diff --git a/dma.c b/dma.c
index e2d86eec..8d1074f1 100644
--- a/dma.c
+++ b/dma.c
@@ -88,6 +88,7 @@ mt76_dma_alloc_queue(struct mt76_dev *dev, struct mt76_queue *q,
 	int i;
 
 	spin_lock_init(&q->lock);
+	spin_lock_init(&q->cleanup_lock);
 
 	q->regs = dev->mmio.regs + ring_base + idx * MT_RING_SIZE;
 	q->ndesc = n_desc;
@@ -225,6 +226,7 @@ mt76_dma_tx_cleanup(struct mt76_dev *dev, struct mt76_queue *q, bool flush)
 	if (!q)
 		return;
 
+	spin_lock_bh(&q->cleanup_lock);
 	if (flush)
 		last = -1;
 	else
@@ -243,6 +245,7 @@ mt76_dma_tx_cleanup(struct mt76_dev *dev, struct mt76_queue *q, bool flush)
 			last = readl(&q->regs->dma_idx);
 
 	}
+	spin_unlock_bh(&q->cleanup_lock);
 
 	if (flush) {
 		spin_lock_bh(&q->lock);
diff --git a/mt76.h b/mt76.h
index d1997a98..58e7145e 100644
--- a/mt76.h
+++ b/mt76.h
@@ -126,6 +126,7 @@ struct mt76_queue {
 	struct mt76_queue_regs __iomem *regs;
 
 	spinlock_t lock;
+	spinlock_t cleanup_lock;
 	struct mt76_queue_entry *entry;
 	struct mt76_desc *desc;
 
diff --git a/tx.c b/tx.c
index 41dd0a80..234c8202 100644
--- a/tx.c
+++ b/tx.c
@@ -444,6 +444,7 @@ static int
 mt76_txq_schedule_list(struct mt76_phy *phy, enum mt76_txq_id qid)
 {
 	struct mt76_queue *q = phy->q_tx[qid];
+	struct mt76_dev *dev = phy->dev;
 	struct ieee80211_txq *txq;
 	struct mt76_txq *mtxq;
 	struct mt76_wcid *wcid;
@@ -460,6 +461,13 @@ mt76_txq_schedule_list(struct mt76_phy *phy, enum mt76_txq_id qid)
 		if (q->stopped || q->blocked)
 			break;
 
+		if (dev->queue_ops->tx_cleanup &&
+		    q->queued + 2 * MT_TXQ_FREE_THR >= q->ndesc) {
+			spin_unlock_bh(&q->lock);
+			dev->queue_ops->tx_cleanup(dev, q, false);
+			spin_lock_bh(&q->lock);
+		}
+
 		if (q->queued + MT_TXQ_FREE_THR >= q->ndesc)
 			break;