&c->journal.reclaim_lock);
-			ret = bch2_journal_reclaim(&c->journal);
-			mutex_unlock(&c->journal.reclaim_lock);
-		} while (!ret && bch2_btree_key_cache_must_wait(c));
+		wait_event(c->journal.reclaim_wait,
+			   (ret = journal_reclaim_wait_done(c)));
+		if (ret < 0)
+			return ret;
 
-		if (!ret && bch2_trans_relock(trans))
+		if (bch2_trans_relock(trans))
 			return 0;
 
 		trace_trans_restart_journal_reclaim(trans->ip);

fs/bcachefs/journal.c

@@ -1116,6 +1116,7 @@ int bch2_fs_journal_init(struct journal *j)
 	spin_lock_init(&j->err_lock);
 	init_waitqueue_head(&j->wait);
 	INIT_DELAYED_WORK(&j->write_work, journal_write_work);
+	init_waitqueue_head(&j->reclaim_wait);
 	init_waitqueue_head(&j->pin_flush_wait);
 	mutex_init(&j->reclaim_lock);
 	mutex_init(&j->discard_lock);

fs/bcachefs/journal_reclaim.c

@@ -604,6 +604,9 @@ static int __bch2_journal_reclaim(struct journal *j, bool direct)
 
 		min_nr = max(min_nr, bch2_nr_btree_keys_want_flush(c));
 
+		/* Don't do too many without delivering wakeup: */
+		min_nr = min(min_nr, 128UL);
+
 		trace_journal_reclaim_start(c,
 				min_nr,
 				j->prereserved.reserved,
@@ -620,6 +623,9 @@ static int __bch2_journal_reclaim(struct journal *j, bool direct)
 		else
 			j->nr_background_reclaim += nr_flushed;
 		trace_journal_reclaim_finish(c, nr_flushed);
+
+		if (nr_flushed)
+			wake_up(&j->reclaim_wait);
 	} while (min_nr && nr_flushed);
 
 	memalloc_noreclaim_restore(flags);

fs/bcachefs/journal_types.h

@@ -243,6 +243,7 @@ struct journal {
 	spinlock_t		err_lock;
 
 	struct mutex		reclaim_lock;
+	wait_queue_head_t	reclaim_wait;
 	struct task_struct	*reclaim_thread;
 	bool			reclaim_kicked;
 	u64			nr_direct_reclaim;