mt76: mt7663s: fix the possible device hang in high traffic

Use the additional memory barrier to ensure the skb list up-to-date
between the skb producer and consumer to avoid the invalid skb content
written into sdio controller and then cause device hang due to mcu assert
caught by WR_TIMEOUT_INT.

Fixes: 1522ff73 ("mt76: mt7663s: introduce sdio tx aggregation")
Signed-off-by: Sean Wang <sean.wang@mediatek.com>
Signed-off-by: Felix Fietkau <nbd@nbd.name>

drivers/net/wireless/mediatek/mt76/mt7615/sdio_txrx.c

@@ -225,6 +225,8 @@ static int mt7663s_tx_run_queue(struct mt76_dev *dev, struct mt76_queue *q)
 		struct mt76_queue_entry *e = &q->entry[q->first];
 		struct sk_buff *iter;
 
+		smp_rmb();
+
 		if (!test_bit(MT76_STATE_MCU_RUNNING, &dev->phy.state)) {
 			__skb_put_zero(e->skb, 4);
 			err = __mt7663s_xmit_queue(dev, e->skb->data,

drivers/net/wireless/mediatek/mt76/sdio.c

@@ -256,6 +256,9 @@ mt76s_tx_queue_skb(struct mt76_dev *dev, struct mt76_queue *q,
 
 	q->entry[q->head].skb = tx_info.skb;
 	q->entry[q->head].buf_sz = len;
+
+	smp_wmb();
+
 	q->head = (q->head + 1) % q->ndesc;
 	q->queued++;