mt76: mt7921e: process txfree and txstatus without allocating skbs

Similar to mt7915 driver, process txfree and txstatus without allocating
skbs in order to reduce pressure on the memory allocator
Signed-off-by: Lorenzo Bianconi <lorenzo@kernel.org>
Signed-off-by: Felix Fietkau <nbd@nbd.name>

drivers/net/wireless/mediatek/mt76/mt7921/mac.c

@@ -1195,6 +1195,7 @@ void mt7921_mac_add_txs(struct mt7921_dev *dev, void *data)
 out:
 	rcu_read_unlock();
 }
+EXPORT_SYMBOL_GPL(mt7921_mac_add_txs);
 
 void mt7921_queue_rx_skb(struct mt76_dev *mdev, enum mt76_rxq_id q,
 			 struct sk_buff *skb)

drivers/net/wireless/mediatek/mt76/mt7921/mt7921.h

@@ -434,6 +434,7 @@ int mt7921_mcu_fill_message(struct mt76_dev *mdev, struct sk_buff *skb,
 int mt7921_mcu_parse_response(struct mt76_dev *mdev, int cmd,
 			      struct sk_buff *skb, int seq);
 
+bool mt7921e_rx_check(struct mt76_dev *mdev, void *data, int len);
 void mt7921e_queue_rx_skb(struct mt76_dev *mdev, enum mt76_rxq_id q,
 			  struct sk_buff *skb);
 int mt7921e_driver_own(struct mt7921_dev *dev);

drivers/net/wireless/mediatek/mt76/mt7921/pci.c

@@ -134,6 +134,7 @@ static int mt7921_pci_probe(struct pci_dev *pdev,
 		.token_size = MT7921_TOKEN_SIZE,
 		.tx_prepare_skb = mt7921e_tx_prepare_skb,
 		.tx_complete_skb = mt7921e_tx_complete_skb,
+		.rx_check = mt7921e_rx_check,
 		.rx_skb = mt7921e_queue_rx_skb,
 		.rx_poll_complete = mt7921_rx_poll_complete,
 		.sta_ps = mt7921_sta_ps,

drivers/net/wireless/mediatek/mt76/mt7921/pci_mac.c

@@ -148,14 +148,15 @@ mt7921_txwi_free(struct mt7921_dev *dev, struct mt76_txwi_cache *t,
 }
 
 static void
-mt7921_mac_tx_free(struct mt7921_dev *dev, struct sk_buff *skb)
+mt7921e_mac_tx_free(struct mt7921_dev *dev, void *data, int len)
 {
-	struct mt7921_tx_free *free = (struct mt7921_tx_free *)skb->data;
+	struct mt7921_tx_free *free = (struct mt7921_tx_free *)data;
 	struct mt76_dev *mdev = &dev->mt76;
 	struct mt76_txwi_cache *txwi;
 	struct ieee80211_sta *sta = NULL;
+	struct sk_buff *skb, *tmp;
+	void *end = data + len;
 	LIST_HEAD(free_list);
-	struct sk_buff *tmp;
 	bool wake = false;
 	u8 i, count;
 
@@ -168,6 +169,9 @@ mt7921_mac_tx_free(struct mt7921_dev *dev, struct sk_buff *skb)
 	 * Should avoid accessing WTBL to get Tx airtime, and use it instead.
 	 */
 	count = FIELD_GET(MT_TX_FREE_MSDU_CNT, le16_to_cpu(free->ctrl));
+	if (WARN_ON_ONCE((void *)&free->info[count] > end))
+		return;
+
 	for (i = 0; i < count; i++) {
 		u32 msdu, info = le32_to_cpu(free->info[i]);
 		u8 stat;
@@ -208,8 +212,6 @@ mt7921_mac_tx_free(struct mt7921_dev *dev, struct sk_buff *skb)
 	if (wake)
 		mt76_set_tx_blocked(&dev->mt76, false);
 
-	napi_consume_skb(skb, 1);
-
 	list_for_each_entry_safe(skb, tmp, &free_list, list) {
 		skb_list_del_init(skb);
 		napi_consume_skb(skb, 1);
@@ -222,6 +224,27 @@ mt7921_mac_tx_free(struct mt7921_dev *dev, struct sk_buff *skb)
 	mt76_worker_schedule(&dev->mt76.tx_worker);
 }
 
+bool mt7921e_rx_check(struct mt76_dev *mdev, void *data, int len)
+{
+	struct mt7921_dev *dev = container_of(mdev, struct mt7921_dev, mt76);
+	__le32 *rxd = (__le32 *)data;
+	__le32 *end = (__le32 *)&rxd[len / 4];
+	enum rx_pkt_type type;
+
+	type = FIELD_GET(MT_RXD0_PKT_TYPE, le32_to_cpu(rxd[0]));
+	switch (type) {
+	case PKT_TYPE_TXRX_NOTIFY:
+		mt7921e_mac_tx_free(dev, data, len);
+		return false;
+	case PKT_TYPE_TXS:
+		for (rxd += 2; rxd + 8 <= end; rxd += 8)
+			mt7921_mac_add_txs(dev, rxd);
+		return false;
+	default:
+		return true;
+	}
+}
+
 void mt7921e_queue_rx_skb(struct mt76_dev *mdev, enum mt76_rxq_id q,
 			  struct sk_buff *skb)
 {
@@ -233,7 +256,8 @@ void mt7921e_queue_rx_skb(struct mt76_dev *mdev, enum mt76_rxq_id q,
 
 	switch (type) {
 	case PKT_TYPE_TXRX_NOTIFY:
-		mt7921_mac_tx_free(dev, skb);
+		mt7921e_mac_tx_free(dev, skb->data, skb->len);
+		napi_consume_skb(skb, 1);
 		break;
 	default:
 		mt7921_queue_rx_skb(mdev, q, skb);