From c9cbe5b9cc6e0d17352814aafe6514a6623bbd12 Mon Sep 17 00:00:00 2001
From: Bo Jiao <Bo.Jiao@mediatek.com>
Date: Mon, 6 Feb 2023 13:50:56 +0800
Subject: [PATCH 2004/2012] wifi: mt76: mt7996: wed: add wed3.0 rx support

add hardware rro support, This is the preliminary patch for WED3.0 support.

Signed-off-by: Bo Jiao <Bo.Jiao@mediatek.com>
Change-Id: I7e113b1392bcf085ec02c8a44ffbb7cf7c3fa027
Signed-off-by: sujuan.chen <sujuan.chen@mediatek.com>
---
 dma.c           | 197 +++++++++++++++++++++++++++++++++++++-----------
 dma.h           |  12 +++
 mac80211.c      |   1 +
 mt76.h          |  63 ++++++++++++++--
 mt7996/dma.c    | 161 ++++++++++++++++++++++++++++++++++-----
 mt7996/init.c   | 130 ++++++++++++++++++++++++++++++--
 mt7996/mac.c    |  42 +++++++++--
 mt7996/mcu.c    |   8 +-
 mt7996/mmio.c   |  44 +++++++++--
 mt7996/mt7996.h |  58 ++++++++++++++
 mt7996/pci.c    |   3 +-
 mt7996/regs.h   |  69 ++++++++++++++++-
 12 files changed, 693 insertions(+), 95 deletions(-)

diff --git a/dma.c b/dma.c
index c2dbe6f6b..8097a3121 100644
--- a/dma.c
+++ b/dma.c
@@ -198,46 +198,65 @@ EXPORT_SYMBOL_GPL(mt76_free_pending_rxwi);
 static void
 mt76_dma_sync_idx(struct mt76_dev *dev, struct mt76_queue *q)
 {
+	int ndesc = q->ndesc;
+
+	if (q->flags & MT_QFLAG_MAGIC)
+		ndesc |= MT_DMA_MAGIC_EN;
+
 	Q_WRITE(dev, q, desc_base, q->desc_dma);
-	Q_WRITE(dev, q, ring_size, q->ndesc);
+	Q_WRITE(dev, q, ring_size, ndesc);
 	q->head = Q_READ(dev, q, dma_idx);
 	q->tail = q->head;
 }
 
 static void
-mt76_dma_queue_reset(struct mt76_dev *dev, struct mt76_queue *q)
+mt76_dma_queue_reset(struct mt76_dev *dev, struct mt76_queue *q, bool skip)
 {
 	int i;
 
 	if (!q || !q->ndesc)
 		return;
 
+	if (!q->desc)
+		goto done;
+
 	/* clear descriptors */
 	for (i = 0; i < q->ndesc; i++)
 		q->desc[i].ctrl = cpu_to_le32(MT_DMA_CTL_DMA_DONE);
 
+	if (skip)
+		goto sync;
+
+done:
 	Q_WRITE(dev, q, cpu_idx, 0);
 	Q_WRITE(dev, q, dma_idx, 0);
+sync:
 	mt76_dma_sync_idx(dev, q);
 }
 
 static int
 mt76_dma_add_rx_buf(struct mt76_dev *dev, struct mt76_queue *q,
-		    struct mt76_queue_buf *buf, void *data)
+		    struct mt76_queue_buf *buf, void *data,
+		    struct mt76_rxwi_cache *rxwi)
 {
-	struct mt76_desc *desc = &q->desc[q->head];
+	struct mt76_desc *desc;
 	struct mt76_queue_entry *entry = &q->entry[q->head];
-	struct mt76_rxwi_cache *rxwi = NULL;
 	u32 buf1 = 0, ctrl;
 	int idx = q->head;
 	int rx_token;
 
+	if (mt76_queue_is_rro_ind(q))
+		goto done;
+
+	desc = &q->desc[q->head];
 	ctrl = FIELD_PREP(MT_DMA_CTL_SD_LEN0, buf[0].len);
 
 	if (mt76_queue_is_wed_rx(q)) {
-		rxwi = mt76_get_rxwi(dev);
-		if (!rxwi)
-			return -ENOMEM;
+		if (!rxwi) {
+			rxwi = mt76_get_rxwi(dev);
+			if (!rxwi)
+				return -ENOMEM;
+		}
 
 		rx_token = mt76_rx_token_consume(dev, data, rxwi, buf->addr);
 		if (rx_token < 0) {
@@ -254,6 +273,7 @@ mt76_dma_add_rx_buf(struct mt76_dev *dev, struct mt76_queue *q,
 	WRITE_ONCE(desc->ctrl, cpu_to_le32(ctrl));
 	WRITE_ONCE(desc->info, 0);
 
+done:
 	entry->dma_addr[0] = buf->addr;
 	entry->dma_len[0] = buf->len;
 	entry->rxwi = rxwi;
@@ -398,14 +418,15 @@ mt76_dma_tx_cleanup(struct mt76_dev *dev, struct mt76_queue *q, bool flush)
 
 static void *
 mt76_dma_get_buf(struct mt76_dev *dev, struct mt76_queue *q, int idx,
-		 int *len, u32 *info, bool *more, bool *drop)
+		 int *len, u32 *info, bool *more, bool *drop, bool flush)
 {
 	struct mt76_queue_entry *e = &q->entry[idx];
 	struct mt76_desc *desc = &q->desc[idx];
 	void *buf;
+	u32 ctrl;
 
+	ctrl = le32_to_cpu(READ_ONCE(desc->ctrl));
 	if (len) {
-		u32 ctrl = le32_to_cpu(READ_ONCE(desc->ctrl));
 		*len = FIELD_GET(MT_DMA_CTL_SD_LEN0, ctrl);
 		*more = !(ctrl & MT_DMA_CTL_LAST_SEC0);
 	}
@@ -413,6 +434,12 @@ mt76_dma_get_buf(struct mt76_dev *dev, struct mt76_queue *q, int idx,
 	if (info)
 		*info = le32_to_cpu(desc->info);
 
+	if (drop) {
+		*drop = !!(ctrl & (MT_DMA_CTL_TO_HOST_A | MT_DMA_CTL_DROP));
+		if (ctrl & MT_DMA_CTL_VER_MASK)
+			*drop = !!(ctrl & MT_DMA_CTL_PN_CHK_FAIL);
+	}
+
 	if (mt76_queue_is_wed_rx(q)) {
 		u32 buf1 = le32_to_cpu(desc->buf1);
 		u32 token = FIELD_GET(MT_DMA_CTL_TOKEN, buf1);
@@ -425,20 +452,46 @@ mt76_dma_get_buf(struct mt76_dev *dev, struct mt76_queue *q, int idx,
 				 SKB_WITH_OVERHEAD(q->buf_size),
 				 DMA_FROM_DEVICE);
 
-		buf = r->ptr;
-		r->dma_addr = 0;
-		r->ptr = NULL;
-
-		mt76_put_rxwi(dev, r);
-
-		if (drop) {
-			u32 ctrl = le32_to_cpu(READ_ONCE(desc->ctrl));
-
-			*drop = !!(ctrl & (MT_DMA_CTL_TO_HOST_A |
-					   MT_DMA_CTL_DROP));
+		if (flush) {
+			buf = r->ptr;
+			r->dma_addr = 0;
+			r->ptr = NULL;
+
+			mt76_put_rxwi(dev, r);
+		} else {
+			struct mt76_queue_buf qbuf;
+
+			buf = page_frag_alloc(&q->rx_page, q->buf_size, GFP_ATOMIC);
+			if (!buf)
+				return NULL;
+
+			memcpy(buf, r->ptr, SKB_WITH_OVERHEAD(q->buf_size));
+
+			r->dma_addr = dma_map_single(dev->dma_dev, r->ptr,
+						     SKB_WITH_OVERHEAD(q->buf_size),
+						     DMA_FROM_DEVICE);
+			if (unlikely(dma_mapping_error(dev->dma_dev, r->dma_addr))) {
+				skb_free_frag(r->ptr);
+				mt76_put_rxwi(dev, r);
+				return NULL;
+			}
+
+			qbuf.addr = r->dma_addr;
+			qbuf.len = SKB_WITH_OVERHEAD(q->buf_size);
+			qbuf.skip_unmap = false;
+
+			if (mt76_dma_add_rx_buf(dev, q, &qbuf, r->ptr, r) < 0) {
+				dma_unmap_single(dev->dma_dev, r->dma_addr,
+						 SKB_WITH_OVERHEAD(q->buf_size),
+						 DMA_FROM_DEVICE);
+				skb_free_frag(r->ptr);
+				mt76_put_rxwi(dev, r);
+				return NULL;
+			}
+		}
 
+		if (drop)
 			*drop |= !!(buf1 & MT_DMA_CTL_WO_DROP);
-		}
 	} else {
 		buf = e->buf;
 		e->buf = NULL;
@@ -460,15 +513,20 @@ mt76_dma_dequeue(struct mt76_dev *dev, struct mt76_queue *q, bool flush,
 	if (!q->queued)
 		return NULL;
 
-	if (flush)
-		q->desc[idx].ctrl |= cpu_to_le32(MT_DMA_CTL_DMA_DONE);
-	else if (!(q->desc[idx].ctrl & cpu_to_le32(MT_DMA_CTL_DMA_DONE)))
-		return NULL;
+	if (q->flags & MT_QFLAG_RRO) {
+		goto done;
+	} else {
+		if (flush)
+			q->desc[idx].ctrl |= cpu_to_le32(MT_DMA_CTL_DMA_DONE);
+		else if (!(q->desc[idx].ctrl & cpu_to_le32(MT_DMA_CTL_DMA_DONE)))
+			return NULL;
+	}
 
+done:
 	q->tail = (q->tail + 1) % q->ndesc;
 	q->queued--;
 
-	return mt76_dma_get_buf(dev, q, idx, len, info, more, drop);
+	return mt76_dma_get_buf(dev, q, idx, len, info, more, drop, flush);
 }
 
 static int
@@ -617,7 +675,10 @@ mt76_dma_rx_fill(struct mt76_dev *dev, struct mt76_queue *q)
 
 	while (q->queued < q->ndesc - 1) {
 		struct mt76_queue_buf qbuf;
-		void *buf;
+		void *buf = NULL;
+
+		if (mt76_queue_is_rro_ind(q))
+			goto done;
 
 		buf = page_frag_alloc(&q->rx_page, q->buf_size, GFP_ATOMIC);
 		if (!buf)
@@ -629,10 +690,11 @@ mt76_dma_rx_fill(struct mt76_dev *dev, struct mt76_queue *q)
 			break;
 		}
 
+done:
 		qbuf.addr = addr + offset;
 		qbuf.len = len - offset;
 		qbuf.skip_unmap = false;
-		if (mt76_dma_add_rx_buf(dev, q, &qbuf, buf) < 0) {
+		if (mt76_dma_add_rx_buf(dev, q, &qbuf, buf, NULL) < 0) {
 			dma_unmap_single(dev->dma_dev, addr, len,
 					 DMA_FROM_DEVICE);
 			skb_free_frag(buf);
@@ -641,7 +703,7 @@ mt76_dma_rx_fill(struct mt76_dev *dev, struct mt76_queue *q)
 		frames++;
 	}
 
-	if (frames)
+	if (frames || mt76_queue_is_wed_rx(q))
 		mt76_dma_kick_queue(dev, q);
 
 	spin_unlock_bh(&q->lock);
@@ -654,7 +716,7 @@ int mt76_dma_wed_setup(struct mt76_dev *dev, struct mt76_queue *q, bool reset)
 #ifdef CONFIG_NET_MEDIATEK_SOC_WED
 	struct mtk_wed_device *wed = &dev->mmio.wed;
 	int ret, type, ring;
-	u8 flags;
+	u16 flags;
 
 	if (!q || !q->ndesc)
 		return -EINVAL;
@@ -681,7 +743,7 @@ int mt76_dma_wed_setup(struct mt76_dev *dev, struct mt76_queue *q, bool reset)
 	case MT76_WED_Q_TXFREE:
 		/* WED txfree queue needs ring to be initialized before setup */
 		q->flags = 0;
-		mt76_dma_queue_reset(dev, q);
+		mt76_dma_queue_reset(dev, q, false);
 		mt76_dma_rx_fill(dev, q);
 		q->flags = flags;
 
@@ -690,9 +752,31 @@ int mt76_dma_wed_setup(struct mt76_dev *dev, struct mt76_queue *q, bool reset)
 			q->wed_regs = wed->txfree_ring.reg_base;
 		break;
 	case MT76_WED_Q_RX:
-		ret = mtk_wed_device_rx_ring_setup(wed, ring, q->regs, reset);
-		if (!ret)
-			q->wed_regs = wed->rx_ring[ring].reg_base;
+		if (q->flags & MT_QFLAG_RRO) {
+			q->flags &= ~0x1f;
+
+			ring = FIELD_GET(MT_QFLAG_RRO_RING, q->flags);
+			type = FIELD_GET(MT_QFLAG_RRO_TYPE, q->flags);
+			if (type == MT76_RRO_Q_DATA) {
+				mt76_dma_queue_reset(dev, q, true);
+				ret = mtk_wed_device_rro_rx_ring_setup(wed, ring, q->regs);
+			} else if (type == MT76_RRO_Q_MSDU_PG) {
+				mt76_dma_queue_reset(dev, q, true);
+				ret = mtk_wed_device_msdu_pg_rx_ring_setup(wed, ring, q->regs);
+			} else if (type == MT76_RRO_Q_IND) {
+				mt76_dma_queue_reset(dev, q, false);
+				mt76_dma_rx_fill(dev, q);
+				ret = mtk_wed_device_ind_rx_ring_setup(wed, q->regs);
+			}
+			if (type != MT76_RRO_Q_IND) {
+				q->head = q->ndesc - 1;
+				q->queued = q->ndesc - 1;
+			}
+		} else {
+			ret = mtk_wed_device_rx_ring_setup(wed, ring, q->regs, 0);
+			if (!ret)
+				q->wed_regs = wed->rx_ring[ring].reg_base;
+		}
 		break;
 	default:
 		ret = -EINVAL;
@@ -721,10 +805,25 @@ mt76_dma_alloc_queue(struct mt76_dev *dev, struct mt76_queue *q,
 	q->hw_idx = idx;
 
 	size = q->ndesc * sizeof(struct mt76_desc);
+	if (mt76_queue_is_rro_ind(q))
+		size = q->ndesc * sizeof(struct mt76_rro_desc);
+
 	q->desc = dmam_alloc_coherent(dev->dma_dev, size, &q->desc_dma, GFP_KERNEL);
 	if (!q->desc)
 		return -ENOMEM;
 
+	if (mt76_queue_is_rro_ind(q)) {
+		struct mt76_rro_ind *cmd;
+		int i;
+
+		q->rro_desc = (struct mt76_rro_desc *)(q->desc);
+		q->desc = NULL;
+		for (i = 0; i < q->ndesc; i++) {
+			cmd = (struct mt76_rro_ind *) &q->rro_desc[i];
+			cmd->magic_cnt = MT_DMA_IND_CMD_MAGIC_CNT - 1;
+		}
+	}
+
 	size = q->ndesc * sizeof(*q->entry);
 	q->entry = devm_kzalloc(dev->dev, size, GFP_KERNEL);
 	if (!q->entry)
@@ -734,8 +833,11 @@ mt76_dma_alloc_queue(struct mt76_dev *dev, struct mt76_queue *q,
 	if (ret)
 		return ret;
 
-	if (!mt76_queue_is_txfree(q))
-		mt76_dma_queue_reset(dev, q);
+	if (!mtk_wed_device_active(&dev->mmio.wed) ||
+	    (!mt76_queue_is_wed_txfree(q) &&
+	     !(mtk_wed_get_rx_capa(&dev->mmio.wed) &&
+	       q->flags & MT_QFLAG_RRO)))
+		mt76_dma_queue_reset(dev, q, false);
 
 	return 0;
 }
@@ -753,13 +855,13 @@ mt76_dma_rx_cleanup(struct mt76_dev *dev, struct mt76_queue *q)
 	spin_lock_bh(&q->lock);
 
 	do {
+		if (q->flags & MT_QFLAG_RRO)
+			break;
+
 		buf = mt76_dma_dequeue(dev, q, true, NULL, NULL, &more, NULL);
 		if (!buf)
 			break;
 
-		if (q->flags & MT_QFLAG_RRO)
-			continue;
-
 		skb_free_frag(buf);
 	} while (1);
 
@@ -770,8 +872,7 @@ mt76_dma_rx_cleanup(struct mt76_dev *dev, struct mt76_queue *q)
 
 	spin_unlock_bh(&q->lock);
 
-	if (((q->flags & MT_QFLAG_WED) &&
-	     FIELD_GET(MT_QFLAG_WED_TYPE, q->flags) == MT76_WED_Q_RX) ||
+	if (mt76_queue_is_wed_rx(q) ||
 	    (q->flags & MT_QFLAG_RRO))
 		return;
 
@@ -792,9 +893,13 @@ mt76_dma_rx_reset(struct mt76_dev *dev, enum mt76_rxq_id qid)
 	if (!q->ndesc)
 		return;
 
+	if (!q->desc)
+		goto done;
+
 	for (i = 0; i < q->ndesc; i++)
 		q->desc[i].ctrl = cpu_to_le32(MT_DMA_CTL_DMA_DONE);
 
+done:
 	mt76_dma_rx_cleanup(dev, q);
 
 	/* reset WED rx queues */
@@ -841,8 +946,8 @@ mt76_dma_rx_process(struct mt76_dev *dev, struct mt76_queue *q, int budget)
 	bool check_ddone = false;
 	bool more;
 
-	if (IS_ENABLED(CONFIG_NET_MEDIATEK_SOC_WED) &&
-	    q->flags == MT_WED_Q_TXFREE) {
+	if ((IS_ENABLED(CONFIG_NET_MEDIATEK_SOC_WED) &&
+	     q->flags == MT_WED_Q_TXFREE)) {
 		dma_idx = Q_READ(dev, q, dma_idx);
 		check_ddone = true;
 	}
@@ -1004,7 +1109,8 @@ void mt76_dma_cleanup(struct mt76_dev *dev)
 	mt76_for_each_q_rx(dev, i) {
 		struct mt76_queue *q = &dev->q_rx[i];
 
-		if (mt76_queue_is_wed_rx(q))
+		if (mtk_wed_device_active(&dev->mmio.wed) &&
+		    (q->flags & MT_QFLAG_RRO))
 			continue;
 
 		netif_napi_del(&dev->napi[i]);
@@ -1016,6 +1122,7 @@ void mt76_dma_cleanup(struct mt76_dev *dev)
 
 	if (mtk_wed_device_active(&dev->mmio.wed_ext))
 		mtk_wed_device_detach(&dev->mmio.wed_ext);
+
 	mt76_free_pending_txwi(dev);
 	mt76_free_pending_rxwi(dev);
 }
diff --git a/dma.h b/dma.h
index 1b090d78c..480370928 100644
--- a/dma.h
+++ b/dma.h
@@ -25,6 +25,13 @@
 #define MT_DMA_PPE_ENTRY		GENMASK(30, 16)
 #define MT_DMA_INFO_PPE_VLD		BIT(31)
 
+#define MT_DMA_CTL_PN_CHK_FAIL		BIT(13)
+#define MT_DMA_CTL_VER_MASK		BIT(7)
+
+#define MT_DMA_MAGIC_EN		BIT(13)
+
+#define MT_DMA_IND_CMD_MAGIC_CNT	8
+
 #define MT_DMA_HDR_LEN			4
 #define MT_RX_INFO_LEN			4
 #define MT_FCE_INFO_LEN			4
@@ -37,6 +44,11 @@ struct mt76_desc {
 	__le32 info;
 } __packed __aligned(4);
 
+struct mt76_rro_desc {
+	__le32 buf0;
+	__le32 buf1;
+} __packed __aligned(4);
+
 enum mt76_qsel {
 	MT_QSEL_MGMT,
 	MT_QSEL_HCCA,
diff --git a/mac80211.c b/mac80211.c
index 7cd9b6fc7..3070321d5 100644
--- a/mac80211.c
+++ b/mac80211.c
@@ -735,6 +735,7 @@ static void mt76_rx_release_amsdu(struct mt76_phy *phy, enum mt76_rxq_id q)
 			return;
 		}
 	}
+
 	__skb_queue_tail(&dev->rx_skb[q], skb);
 }
 
diff --git a/mt76.h b/mt76.h
index 3b2a658db..3954d01c5 100644
--- a/mt76.h
+++ b/mt76.h
@@ -48,6 +48,18 @@
 
 #define MT76_TOKEN_FREE_THR	64
 
+#define MT_QFLAG_RRO_RING	GENMASK(6, 5)
+#define MT_QFLAG_RRO_TYPE	GENMASK(8, 7)
+#define MT_QFLAG_RRO		BIT(9)
+#define MT_QFLAG_MAGIC		BIT(10)
+
+#define __MT_RRO_Q(_type, _n)	(MT_QFLAG_RRO | \
+				 FIELD_PREP(MT_QFLAG_RRO_TYPE, _type) | \
+				 FIELD_PREP(MT_QFLAG_RRO_RING, _n))
+#define MT_RRO_Q_DATA(_n)	__MT_RRO_Q(MT76_RRO_Q_DATA, _n)
+#define MT_RRO_Q_MSDU_PG(_n)	__MT_RRO_Q(MT76_RRO_Q_MSDU_PG, _n)
+#define MT_RRO_Q_IND		__MT_RRO_Q(MT76_RRO_Q_IND, 0)
+
 #define MT_QFLAG_WED_RING	GENMASK(1, 0)
 #define MT_QFLAG_WED_TYPE	GENMASK(3, 2)
 #define MT_QFLAG_WED		BIT(4)
@@ -82,6 +94,12 @@ enum mt76_wed_type {
 	MT76_WED_Q_RX,
 };
 
+enum mt76_RRO_type {
+	MT76_RRO_Q_DATA,
+	MT76_RRO_Q_MSDU_PG,
+	MT76_RRO_Q_IND,
+};
+
 struct mt76_bus_ops {
 	u32 (*rr)(struct mt76_dev *dev, u32 offset);
 	void (*wr)(struct mt76_dev *dev, u32 offset, u32 val);
@@ -128,6 +146,16 @@ enum mt76_rxq_id {
 	MT_RXQ_MAIN_WA,
 	MT_RXQ_BAND2,
 	MT_RXQ_BAND2_WA,
+	MT_RXQ_RRO_BAND0,
+	MT_RXQ_RRO_BAND1,
+	MT_RXQ_RRO_BAND2,
+	MT_RXQ_MSDU_PAGE_BAND0,
+	MT_RXQ_MSDU_PAGE_BAND1,
+	MT_RXQ_MSDU_PAGE_BAND2,
+	MT_RXQ_TXFREE_BAND0,
+	MT_RXQ_TXFREE_BAND1,
+	MT_RXQ_TXFREE_BAND2,
+	MT_RXQ_RRO_IND,
 	__MT_RXQ_MAX
 };
 
@@ -206,6 +234,7 @@ struct mt76_queue {
 	spinlock_t lock;
 	spinlock_t cleanup_lock;
 	struct mt76_queue_entry *entry;
+	struct mt76_rro_desc *rro_desc;
 	struct mt76_desc *desc;
 
 	u16 first;
@@ -219,8 +248,8 @@ struct mt76_queue {
 
 	u8 buf_offset;
 	u8 hw_idx;
-	u8 flags;
-
+	u8 magic_cnt;
+	u32 flags;
 	u32 wed_regs;
 
 	dma_addr_t desc_dma;
@@ -274,7 +303,7 @@ struct mt76_queue_ops {
 
 	void (*kick)(struct mt76_dev *dev, struct mt76_queue *q);
 
-	void (*reset_q)(struct mt76_dev *dev, struct mt76_queue *q);
+	void (*reset_q)(struct mt76_dev *dev, struct mt76_queue *q, bool skip);
 };
 
 enum mt76_phy_type {
@@ -375,6 +404,17 @@ struct mt76_txq {
 	bool aggr;
 };
 
+struct mt76_rro_ind {
+	u32 se_id	: 12;
+	u32 rsv		: 4;
+	u32 start_sn	: 12;
+	u32 ind_reason	: 4;
+	u32 ind_cnt	: 13;
+	u32 win_sz	: 3;
+	u32 rsv2	: 13;
+	u32 magic_cnt	: 3;
+};
+
 struct mt76_txwi_cache {
 	struct list_head list;
 	dma_addr_t dma_addr;
@@ -1629,12 +1669,19 @@ static inline bool mt76_queue_is_wed_rx(struct mt76_queue *q)
 	return (q->flags & MT_QFLAG_WED) &&
 	       FIELD_GET(MT_QFLAG_WED_TYPE, q->flags) == MT76_WED_Q_RX;
 }
-static inline bool mt76_queue_is_txfree(struct mt76_queue *q)
+
+static inline bool mt76_queue_is_wed_txfree(struct mt76_queue *q)
 {
 	return (q->flags & MT_QFLAG_WED) &&
 	       FIELD_GET(MT_QFLAG_WED_TYPE, q->flags) == MT76_WED_Q_TXFREE;
 }
 
+static inline bool mt76_queue_is_rro_ind(struct mt76_queue *q)
+{
+	return (q->flags & MT_QFLAG_RRO) &&
+	       FIELD_GET(MT_QFLAG_RRO_TYPE, q->flags) == MT76_RRO_Q_IND;
+}
+
 struct mt76_txwi_cache *
 mt76_token_release(struct mt76_dev *dev, int token, bool *wake);
 int mt76_token_consume(struct mt76_dev *dev, struct mt76_txwi_cache **ptxwi);
@@ -1653,10 +1700,14 @@ static inline void mt76_set_tx_blocked(struct mt76_dev *dev, bool blocked)
 static inline int
 mt76_token_get(struct mt76_dev *dev, struct mt76_txwi_cache **ptxwi)
 {
-	int token;
+	int token, start = 0;
+
+	if (mtk_wed_device_active(&dev->mmio.wed))
+		start = dev->mmio.wed.wlan.nbuf;
 
 	spin_lock_bh(&dev->token_lock);
-	token = idr_alloc(&dev->token, *ptxwi, 0, dev->token_size, GFP_ATOMIC);
+	token = idr_alloc(&dev->token, *ptxwi, start, start + dev->token_size,
+			  GFP_ATOMIC);
 	spin_unlock_bh(&dev->token_lock);
 
 	return token;
diff --git a/mt7996/dma.c b/mt7996/dma.c
index 3c8f617e0..309cc242e 100644
--- a/mt7996/dma.c
+++ b/mt7996/dma.c
@@ -64,6 +64,29 @@ static void mt7996_dma_config(struct mt7996_dev *dev)
 	RXQ_CONFIG(MT_RXQ_BAND2, WFDMA0, MT_INT_RX_DONE_BAND2, MT7996_RXQ_BAND2);
 	RXQ_CONFIG(MT_RXQ_BAND2_WA, WFDMA0, MT_INT_RX_DONE_WA_TRI, MT7996_RXQ_MCU_WA_TRI);
 
+	if (dev->rro_support) {
+		/* band0 */
+		RXQ_CONFIG(MT_RXQ_RRO_BAND0, WFDMA0, MT_INT_RX_DONE_RRO_BAND0,
+			   MT7996_RXQ_RRO_BAND0);
+		RXQ_CONFIG(MT_RXQ_MSDU_PAGE_BAND0, WFDMA0, MT_INT_RX_DONE_MSDU_PG_BAND0,
+			   MT7996_RXQ_MSDU_PG_BAND0);
+		RXQ_CONFIG(MT_RXQ_TXFREE_BAND0, WFDMA0, MT_INT_RX_TXFREE_MAIN,
+			   MT7996_RXQ_TXFREE0);
+		/* band1 */
+		RXQ_CONFIG(MT_RXQ_MSDU_PAGE_BAND1, WFDMA0, MT_INT_RX_DONE_MSDU_PG_BAND1,
+			   MT7996_RXQ_MSDU_PG_BAND1);
+		/* band2 */
+		RXQ_CONFIG(MT_RXQ_RRO_BAND2, WFDMA0, MT_INT_RX_DONE_RRO_BAND2,
+			   MT7996_RXQ_RRO_BAND2);
+		RXQ_CONFIG(MT_RXQ_MSDU_PAGE_BAND2, WFDMA0, MT_INT_RX_DONE_MSDU_PG_BAND2,
+			   MT7996_RXQ_MSDU_PG_BAND2);
+		RXQ_CONFIG(MT_RXQ_TXFREE_BAND2, WFDMA0, MT_INT_RX_TXFREE_TRI,
+			   MT7996_RXQ_TXFREE2);
+
+		RXQ_CONFIG(MT_RXQ_RRO_IND, WFDMA0, MT_INT_RX_DONE_RRO_IND,
+			   MT7996_RXQ_RRO_IND);
+	}
+
 	/* data tx queue */
 	TXQ_CONFIG(0, WFDMA0, MT_INT_TX_DONE_BAND0, MT7996_TXQ_BAND0);
 	TXQ_CONFIG(1, WFDMA0, MT_INT_TX_DONE_BAND1, MT7996_TXQ_BAND1);
@@ -102,6 +125,22 @@ static void __mt7996_dma_prefetch(struct mt7996_dev *dev, u32 ofs)
 	mt76_wr(dev, MT_RXQ_BAND1_CTRL(MT_RXQ_BAND2_WA) + ofs, PREFETCH(0x2));
 	mt76_wr(dev, MT_RXQ_BAND1_CTRL(MT_RXQ_MAIN) + ofs, PREFETCH(0x10));
 	mt76_wr(dev, MT_RXQ_BAND1_CTRL(MT_RXQ_BAND2) + ofs, PREFETCH(0x10));
+	if (dev->rro_support) {
+		mt76_wr(dev, MT_RXQ_BAND1_CTRL(MT_RXQ_RRO_BAND0) + ofs,
+			PREFETCH(0x10));
+		mt76_wr(dev, MT_RXQ_BAND1_CTRL(MT_RXQ_RRO_BAND2) + ofs,
+			PREFETCH(0x10));
+		mt76_wr(dev, MT_RXQ_BAND1_CTRL(MT_RXQ_MSDU_PAGE_BAND0) + ofs,
+			PREFETCH(0x4));
+		mt76_wr(dev, MT_RXQ_BAND1_CTRL(MT_RXQ_MSDU_PAGE_BAND1) + ofs,
+			PREFETCH(0x4));
+		mt76_wr(dev, MT_RXQ_BAND1_CTRL(MT_RXQ_MSDU_PAGE_BAND2) + ofs,
+			PREFETCH(0x4));
+		mt76_wr(dev, MT_RXQ_BAND1_CTRL(MT_RXQ_TXFREE_BAND0) + ofs,
+			PREFETCH(0x4));
+		mt76_wr(dev, MT_RXQ_BAND1_CTRL(MT_RXQ_TXFREE_BAND2) + ofs,
+			PREFETCH(0x4));
+	}
 #undef PREFETCH
 
 	mt76_set(dev, WF_WFDMA0_GLO_CFG_EXT1 + ofs, WF_WFDMA0_GLO_CFG_EXT1_CALC_MODE);
@@ -161,6 +200,7 @@ static void mt7996_dma_disable(struct mt7996_dev *dev, bool reset)
 
 void mt7996_dma_start(struct mt7996_dev *dev, bool reset, bool wed_reset)
 {
+	struct mtk_wed_device *wed = &dev->mt76.mmio.wed;
 	u32 hif1_ofs = 0;
 	u32 irq_mask;
 
@@ -169,11 +209,16 @@ void mt7996_dma_start(struct mt7996_dev *dev, bool reset, bool wed_reset)
 
 	/* enable WFDMA Tx/Rx */
 	if (!reset) {
-		mt76_set(dev, MT_WFDMA0_GLO_CFG,
-			 MT_WFDMA0_GLO_CFG_TX_DMA_EN |
-			 MT_WFDMA0_GLO_CFG_RX_DMA_EN |
-			 MT_WFDMA0_GLO_CFG_OMIT_TX_INFO |
-			 MT_WFDMA0_GLO_CFG_OMIT_RX_INFO_PFET2);
+		if (mtk_wed_device_active(wed) && mtk_wed_get_rx_capa(wed))
+			mt76_set(dev, MT_WFDMA0_GLO_CFG,
+				 MT_WFDMA0_GLO_CFG_TX_DMA_EN |
+				 MT_WFDMA0_GLO_CFG_OMIT_TX_INFO);
+		else
+			mt76_set(dev, MT_WFDMA0_GLO_CFG,
+				MT_WFDMA0_GLO_CFG_TX_DMA_EN |
+				MT_WFDMA0_GLO_CFG_RX_DMA_EN |
+				MT_WFDMA0_GLO_CFG_OMIT_TX_INFO |
+				MT_WFDMA0_GLO_CFG_OMIT_RX_INFO_PFET2);
 
 		if (dev->hif2)
 			mt76_set(dev, MT_WFDMA0_GLO_CFG + hif1_ofs,
@@ -195,14 +240,14 @@ void mt7996_dma_start(struct mt7996_dev *dev, bool reset, bool wed_reset)
 	if (mt7996_band_valid(dev, MT_BAND2))
 		irq_mask |= MT_INT_BAND2_RX_DONE;
 
-	if (mtk_wed_device_active(&dev->mt76.mmio.wed) && wed_reset) {
+	if (mtk_wed_device_active(wed) && wed_reset) {
 		u32 wed_irq_mask = irq_mask;
 
 		wed_irq_mask |= MT_INT_TX_DONE_BAND0 | MT_INT_TX_DONE_BAND1;
 
 		mt76_wr(dev, MT_INT_MASK_CSR, wed_irq_mask);
 
-		mtk_wed_device_start(&dev->mt76.mmio.wed, wed_irq_mask);
+		mtk_wed_device_start(wed, wed_irq_mask);
 	}
 
 	irq_mask = reset ? MT_INT_MCU_CMD : irq_mask;
@@ -296,7 +341,8 @@ static void mt7996_dma_enable(struct mt7996_dev *dev, bool reset)
 		/* fix hardware limitation, pcie1's rx ring3 is not available
 		 * so, redirect pcie0 rx ring3 interrupt to pcie1
 		 */
-		if (mtk_wed_device_active(&dev->mt76.mmio.wed) && dev->rro_support)
+		if (mtk_wed_device_active(&dev->mt76.mmio.wed) &&
+		    dev->rro_support)
 			mt76_set(dev, MT_WFDMA0_RX_INT_PCIE_SEL + hif1_ofs,
 				 MT_WFDMA0_RX_INT_SEL_RING6);
 		else
@@ -307,6 +353,78 @@ static void mt7996_dma_enable(struct mt7996_dev *dev, bool reset)
 	mt7996_dma_start(dev, reset, true);
 }
 
+int mt7996_dma_rro_init(struct mt7996_dev *dev)
+{
+	int ret;
+	u32 hif1_ofs = 0;
+	u32 wed_irq_mask;
+	struct mtk_wed_device *wed = &dev->mt76.mmio.wed;
+
+	if (dev->hif2)
+		hif1_ofs = MT_WFDMA0_PCIE1(0) - MT_WFDMA0(0);
+
+	/* ind cmd */
+	dev->mt76.q_rx[MT_RXQ_RRO_IND].flags = MT_RRO_Q_IND | MT_WED_Q_RX(0);
+	dev->mt76.q_rx[MT_RXQ_RRO_IND].flags |= MT_WED_Q_RX(0);
+	ret = mt76_queue_alloc(dev, &dev->mt76.q_rx[MT_RXQ_RRO_IND],
+			       MT_RXQ_ID(MT_RXQ_RRO_IND),
+			       MT7996_RX_RING_SIZE,
+			       0, MT_RXQ_RRO_IND_RING_BASE);
+	if (ret)
+		return ret;
+
+	/* rx msdu page queue for band0 */
+	dev->mt76.q_rx[MT_RXQ_MSDU_PAGE_BAND0].flags = MT_RRO_Q_MSDU_PG(0);
+	dev->mt76.q_rx[MT_RXQ_MSDU_PAGE_BAND0].flags |= MT_QFLAG_MAGIC;
+	dev->mt76.q_rx[MT_RXQ_MSDU_PAGE_BAND0].flags |= MT_WED_Q_RX(0);
+	ret = mt76_queue_alloc(dev, &dev->mt76.q_rx[MT_RXQ_MSDU_PAGE_BAND0],
+			       MT_RXQ_ID(MT_RXQ_MSDU_PAGE_BAND0),
+			       MT7996_RX_RING_SIZE,
+			       MT7996_RX_MSDU_PAGE_SIZE,
+			       MT_RXQ_RING_BASE(MT_RXQ_MSDU_PAGE_BAND0));
+	if (ret)
+		return ret;
+
+	if (mt7996_band_valid(dev, MT_BAND1)) {
+		/* rx msdu page queue for band1 */
+		dev->mt76.q_rx[MT_RXQ_MSDU_PAGE_BAND1].flags = MT_RRO_Q_MSDU_PG(1);
+		dev->mt76.q_rx[MT_RXQ_MSDU_PAGE_BAND1].flags |= MT_QFLAG_MAGIC;
+		dev->mt76.q_rx[MT_RXQ_MSDU_PAGE_BAND1].flags |= MT_WED_Q_RX(1);
+		ret = mt76_queue_alloc(dev, &dev->mt76.q_rx[MT_RXQ_MSDU_PAGE_BAND1],
+				       MT_RXQ_ID(MT_RXQ_MSDU_PAGE_BAND1),
+				       MT7996_RX_RING_SIZE,
+				       MT7996_RX_MSDU_PAGE_SIZE,
+				       MT_RXQ_RING_BASE(MT_RXQ_MSDU_PAGE_BAND1));
+		if (ret)
+			return ret;
+	}
+
+	if (mt7996_band_valid(dev, MT_BAND2)) {
+		/* rx msdu page queue for band2 */
+		dev->mt76.q_rx[MT_RXQ_MSDU_PAGE_BAND2].flags = MT_RRO_Q_MSDU_PG(2);
+		dev->mt76.q_rx[MT_RXQ_MSDU_PAGE_BAND2].flags |= MT_QFLAG_MAGIC;
+		dev->mt76.q_rx[MT_RXQ_MSDU_PAGE_BAND2].flags |= MT_WED_Q_RX(0);
+		ret = mt76_queue_alloc(dev, &dev->mt76.q_rx[MT_RXQ_MSDU_PAGE_BAND2],
+				       MT_RXQ_ID(MT_RXQ_MSDU_PAGE_BAND2),
+				       MT7996_RX_RING_SIZE,
+				       MT7996_RX_MSDU_PAGE_SIZE,
+				       MT_RXQ_RING_BASE(MT_RXQ_MSDU_PAGE_BAND2));
+		if (ret)
+			return ret;
+	}
+
+	wed_irq_mask = dev->mt76.mmio.irqmask |
+		       MT_INT_RRO_RX_DONE |
+		       MT_INT_TX_DONE_BAND2;
+
+	mt76_wr(dev, MT_INT_MASK_CSR, wed_irq_mask);
+
+	mtk_wed_device_start_hwrro(wed, wed_irq_mask, false);
+	mt7996_irq_enable(dev, wed_irq_mask);
+
+	return 0;
+}
+
 int mt7996_dma_init(struct mt7996_dev *dev)
 {
 	struct mtk_wed_device *wed = &dev->mt76.mmio.wed;
@@ -376,6 +494,9 @@ int mt7996_dma_init(struct mt7996_dev *dev)
 		return ret;
 
 	/* rx data queue for band0 and band1 */
+	if (mtk_wed_device_active(wed) && mtk_wed_get_rx_capa(wed))
+		dev->mt76.q_rx[MT_RXQ_MAIN].flags = MT_WED_Q_RX(0);
+
 	ret = mt76_queue_alloc(dev, &dev->mt76.q_rx[MT_RXQ_MAIN],
 			       MT_RXQ_ID(MT_RXQ_MAIN),
 			       MT7996_RX_RING_SIZE,
@@ -399,8 +520,9 @@ int mt7996_dma_init(struct mt7996_dev *dev)
 	if (mt7996_band_valid(dev, MT_BAND2)) {
 		/* rx data queue for band2 */
 		rx_base = MT_RXQ_RING_BASE(MT_RXQ_BAND2) + hif1_ofs;
-		if (mtk_wed_device_active(wed))
-			rx_base = MT_RXQ_RING_BASE(MT_RXQ_BAND2);
+		if (mtk_wed_device_active(wed_ext) && mtk_wed_get_rx_capa(wed_ext))
+			dev->mt76.q_rx[MT_RXQ_BAND2].flags = MT_WED_Q_RX(0) |
+							     MT_QFLAG_WED_EXT;
 
 		ret = mt76_queue_alloc(dev, &dev->mt76.q_rx[MT_RXQ_BAND2],
 				       MT_RXQ_ID(MT_RXQ_BAND2),
@@ -425,11 +547,12 @@ int mt7996_dma_init(struct mt7996_dev *dev)
 			return ret;
 	}
 
-
-	if (dev->rro_support) {
+	if (mtk_wed_device_active(wed) && mtk_wed_get_rx_capa(wed) &&
+	    dev->rro_support) {
 		/* rx rro data queue for band0 */
 		dev->mt76.q_rx[MT_RXQ_RRO_BAND0].flags = MT_RRO_Q_DATA(0);
 		dev->mt76.q_rx[MT_RXQ_RRO_BAND0].flags |= MT_QFLAG_MAGIC;
+		dev->mt76.q_rx[MT_RXQ_RRO_BAND0].flags |= MT_WED_Q_RX(0);
 		ret = mt76_queue_alloc(dev, &dev->mt76.q_rx[MT_RXQ_RRO_BAND0],
 				       MT_RXQ_ID(MT_RXQ_RRO_BAND0),
 				       MT7996_RX_RING_SIZE,
@@ -439,8 +562,7 @@ int mt7996_dma_init(struct mt7996_dev *dev)
 			return ret;
 
 		/* tx free notify event from WA for band0 */
-		if (mtk_wed_device_active(wed))
-			dev->mt76.q_rx[MT_RXQ_TXFREE_BAND0].flags = MT_WED_Q_TXFREE;
+		dev->mt76.q_rx[MT_RXQ_TXFREE_BAND0].flags = MT_WED_Q_TXFREE;
 		ret = mt76_queue_alloc(dev, &dev->mt76.q_rx[MT_RXQ_TXFREE_BAND0],
 				       MT_RXQ_ID(MT_RXQ_TXFREE_BAND0),
 				       MT7996_RX_MCU_RING_SIZE,
@@ -453,6 +575,7 @@ int mt7996_dma_init(struct mt7996_dev *dev)
 			/* rx rro data queue for band2 */
 			dev->mt76.q_rx[MT_RXQ_RRO_BAND2].flags = MT_RRO_Q_DATA(1);
 			dev->mt76.q_rx[MT_RXQ_RRO_BAND2].flags |= MT_QFLAG_MAGIC;
+			dev->mt76.q_rx[MT_RXQ_RRO_BAND2].flags |= MT_WED_Q_RX(1);
 			ret = mt76_queue_alloc(dev, &dev->mt76.q_rx[MT_RXQ_RRO_BAND2],
 					       MT_RXQ_ID(MT_RXQ_RRO_BAND2),
 					       MT7996_RX_RING_SIZE,
@@ -530,18 +653,18 @@ void mt7996_dma_reset(struct mt7996_dev *dev, bool force)
 
 	/* reset hw queues */
 	for (i = 0; i < __MT_TXQ_MAX; i++) {
-		mt76_queue_reset(dev, dev->mphy.q_tx[i]);
+		mt76_queue_reset(dev, dev->mphy.q_tx[i], false);
 		if (phy2)
-			mt76_queue_reset(dev, phy2->q_tx[i]);
+			mt76_queue_reset(dev, phy2->q_tx[i], false);
 		if (phy3)
-			mt76_queue_reset(dev, phy3->q_tx[i]);
+			mt76_queue_reset(dev, phy3->q_tx[i], false);
 	}
 
 	for (i = 0; i < __MT_MCUQ_MAX; i++)
-		mt76_queue_reset(dev, dev->mt76.q_mcu[i]);
+		mt76_queue_reset(dev, dev->mt76.q_mcu[i], false);
 
 	mt76_for_each_q_rx(&dev->mt76, i) {
-		mt76_queue_reset(dev, &dev->mt76.q_rx[i]);
+		mt76_queue_reset(dev, &dev->mt76.q_rx[i], false);
 	}
 
 	mt76_tx_status_check(&dev->mt76, true);
diff --git a/mt7996/init.c b/mt7996/init.c
index f2d43d3dc..3a749475e 100644
--- a/mt7996/init.c
+++ b/mt7996/init.c
@@ -502,8 +502,13 @@ void mt7996_mac_init(struct mt7996_dev *dev)
 
 	/* rro module init */
 	mt7996_mcu_set_rro(dev, UNI_RRO_SET_PLATFORM_TYPE, 2);
-	mt7996_mcu_set_rro(dev, UNI_RRO_SET_BYPASS_MODE, 3);
-	mt7996_mcu_set_rro(dev, UNI_RRO_SET_TXFREE_PATH, 1);
+	if (dev->rro_support) {
+		mt7996_mcu_set_rro(dev, UNI_RRO_SET_BYPASS_MODE, 1);
+		mt7996_mcu_set_rro(dev, UNI_RRO_SET_TXFREE_PATH, 0);
+	} else {
+		mt7996_mcu_set_rro(dev, UNI_RRO_SET_BYPASS_MODE, 3);
+		mt7996_mcu_set_rro(dev, UNI_RRO_SET_TXFREE_PATH, 1);
+	}
 
 	mt7996_mcu_wa_cmd(dev, MCU_WA_PARAM_CMD(SET),
 			  MCU_WA_PARAM_HW_PATH_HIF_VER,
@@ -656,6 +661,114 @@ void mt7996_wfsys_reset(struct mt7996_dev *dev)
 	msleep(20);
 }
 
+static int mt7996_rro_init(struct mt7996_dev *dev)
+{
+	struct mt7996_rro_addr *ptr;
+	struct mt7996_rro_cfg *rro = &dev->rro;
+	struct mtk_wed_device *wed = &dev->mt76.mmio.wed;
+	u32 size, val = 0, reg = MT_RRO_ADDR_ELEM_SEG_ADDR0;
+	int i, j;
+	void *buf;
+
+	for (i = 0; i < MT7996_RRO_BA_BITMAP_CR_CNT; i++) {
+		buf = dmam_alloc_coherent(dev->mt76.dma_dev,
+					  MT7996_BA_BITMAP_SZ_PER_CR,
+					  &rro->ba_bitmap_cache_pa[i],
+					  GFP_KERNEL);
+		if (!buf)
+			return -ENOMEM;
+
+		rro->ba_bitmap_cache_va[i] = buf;
+	}
+
+	rro->win_sz = MT7996_RRO_WIN_SIZE_MAX;
+	for (i = 0; i < MT7996_RRO_ADDR_ELEM_CR_CNT; i++) {
+		size = MT7996_RRO_SESSION_PER_CR *
+		       rro->win_sz * sizeof(struct mt7996_rro_addr);
+
+		buf = dmam_alloc_coherent(dev->mt76.dma_dev, size,
+					  &rro->addr_elem_alloc_pa[i],
+					  GFP_KERNEL);
+		if (!buf)
+			return -ENOMEM;
+		rro->addr_elem_alloc_va[i] = buf;
+
+		memset(rro->addr_elem_alloc_va[i], 0, size);
+
+		ptr = rro->addr_elem_alloc_va[i];
+		for (j = 0; j < MT7996_RRO_SESSION_PER_CR * rro->win_sz; j++, ptr++)
+			ptr->signature = 0xff;
+
+		wed->wlan.ind_cmd.addr_elem_phys[i] = rro->addr_elem_alloc_pa[i];
+	}
+
+	rro->particular_se_id = MT7996_RRO_SESSION_MAX;
+	size = rro->win_sz * sizeof(struct mt7996_rro_addr);
+	buf = dmam_alloc_coherent(dev->mt76.dma_dev, size,
+				  &rro->particular_session_pa,
+				  GFP_KERNEL);
+	if (!buf)
+		return -ENOMEM;
+
+	rro->particular_session_va = buf;
+	ptr = rro->particular_session_va;
+	for (j = 0; j < rro->win_sz; j++, ptr++)
+		ptr->signature = 0xff;
+
+	INIT_LIST_HEAD(&rro->pg_addr_cache);
+	for (i = 0; i < MT7996_RRO_MSDU_PG_HASH_SIZE; i++)
+		INIT_LIST_HEAD(&rro->pg_hash_head[i]);
+
+	/* rro hw init */
+	/* TODO: remove line after WM has set */
+	mt76_clear(dev, WF_RRO_AXI_MST_CFG, WF_RRO_AXI_MST_CFG_DIDX_OK);
+
+	/* setup BA bitmap cache address */
+	mt76_wr(dev, MT_RRO_BA_BITMAP_BASE0,
+		rro->ba_bitmap_cache_pa[0]);
+	mt76_wr(dev, MT_RRO_BA_BITMAP_BASE1, 0);
+	mt76_wr(dev, MT_RRO_BA_BITMAP_BASE_EXT0,
+		rro->ba_bitmap_cache_pa[1]);
+	mt76_wr(dev, MT_RRO_BA_BITMAP_BASE_EXT1, 0);
+
+	/* setup Address element address */
+	for (i = 0; i < MT7996_RRO_ADDR_ELEM_CR_CNT; i++) {
+		mt76_wr(dev, reg, rro->addr_elem_alloc_pa[i] >> 4);
+		reg += 4;
+	}
+
+	/* setup Address element address - separate address segment mode */
+	mt76_wr(dev, MT_RRO_ADDR_ARRAY_BASE1,
+		MT_RRO_ADDR_ARRAY_ELEM_ADDR_SEG_MODE);
+
+	wed->wlan.ind_cmd.win_size = ffs(rro->win_sz) - 6;
+	wed->wlan.ind_cmd.particular_sid = rro->particular_se_id;
+	wed->wlan.ind_cmd.particular_se_phys = rro->particular_session_pa;
+	wed->wlan.ind_cmd.se_group_nums = MT7996_RRO_ADDR_ELEM_CR_CNT;
+	wed->wlan.ind_cmd.ack_sn_addr = MT_RRO_ACK_SN_CTRL;
+
+	mt76_wr(dev, MT_RRO_IND_CMD_SIGNATURE_BASE0, 0x15010e00);
+	mt76_set(dev, MT_RRO_IND_CMD_SIGNATURE_BASE1,
+		 MT_RRO_IND_CMD_SIGNATURE_BASE1_EN);
+
+	/* particular session configure */
+	/* use max session idx + 1 as particular session id */
+	mt76_wr(dev, MT_RRO_PARTICULAR_CFG0,
+		rro->particular_session_pa);
+
+	val = FIELD_PREP(MT_RRO_PARTICULAR_SID,
+			 MT7996_RRO_SESSION_MAX);
+	val |= MT_RRO_PARTICULAR_CONFG_EN;
+	mt76_wr(dev, MT_RRO_PARTICULAR_CFG1, val);
+
+	/* interrupt enable */
+	mt76_wr(dev, MT_RRO_HOST_INT_ENA,
+		MT_RRO_HOST_INT_ENA_HOST_RRO_DONE_ENA);
+
+	/* rro ind cmd queue init */
+	return mt7996_dma_rro_init(dev);
+}
+
 static int mt7996_init_hardware(struct mt7996_dev *dev)
 {
 	int ret, idx;
@@ -687,6 +800,13 @@ static int mt7996_init_hardware(struct mt7996_dev *dev)
 	if (ret)
 		return ret;
 
+	if (mtk_wed_device_active(&dev->mt76.mmio.wed) &&
+	    dev->rro_support) {
+		ret = mt7996_rro_init(dev);
+		if (ret)
+			return ret;
+	}
+
 	ret = mt7996_eeprom_init(dev);
 	if (ret < 0)
 		return ret;
@@ -1131,10 +1251,10 @@ int mt7996_register_device(struct mt7996_dev *dev)
 	ieee80211_queue_work(mt76_hw(dev), &dev->init_work);
 
 	if (mtk_wed_device_active(&dev->mt76.mmio.wed_ext)) {
-		mt76_wr(dev, MT_INT1_MASK_CSR,
-			dev->mt76.mmio.irqmask|MT_INT_TX_DONE_BAND2);
+		mt76_wr(dev, MT_INT_PCIE1_MASK_CSR,
+			MT_INT_TRX_DONE_EXT);
 		mtk_wed_device_start(&dev->mt76.mmio.wed_ext,
-				     dev->mt76.mmio.irqmask |MT_INT_TX_DONE_BAND2);
+				     MT_INT_TRX_DONE_EXT);
 	}
 
 	dev->recovery.hw_init_done = true;
diff --git a/mt7996/mac.c b/mt7996/mac.c
index e57bdee21..08a32195b 100644
--- a/mt7996/mac.c
+++ b/mt7996/mac.c
@@ -393,8 +393,37 @@ mt7996_mac_fill_rx_rate(struct mt7996_dev *dev,
 	return 0;
 }
 
+static void
+mt7996_wed_check_ppe(struct mt7996_dev *dev, struct mt76_queue *q,
+		     struct mt7996_sta *msta, struct sk_buff *skb,
+		     u32 info)
+{
+	struct ieee80211_vif *vif;
+	struct wireless_dev *wdev;
+
+	if (!msta || !msta->vif)
+		return;
+
+	if (!mt76_queue_is_wed_rx(q))
+		return;
+
+	if (!(info & MT_DMA_INFO_PPE_VLD))
+		return;
+
+	vif = container_of((void *)msta->vif, struct ieee80211_vif,
+			   drv_priv);
+	wdev = ieee80211_vif_to_wdev(vif);
+	skb->dev = wdev->netdev;
+
+	mtk_wed_device_ppe_check(&dev->mt76.mmio.wed, skb,
+				 FIELD_GET(MT_DMA_PPE_CPU_REASON, info),
+				 FIELD_GET(MT_DMA_PPE_ENTRY, info));
+}
+
+
 static int
-mt7996_mac_fill_rx(struct mt7996_dev *dev, struct sk_buff *skb)
+mt7996_mac_fill_rx(struct mt7996_dev *dev, enum mt76_rxq_id q,
+		   struct sk_buff *skb, u32 *info)
 {
 	struct mt76_rx_status *status = (struct mt76_rx_status *)skb->cb;
 	struct mt76_phy *mphy = &dev->mt76.phy;
@@ -419,7 +448,10 @@ mt7996_mac_fill_rx(struct mt7996_dev *dev, struct sk_buff *skb)
 	u16 seq_ctrl = 0;
 	__le16 fc = 0;
 	int idx;
+	u8 hw_aggr = false;
+	struct mt7996_sta *msta = NULL;
 
+	hw_aggr = status->aggr;
 	memset(status, 0, sizeof(*status));
 
 	band_idx = FIELD_GET(MT_RXD1_NORMAL_BAND_IDX, rxd1);
@@ -446,8 +478,6 @@ mt7996_mac_fill_rx(struct mt7996_dev *dev, struct sk_buff *skb)
 	status->wcid = mt7996_rx_get_wcid(dev, idx, unicast);
 
 	if (status->wcid) {
-		struct mt7996_sta *msta;
-
 		msta = container_of(status->wcid, struct mt7996_sta, wcid);
 		spin_lock_bh(&dev->mt76.sta_poll_lock);
 		if (list_empty(&msta->wcid.poll_list))
@@ -656,13 +686,15 @@ mt7996_mac_fill_rx(struct mt7996_dev *dev, struct sk_buff *skb)
 #endif
 	} else {
 		status->flag |= RX_FLAG_8023;
+		mt7996_wed_check_ppe(dev, &dev->mt76.q_rx[q], msta, skb,
+				     *info);
 	}
 
 	if (rxv && mode >= MT_PHY_TYPE_HE_SU && mode < MT_PHY_TYPE_EHT_SU &&
 	    !(status->flag & RX_FLAG_8023))
 		mt76_connac3_mac_decode_he_radiotap(skb, rxv, mode);
 
-	if (!status->wcid || !ieee80211_is_data_qos(fc))
+	if (!status->wcid || !ieee80211_is_data_qos(fc) || hw_aggr)
 		return 0;
 
 	status->aggr = unicast &&
@@ -1406,7 +1438,7 @@ void mt7996_queue_rx_skb(struct mt76_dev *mdev, enum mt76_rxq_id q,
 		dev_kfree_skb(skb);
 		break;
 	case PKT_TYPE_NORMAL:
-		if (!mt7996_mac_fill_rx(dev, skb)) {
+		if (!mt7996_mac_fill_rx(dev, q, skb, info)) {
 			mt76_rx(&dev->mt76, q, skb);
 			return;
 		}
diff --git a/mt7996/mcu.c b/mt7996/mcu.c
index 5f18de031..2fc22d576 100644
--- a/mt7996/mcu.c
+++ b/mt7996/mcu.c
@@ -1063,7 +1063,7 @@ int mt7996_mcu_set_timing(struct mt7996_phy *phy, struct ieee80211_vif *vif)
 static int
 mt7996_mcu_sta_ba(struct mt76_dev *dev, struct mt76_vif *mvif,
 		  struct ieee80211_ampdu_params *params,
-		  bool enable, bool tx)
+		  bool enable, bool tx, bool rro_enable)
 {
 	struct mt76_wcid *wcid = (struct mt76_wcid *)params->sta->drv_priv;
 	struct sta_rec_ba_uni *ba;
@@ -1084,6 +1084,8 @@ mt7996_mcu_sta_ba(struct mt76_dev *dev, struct mt76_vif *mvif,
 	ba->ba_en = enable << params->tid;
 	ba->amsdu = params->amsdu;
 	ba->tid = params->tid;
+	if (rro_enable && !tx && enable)
+		ba->ba_rdd_rro = true;
 
 	return mt76_mcu_skb_send_msg(dev, skb,
 				     MCU_WMWA_UNI_CMD(STA_REC_UPDATE), true);
@@ -1101,7 +1103,7 @@ int mt7996_mcu_add_tx_ba(struct mt7996_dev *dev,
 		msta->wcid.amsdu = false;
 
 	return mt7996_mcu_sta_ba(&dev->mt76, &mvif->mt76, params,
-				 enable, true);
+				 enable, true, dev->rro_support);
 }
 
 int mt7996_mcu_add_rx_ba(struct mt7996_dev *dev,
@@ -1112,7 +1114,7 @@ int mt7996_mcu_add_rx_ba(struct mt7996_dev *dev,
 	struct mt7996_vif *mvif = msta->vif;
 
 	return mt7996_mcu_sta_ba(&dev->mt76, &mvif->mt76, params,
-				 enable, false);
+				 enable, false, dev->rro_support);
 }
 
 static void
diff --git a/mt7996/mmio.c b/mt7996/mmio.c
index ad2482ef2..1805d892f 100644
--- a/mt7996/mmio.c
+++ b/mt7996/mmio.c
@@ -336,7 +336,8 @@ int mt7996_mmio_wed_init(struct mt7996_dev *dev, void *pdev_ptr,
 
 	dev->rro_support = true;
 
-	hif1_ofs = MT_WFDMA0_PCIE1(0) - MT_WFDMA0(0);
+	if (dev->hif2)
+		hif1_ofs = MT_WFDMA0_PCIE1(0) - MT_WFDMA0(0);
 
 	if (hif2)
 		wed = &dev->mt76.mmio.wed_ext;
@@ -369,9 +370,15 @@ int mt7996_mmio_wed_init(struct mt7996_dev *dev, void *pdev_ptr,
 			wed->wlan.txfree_tbit = ffs(MT_INT_RX_DONE_WA_TRI) - 1;
 		}
 
+		wed->wlan.wpdma_rx_glo = wed->wlan.phy_base + hif1_ofs + MT_WFDMA0_GLO_CFG;
+		wed->wlan.wpdma_rx[0] = wed->wlan.phy_base + hif1_ofs +
+					MT_RXQ_RING_BASE(MT7996_RXQ_BAND2) +
+					MT7996_RXQ_BAND2 * MT_RING_SIZE;
+
 		wed->wlan.chip_id = 0x7991;
 		wed->wlan.tx_tbit[0] = ffs(MT_INT_TX_DONE_BAND2) - 1;
 	} else {
+		wed->wlan.hwrro = dev->rro_support; /* default on */
 		wed->wlan.wpdma_int = wed->wlan.phy_base + MT_INT_SOURCE_CSR;
 		wed->wlan.wpdma_mask = wed->wlan.phy_base + MT_INT_MASK_CSR;
 		wed->wlan.wpdma_tx = wed->wlan.phy_base + MT_TXQ_RING_BASE(0) +
@@ -383,13 +390,33 @@ int mt7996_mmio_wed_init(struct mt7996_dev *dev, void *pdev_ptr,
 					MT_RXQ_RING_BASE(MT7996_RXQ_BAND0) +
 					MT7996_RXQ_BAND0 * MT_RING_SIZE;
 
+		wed->wlan.wpdma_rx_rro[0] = wed->wlan.phy_base +
+					    MT_RXQ_RING_BASE(MT7996_RXQ_RRO_BAND0) +
+					    MT7996_RXQ_RRO_BAND0 * MT_RING_SIZE;
+		wed->wlan.wpdma_rx_rro[1] = wed->wlan.phy_base + hif1_ofs +
+					    MT_RXQ_RING_BASE(MT7996_RXQ_RRO_BAND2) +
+					    MT7996_RXQ_RRO_BAND2 * MT_RING_SIZE;
+		wed->wlan.wpdma_rx_pg = wed->wlan.phy_base +
+					MT_RXQ_RING_BASE(MT7996_RXQ_MSDU_PG_BAND0) +
+					MT7996_RXQ_MSDU_PG_BAND0 * MT_RING_SIZE;
+
 		wed->wlan.rx_nbuf = 65536;
 		wed->wlan.rx_npkt = 24576;
+		if (dev->hif2)
+			wed->wlan.rx_npkt += 8192;
+
 		wed->wlan.rx_size = SKB_WITH_OVERHEAD(MT_RX_BUF_SIZE);
 
 		wed->wlan.rx_tbit[0] = ffs(MT_INT_RX_DONE_BAND0) - 1;
 		wed->wlan.rx_tbit[1] = ffs(MT_INT_RX_DONE_BAND2) - 1;
 
+		wed->wlan.rro_rx_tbit[0] = ffs(MT_INT_RX_DONE_RRO_BAND0) - 1;
+		wed->wlan.rro_rx_tbit[1] = ffs(MT_INT_RX_DONE_RRO_BAND2) - 1;
+
+		wed->wlan.rx_pg_tbit[0] = ffs(MT_INT_RX_DONE_MSDU_PG_BAND0) - 1;
+		wed->wlan.rx_pg_tbit[1] = ffs(MT_INT_RX_DONE_MSDU_PG_BAND1) - 1;
+		wed->wlan.rx_pg_tbit[2] = ffs(MT_INT_RX_DONE_MSDU_PG_BAND2) - 1;
+
 		wed->wlan.tx_tbit[0] = ffs(MT_INT_TX_DONE_BAND0) - 1;
 		wed->wlan.tx_tbit[1] = ffs(MT_INT_TX_DONE_BAND1) - 1;
 		if (dev->rro_support) {
@@ -401,6 +428,8 @@ int mt7996_mmio_wed_init(struct mt7996_dev *dev, void *pdev_ptr,
 			wed->wlan.wpdma_txfree = wed->wlan.phy_base + MT_RXQ_RING_BASE(0) +
 						  MT7996_RXQ_MCU_WA_MAIN * MT_RING_SIZE;
 		}
+
+		dev->mt76.rx_token_size += wed->wlan.rx_npkt;
 	}
 
 	wed->wlan.nbuf = MT7996_TOKEN_SIZE;
@@ -417,8 +446,6 @@ int mt7996_mmio_wed_init(struct mt7996_dev *dev, void *pdev_ptr,
 	wed->wlan.release_rx_buf = mt7996_mmio_wed_release_rx_buf;
 	wed->wlan.update_wo_rx_stats = NULL;
 
-	dev->mt76.rx_token_size += wed->wlan.rx_npkt;
-
 	if (mtk_wed_device_attach(wed))
 		return 0;
 
@@ -530,12 +557,15 @@ static void mt7996_irq_tasklet(struct tasklet_struct *t)
 					       dev->mt76.mmio.irqmask);
 		if (intr1 & MT_INT_RX_TXFREE_EXT)
 			napi_schedule(&dev->mt76.napi[MT_RXQ_TXFREE_BAND2]);
+
+		if (intr1 & MT_INT_RX_DONE_BAND2_EXT)
+			napi_schedule(&dev->mt76.napi[MT_RXQ_BAND2]);
 	}
 
 	if (mtk_wed_device_active(wed)) {
 		mtk_wed_device_irq_set_mask(wed, 0);
 		intr = mtk_wed_device_irq_get(wed, dev->mt76.mmio.irqmask);
-		intr |= (intr1 & ~MT_INT_RX_TXFREE_EXT);
+		intr |= (intr1 & ~MT_INT_TRX_DONE_EXT);
 	} else {
 		mt76_wr(dev, MT_INT_MASK_CSR, 0);
 		if (dev->hif2)
@@ -581,10 +611,9 @@ static void mt7996_irq_tasklet(struct tasklet_struct *t)
 irqreturn_t mt7996_irq_handler(int irq, void *dev_instance)
 {
 	struct mt7996_dev *dev = dev_instance;
-	struct mtk_wed_device *wed = &dev->mt76.mmio.wed;
 
-	if (mtk_wed_device_active(wed))
-		mtk_wed_device_irq_set_mask(wed, 0);
+	if (mtk_wed_device_active(&dev->mt76.mmio.wed))
+		mtk_wed_device_irq_set_mask(&dev->mt76.mmio.wed, 0);
 	else
 		mt76_wr(dev, MT_INT_MASK_CSR, 0);
 
@@ -616,6 +645,7 @@ struct mt7996_dev *mt7996_mmio_probe(struct device *pdev,
 				SURVEY_INFO_TIME_RX |
 				SURVEY_INFO_TIME_BSS_RX,
 		.token_size = MT7996_TOKEN_SIZE,
+		.rx_token_size = MT7996_RX_TOKEN_SIZE,
 		.tx_prepare_skb = mt7996_tx_prepare_skb,
 		.tx_complete_skb = mt76_connac_tx_complete_skb,
 		.rx_skb = mt7996_queue_rx_skb,
diff --git a/mt7996/mt7996.h b/mt7996/mt7996.h
index d09358305..cf2a66df2 100644
--- a/mt7996/mt7996.h
+++ b/mt7996/mt7996.h
@@ -40,6 +40,7 @@
 #define MT7996_EEPROM_SIZE		7680
 #define MT7996_EEPROM_BLOCK_SIZE	16
 #define MT7996_TOKEN_SIZE		16384
+#define MT7996_RX_TOKEN_SIZE		16384
 #define MT7996_SW_TOKEN_SIZE		1024
 
 #define MT7996_CFEND_RATE_DEFAULT	0x49	/* OFDM 24M */
@@ -65,6 +66,24 @@
 #define MT7996_SKU_RATE_NUM		417
 #define MT7996_SKU_PATH_NUM		494
 
+#define MT7996_RRO_MSDU_PG_HASH_SIZE	127
+#define MT7996_RRO_SESSION_MAX		1024
+#define MT7996_RRO_WIN_SIZE_MAX		1024
+#define MT7996_RRO_ADDR_ELEM_CR_CNT	128
+#define MT7996_RRO_BA_BITMAP_CR_CNT	2
+#define MT7996_RRO_SESSION_PER_CR	(MT7996_RRO_SESSION_MAX /	\
+					 MT7996_RRO_ADDR_ELEM_CR_CNT)
+#define MT7996_BA_BITMAP_SZ_PER_SESSION	128
+#define MT7996_BA_BITMAP_SZ_PER_CR	((MT7996_RRO_SESSION_MAX *		\
+					 MT7996_BA_BITMAP_SZ_PER_SESSION) /	\
+					 MT7996_RRO_BA_BITMAP_CR_CNT)
+#define MT7996_SKB_TRUESIZE(x)		((x) +	\
+					 SKB_DATA_ALIGN(sizeof(struct skb_shared_info)))
+#define MT7996_RX_BUF_SIZE		MT7996_SKB_TRUESIZE(1800)
+#define MT7996_RX_MSDU_PAGE_SIZE	MT7996_SKB_TRUESIZE(128)
+
+#define MT7996_WED_RX_TOKEN_SIZE	32768
+
 struct mt7996_vif;
 struct mt7996_sta;
 struct mt7996_dfs_pulse;
@@ -109,6 +128,16 @@ enum mt7996_rxq_id {
 	MT7996_RXQ_BAND0 = 4,
 	MT7996_RXQ_BAND1 = 4,/* unused */
 	MT7996_RXQ_BAND2 = 5,
+	MT7996_RXQ_RRO_BAND0 = 8,
+	MT7996_RXQ_RRO_BAND1 = 8,/* unused */
+	MT7996_RXQ_RRO_BAND2 = 6,
+	MT7996_RXQ_MSDU_PG_BAND0 = 10,
+	MT7996_RXQ_MSDU_PG_BAND1 = 11,
+	MT7996_RXQ_MSDU_PG_BAND2 = 12,
+	MT7996_RXQ_TXFREE0 = 9,
+	MT7996_RXQ_TXFREE1 = 9,
+	MT7996_RXQ_TXFREE2 = 7,
+	MT7996_RXQ_RRO_IND = 0,
 };
 
 struct mt7996_twt_flow {
@@ -216,6 +245,31 @@ struct mt7996_air_monitor_ctrl {
 };
 #endif
 
+struct mt7996_rro_addr {
+	u32 head_pkt_l;
+	u32 head_pkt_h	: 4;
+	u32 seg_cnt	: 11;
+	u32 out_of_range: 1;
+	u32 rsv		: 8;
+	u32 signature	: 8;
+};
+
+struct mt7996_rro_cfg {
+	u32 ind_signature;
+	void *ba_bitmap_cache_va[MT7996_RRO_BA_BITMAP_CR_CNT];
+	void *addr_elem_alloc_va[MT7996_RRO_ADDR_ELEM_CR_CNT];
+	void *particular_session_va;
+	u32 particular_se_id;
+	dma_addr_t ba_bitmap_cache_pa[MT7996_RRO_BA_BITMAP_CR_CNT];
+	dma_addr_t addr_elem_alloc_pa[MT7996_RRO_ADDR_ELEM_CR_CNT];
+	dma_addr_t particular_session_pa;
+	u16 win_sz;
+
+	spinlock_t lock;
+	struct list_head pg_addr_cache;
+	struct list_head pg_hash_head[MT7996_RRO_MSDU_PG_HASH_SIZE];
+};
+
 struct mt7996_phy {
 	struct mt76_phy *mt76;
 	struct mt7996_dev *dev;
@@ -338,6 +392,9 @@ struct mt7996_dev {
 	bool flash_mode:1;
 	bool has_eht:1;
 
+	bool rro_support:1;
+	struct mt7996_rro_cfg rro;
+
 	bool testmode_enable;
 	bool bin_file_mode;
 	u8 eeprom_mode;
@@ -662,6 +719,7 @@ int mt7996_tx_prepare_skb(struct mt76_dev *mdev, void *txwi_ptr,
 			  struct ieee80211_sta *sta,
 			  struct mt76_tx_info *tx_info);
 void mt7996_tx_token_put(struct mt7996_dev *dev);
+int mt7996_dma_rro_init(struct mt7996_dev *dev);
 void mt7996_queue_rx_skb(struct mt76_dev *mdev, enum mt76_rxq_id q,
 			 struct sk_buff *skb, u32 *info);
 bool mt7996_rx_check(struct mt76_dev *mdev, void *data, int len);
diff --git a/mt7996/pci.c b/mt7996/pci.c
index 085408571..9a134fcab 100644
--- a/mt7996/pci.c
+++ b/mt7996/pci.c
@@ -124,6 +124,8 @@ static int mt7996_pci_probe(struct pci_dev *pdev,
 	mdev = &dev->mt76;
 	mt7996_wfsys_reset(dev);
 	hif2 = mt7996_pci_init_hif2(pdev);
+	if (hif2)
+		dev->hif2 = hif2;
 
 	ret = mt7996_mmio_wed_init(dev, pdev, false, &irq);
 	if (ret < 0)
@@ -148,7 +150,6 @@ static int mt7996_pci_probe(struct pci_dev *pdev,
 
 	if (hif2) {
 		hif2_dev = container_of(hif2->dev, struct pci_dev, dev);
-		dev->hif2 = hif2;
 
 		ret = mt7996_mmio_wed_init(dev, hif2_dev, true, &irq);
 		if (ret < 0)
diff --git a/mt7996/regs.h b/mt7996/regs.h
index ca7c2a811..c34357c3e 100644
--- a/mt7996/regs.h
+++ b/mt7996/regs.h
@@ -39,6 +39,40 @@ enum base_rev {
 
 #define __BASE(_id, _band)			(dev->reg.base[(_id)].band_base[(_band)])
 
+
+/* RRO TOP */
+#define MT_RRO_TOP_BASE				0xA000
+#define MT_RRO_TOP(ofs)				(MT_RRO_TOP_BASE + (ofs))
+
+#define MT_RRO_BA_BITMAP_BASE0			MT_RRO_TOP(0x8)
+#define MT_RRO_BA_BITMAP_BASE1			MT_RRO_TOP(0xC)
+#define WF_RRO_AXI_MST_CFG			MT_RRO_TOP(0xB8)
+#define WF_RRO_AXI_MST_CFG_DIDX_OK		BIT(12)
+#define MT_RRO_ADDR_ARRAY_BASE1			MT_RRO_TOP(0x34)
+#define MT_RRO_ADDR_ARRAY_ELEM_ADDR_SEG_MODE	BIT(31)
+
+#define MT_RRO_IND_CMD_SIGNATURE_BASE0		MT_RRO_TOP(0x38)
+#define MT_RRO_IND_CMD_SIGNATURE_BASE1		MT_RRO_TOP(0x3C)
+#define MT_RRO_IND_CMD_0_CTRL0			MT_RRO_TOP(0x40)
+#define MT_RRO_IND_CMD_SIGNATURE_BASE1_EN	BIT(31)
+
+#define MT_RRO_PARTICULAR_CFG0			MT_RRO_TOP(0x5C)
+#define MT_RRO_PARTICULAR_CFG1			MT_RRO_TOP(0x60)
+#define MT_RRO_PARTICULAR_CONFG_EN		BIT(31)
+#define MT_RRO_PARTICULAR_SID			GENMASK(30, 16)
+
+#define MT_RRO_BA_BITMAP_BASE_EXT0		MT_RRO_TOP(0x70)
+#define MT_RRO_BA_BITMAP_BASE_EXT1		MT_RRO_TOP(0x74)
+#define MT_RRO_HOST_INT_ENA			MT_RRO_TOP(0x204)
+#define MT_RRO_HOST_INT_ENA_HOST_RRO_DONE_ENA   BIT(0)
+
+#define MT_RRO_ADDR_ELEM_SEG_ADDR0		MT_RRO_TOP(0x400)
+
+#define MT_RRO_ACK_SN_CTRL			MT_RRO_TOP(0x50)
+#define MT_RRO_ACK_SN_CTRL_SN_MASK		GENMASK(27, 16)
+#define MT_RRO_ACK_SN_CTRL_SESSION_MASK		GENMASK(11, 0)
+
+
 #define MT_MCU_INT_EVENT			0x2108
 #define MT_MCU_INT_EVENT_DMA_STOPPED		BIT(0)
 #define MT_MCU_INT_EVENT_DMA_INIT		BIT(1)
@@ -407,6 +441,7 @@ enum base_rev {
 #define MT_MCUQ_RING_BASE(q)			(MT_Q_BASE(q) + 0x300)
 #define MT_TXQ_RING_BASE(q)			(MT_Q_BASE(__TXQ(q)) + 0x300)
 #define MT_RXQ_RING_BASE(q)			(MT_Q_BASE(__RXQ(q)) + 0x500)
+#define MT_RXQ_RRO_IND_RING_BASE		MT_RRO_TOP(0x40)
 
 #define MT_MCUQ_EXT_CTRL(q)			(MT_Q_BASE(q) +	0x600 +	\
 						 MT_MCUQ_ID(q) * 0x4)
@@ -432,8 +467,19 @@ enum base_rev {
 #define MT_INT_RX_TXFREE_MAIN			BIT(17)
 #define MT_INT_RX_TXFREE_TRI			BIT(15)
 #define MT_INT_MCU_CMD				BIT(29)
+
+#define MT_INT_RX_DONE_BAND2_EXT		BIT(23)
 #define MT_INT_RX_TXFREE_EXT			BIT(26)
 
+#define MT_INT_RX_DONE_RRO_BAND0		BIT(16)
+#define MT_INT_RX_DONE_RRO_BAND1		BIT(16)
+#define MT_INT_RX_DONE_RRO_BAND2		BIT(14)
+#define MT_INT_RX_DONE_RRO_IND			BIT(11)
+#define MT_INT_RX_DONE_MSDU_PG_BAND0		BIT(18)
+#define MT_INT_RX_DONE_MSDU_PG_BAND1		BIT(19)
+#define MT_INT_RX_DONE_MSDU_PG_BAND2		BIT(23)
+
+
 #define MT_INT_RX(q)				(dev->q_int_mask[__RXQ(q)])
 #define MT_INT_TX_MCU(q)			(dev->q_int_mask[(q)])
 
@@ -441,20 +487,31 @@ enum base_rev {
 						 MT_INT_RX(MT_RXQ_MCU_WA))
 
 #define MT_INT_BAND0_RX_DONE			(MT_INT_RX(MT_RXQ_MAIN) |	\
-						 MT_INT_RX(MT_RXQ_MAIN_WA))
+						 MT_INT_RX(MT_RXQ_MAIN_WA) |	\
+						 MT_INT_RX(MT_RXQ_TXFREE_BAND0))
 
 #define MT_INT_BAND1_RX_DONE			(MT_INT_RX(MT_RXQ_BAND1) |	\
 						 MT_INT_RX(MT_RXQ_BAND1_WA) |	\
-						 MT_INT_RX(MT_RXQ_MAIN_WA))
+						 MT_INT_RX(MT_RXQ_MAIN_WA) |	\
+						 MT_INT_RX(MT_RXQ_TXFREE_BAND0))
 
 #define MT_INT_BAND2_RX_DONE			(MT_INT_RX(MT_RXQ_BAND2) |	\
 						 MT_INT_RX(MT_RXQ_BAND2_WA) |	\
-						 MT_INT_RX(MT_RXQ_MAIN_WA))
+						 MT_INT_RX(MT_RXQ_MAIN_WA) |	\
+						 MT_INT_RX(MT_RXQ_TXFREE_BAND0))
+
+#define MT_INT_RRO_RX_DONE			(MT_INT_RX(MT_RXQ_RRO_BAND0) |		\
+						 MT_INT_RX(MT_RXQ_RRO_BAND1) |		\
+						 MT_INT_RX(MT_RXQ_RRO_BAND2) |		\
+						 MT_INT_RX(MT_RXQ_MSDU_PAGE_BAND0) |	\
+						 MT_INT_RX(MT_RXQ_MSDU_PAGE_BAND1) |	\
+						 MT_INT_RX(MT_RXQ_MSDU_PAGE_BAND2))
 
 #define MT_INT_RX_DONE_ALL			(MT_INT_RX_DONE_MCU |		\
 						 MT_INT_BAND0_RX_DONE |		\
 						 MT_INT_BAND1_RX_DONE |		\
-						 MT_INT_BAND2_RX_DONE)
+						 MT_INT_BAND2_RX_DONE |		\
+						 MT_INT_RRO_RX_DONE)
 
 #define MT_INT_TX_DONE_FWDL			BIT(26)
 #define MT_INT_TX_DONE_MCU_WM			BIT(27)
@@ -463,6 +520,10 @@ enum base_rev {
 #define MT_INT_TX_DONE_BAND1			BIT(31)
 #define MT_INT_TX_DONE_BAND2			BIT(15)
 
+#define MT_INT_TRX_DONE_EXT			(MT_INT_TX_DONE_BAND2 |	\
+						 MT_INT_RX_DONE_BAND2_EXT |	\
+						 MT_INT_RX_TXFREE_EXT)
+
 #define MT_INT_TX_DONE_MCU			(MT_INT_TX_MCU(MT_MCUQ_WA) |	\
 						 MT_INT_TX_MCU(MT_MCUQ_WM) |	\
 						 MT_INT_TX_MCU(MT_MCUQ_FWDL))
-- 
2.39.2