wifi: mt76: fix race condition related to checking tx queue fill status

When drv_tx calls race against local tx scheduling, the queue fill status checks
can potentially race, leading to dma queue entries being overwritten.
Fix this by deferring packets from drv_tx calls to the tx worker, in order to
ensure that all regular queue tx comes from the same context.

Signed-off-by: Felix Fietkau <[email protected]>

mac80211.c

@@ -416,6 +416,9 @@ mt76_phy_init(struct mt76_phy *phy, struct ieee80211_hw *hw)
 	struct mt76_dev *dev = phy->dev;
 	struct wiphy *wiphy = hw->wiphy;
 
+	INIT_LIST_HEAD(&phy->tx_list);
+	spin_lock_init(&phy->tx_lock);
+
 	SET_IEEE80211_DEV(hw, dev->dev);
 	SET_IEEE80211_PERM_ADDR(hw, phy->macaddr);
 
@@ -690,6 +693,7 @@ int mt76_register_device(struct mt76_dev *dev, bool vht,
 	int ret;
 
 	dev_set_drvdata(dev->dev, dev);
+	mt76_wcid_init(&dev->global_wcid);
 	ret = mt76_phy_init(phy, hw);
 	if (ret)
 		return ret;
@@ -745,6 +749,7 @@ void mt76_unregister_device(struct mt76_dev *dev)
 	if (IS_ENABLED(CONFIG_MT76_LEDS))
 		mt76_led_cleanup(&dev->phy);
 	mt76_tx_status_check(dev, true);
+	mt76_wcid_cleanup(dev, &dev->global_wcid);
 	ieee80211_unregister_hw(hw);
 }
 EXPORT_SYMBOL_GPL(mt76_unregister_device);
@@ -1485,6 +1490,9 @@ EXPORT_SYMBOL_GPL(mt76_sta_pre_rcu_remove);
 
 void mt76_wcid_init(struct mt76_wcid *wcid)
 {
+	INIT_LIST_HEAD(&wcid->tx_list);
+	skb_queue_head_init(&wcid->tx_pending);
+
 	INIT_LIST_HEAD(&wcid->list);
 	idr_init(&wcid->pktid);
 }
@@ -1492,13 +1500,32 @@ EXPORT_SYMBOL_GPL(mt76_wcid_init);
 
 void mt76_wcid_cleanup(struct mt76_dev *dev, struct mt76_wcid *wcid)
 {
+	struct mt76_phy *phy = dev->phys[wcid->phy_idx];
+	struct ieee80211_hw *hw;
 	struct sk_buff_head list;
+	struct sk_buff *skb;
 
 	mt76_tx_status_lock(dev, &list);
 	mt76_tx_status_skb_get(dev, wcid, -1, &list);
 	mt76_tx_status_unlock(dev, &list);
 
 	idr_destroy(&wcid->pktid);
+
+	spin_lock_bh(&phy->tx_lock);
+
+	if (!list_empty(&wcid->tx_list))
+		list_del_init(&wcid->tx_list);
+
+	spin_lock(&wcid->tx_pending.lock);
+	skb_queue_splice_tail_init(&wcid->tx_pending, &list);
+	spin_unlock(&wcid->tx_pending.lock);
+
+	spin_unlock_bh(&phy->tx_lock);
+
+	while ((skb = __skb_dequeue(&list)) != NULL) {
+		hw = mt76_tx_status_get_hw(dev, skb);
+		ieee80211_free_txskb(hw, skb);
+	}
 }
 EXPORT_SYMBOL_GPL(mt76_wcid_cleanup);
 

mt76.h

@@ -333,6 +333,9 @@ struct mt76_wcid {
 	u32 tx_info;
 	bool sw_iv;
 
+	struct list_head tx_list;
+	struct sk_buff_head tx_pending;
+
 	struct list_head list;
 	struct idr pktid;
 
@@ -708,6 +711,8 @@ struct mt76_phy {
 	unsigned long state;
 	u8 band_idx;
 
+	spinlock_t tx_lock;
+	struct list_head tx_list;
 	struct mt76_queue *q_tx[__MT_TXQ_MAX];
 
 	struct cfg80211_chan_def chandef;

tx.c

@@ -329,40 +329,32 @@ void
 mt76_tx(struct mt76_phy *phy, struct ieee80211_sta *sta,
 	struct mt76_wcid *wcid, struct sk_buff *skb)
 {
-	struct mt76_dev *dev = phy->dev;
 	struct ieee80211_tx_info *info = IEEE80211_SKB_CB(skb);
-	struct ieee80211_hdr *hdr = (struct ieee80211_hdr *)skb->data;
-	struct mt76_queue *q;
-	int qid = skb_get_queue_mapping(skb);
 
 	if (mt76_testmode_enabled(phy)) {
 		ieee80211_free_txskb(phy->hw, skb);
 		return;
 	}
 
-	if (WARN_ON(qid >= MT_TXQ_PSD)) {
+	if (WARN_ON(skb_get_queue_mapping(skb) >= MT_TXQ_PSD))
-		qid = MT_TXQ_BE;
+		skb_set_queue_mapping(skb, MT_TXQ_BE);
-		skb_set_queue_mapping(skb, qid);
-	}
-
-	if ((dev->drv->drv_flags & MT_DRV_HW_MGMT_TXQ) &&
-	    !(info->flags & IEEE80211_TX_CTL_HW_80211_ENCAP) &&
-	    !ieee80211_is_data(hdr->frame_control) &&
-	    !ieee80211_is_bufferable_mmpdu(skb)) {
-		qid = MT_TXQ_PSD;
-	}
 
 	if (wcid && !(wcid->tx_info & MT_WCID_TX_INFO_SET))
 		ieee80211_get_tx_rates(info->control.vif, sta, skb,
 				       info->control.rates, 1);
 
 	info->hw_queue |= FIELD_PREP(MT_TX_HW_QUEUE_PHY, phy->band_idx);
-	q = phy->q_tx[qid];
 
-	spin_lock_bh(&q->lock);
+	spin_lock_bh(&wcid->tx_pending.lock);
-	__mt76_tx_queue_skb(phy, qid, skb, wcid, sta, NULL);
+	__skb_queue_tail(&wcid->tx_pending, skb);
-	dev->queue_ops->kick(dev, q);
+	spin_unlock_bh(&wcid->tx_pending.lock);
-	spin_unlock_bh(&q->lock);
+
+	spin_lock_bh(&phy->tx_lock);
+	if (list_empty(&wcid->tx_list))
+		list_add_tail(&wcid->tx_list, &phy->tx_list);
+	spin_unlock_bh(&phy->tx_lock);
+
+	mt76_worker_schedule(&phy->dev->tx_worker);
 }
 EXPORT_SYMBOL_GPL(mt76_tx);
 
@@ -593,10 +585,86 @@ void mt76_txq_schedule(struct mt76_phy *phy, enum mt76_txq_id qid)
 }
 EXPORT_SYMBOL_GPL(mt76_txq_schedule);
 
+static int
+mt76_txq_schedule_pending_wcid(struct mt76_phy *phy, struct mt76_wcid *wcid)
+{
+	struct mt76_dev *dev = phy->dev;
+	struct ieee80211_sta *sta;
+	struct mt76_queue *q;
+	struct sk_buff *skb;
+	int ret = 0;
+
+	spin_lock(&wcid->tx_pending.lock);
+	while ((skb = skb_peek(&wcid->tx_pending)) != NULL) {
+		struct ieee80211_hdr *hdr = (struct ieee80211_hdr *)skb->data;
+		struct ieee80211_tx_info *info = IEEE80211_SKB_CB(skb);
+		int qid = skb_get_queue_mapping(skb);
+
+		if ((dev->drv->drv_flags & MT_DRV_HW_MGMT_TXQ) &&
+		    !(info->flags & IEEE80211_TX_CTL_HW_80211_ENCAP) &&
+		    !ieee80211_is_data(hdr->frame_control) &&
+		    !ieee80211_is_bufferable_mmpdu(skb))
+			qid = MT_TXQ_PSD;
+
+		q = phy->q_tx[qid];
+		if (mt76_txq_stopped(q)) {
+			ret = -1;
+			break;
+		}
+
+		__skb_unlink(skb, &wcid->tx_pending);
+		spin_unlock(&wcid->tx_pending.lock);
+
+		sta = wcid_to_sta(wcid);
+		spin_lock(&q->lock);
+		__mt76_tx_queue_skb(phy, qid, skb, wcid, sta, NULL);
+		dev->queue_ops->kick(dev, q);
+		spin_unlock(&q->lock);
+
+		spin_lock(&wcid->tx_pending.lock);
+	}
+	spin_unlock(&wcid->tx_pending.lock);
+
+	return ret;
+}
+
+static void mt76_txq_schedule_pending(struct mt76_phy *phy)
+{
+	if (list_empty(&phy->tx_list))
+		return;
+
+	local_bh_disable();
+	rcu_read_lock();
+
+	spin_lock(&phy->tx_lock);
+	while (!list_empty(&phy->tx_list)) {
+		struct mt76_wcid *wcid = NULL;
+		int ret;
+
+		wcid = list_first_entry(&phy->tx_list, struct mt76_wcid, tx_list);
+		list_del_init(&wcid->tx_list);
+
+		spin_unlock(&phy->tx_lock);
+		ret = mt76_txq_schedule_pending_wcid(phy, wcid);
+		spin_lock(&phy->tx_lock);
+
+		if (ret) {
+			if (list_empty(&wcid->tx_list))
+				list_add_tail(&wcid->tx_list, &phy->tx_list);
+			break;
+		}
+	}
+	spin_unlock(&phy->tx_lock);
+
+	rcu_read_unlock();
+	local_bh_enable();
+}
+
 void mt76_txq_schedule_all(struct mt76_phy *phy)
 {
 	int i;
 
+	mt76_txq_schedule_pending(phy);
 	for (i = 0; i <= MT_TXQ_BK; i++)
 		mt76_txq_schedule(phy, i);
 }