diff --git a/dp/inc/cdp_txrx_cmn_struct.h b/dp/inc/cdp_txrx_cmn_struct.h
index 55af6e0983..807574665b 100644
--- a/dp/inc/cdp_txrx_cmn_struct.h
+++ b/dp/inc/cdp_txrx_cmn_struct.h
@@ -1165,6 +1165,7 @@ typedef union cdp_config_param_t {
 	bool cdp_psoc_param_en_rate_stats;
 	int cdp_psoc_param_en_nss_cfg;
 	int cdp_psoc_param_preferred_hw_mode;
+	bool cdp_psoc_param_pext_stats;
 } cdp_config_param_type;
 
 /**
@@ -1267,11 +1268,13 @@ enum cdp_vdev_param_type {
  * @CDP_ENABLE_RATE_STATS: set rate stats enable/disable
  * @CDP_SET_NSS_CFG: set nss cfg
  * @CDP_SET_PREFERRED_HW_MODE: set preferred hw mode
+ * @CDP_CFG_PEER_EXT_STATS: Peer extended stats mode.
  */
 enum cdp_psoc_param_type {
 	CDP_ENABLE_RATE_STATS,
 	CDP_SET_NSS_CFG,
 	CDP_SET_PREFERRED_HW_MODE,
+	CDP_CFG_PEER_EXT_STATS,
 };
 
 #define TXRX_FW_STATS_TXSTATS                     1
diff --git a/dp/wifi3.0/dp_main.c b/dp/wifi3.0/dp_main.c
index cd8e5b5fa9..806e97ba07 100644
--- a/dp/wifi3.0/dp_main.c
+++ b/dp/wifi3.0/dp_main.c
@@ -8013,6 +8013,34 @@ static QDF_STATUS dp_set_pdev_param(struct cdp_soc_t *cdp_soc, uint8_t pdev_id,
 	return QDF_STATUS_SUCCESS;
 }
 
+#ifdef QCA_PEER_EXT_STATS
+static void dp_rx_update_peer_delay_stats(struct dp_soc *soc,
+					  qdf_nbuf_t nbuf)
+{
+	struct dp_peer *peer = NULL;
+	uint16_t peer_id, ring_id;
+	uint8_t tid = qdf_nbuf_get_tid_val(nbuf);
+	struct cdp_peer_ext_stats *pext_stats = NULL;
+
+	peer_id = QDF_NBUF_CB_RX_PEER_ID(nbuf);
+	if (peer_id > soc->max_peers)
+		return;
+
+	peer = dp_peer_find_by_id(soc, peer_id);
+	if (qdf_unlikely(!peer) || qdf_unlikely(!peer->pext_stats))
+		return;
+
+	pext_stats = peer->pext_stats;
+	ring_id = QDF_NBUF_CB_RX_CTX_ID(nbuf);
+	dp_rx_compute_tid_delay(&pext_stats->delay_stats[tid][ring_id], nbuf);
+}
+#else
+static inline void dp_rx_update_peer_delay_stats(struct dp_soc *soc,
+						 qdf_nbuf_t nbuf)
+{
+}
+#endif
+
 /*
  * dp_calculate_delay_stats: function to get rx delay stats
  * @cdp_soc: DP soc handle
@@ -8025,14 +8053,23 @@ static QDF_STATUS
 dp_calculate_delay_stats(struct cdp_soc_t *cdp_soc, uint8_t vdev_id,
 			 qdf_nbuf_t nbuf)
 {
+	struct dp_soc *soc = (struct dp_soc *)cdp_soc;
 	struct dp_vdev *vdev =
 		dp_get_vdev_from_soc_vdev_id_wifi3((struct dp_soc *)cdp_soc,
 						   vdev_id);
-	if (vdev) {
+
+	if (!vdev)
+		return QDF_STATUS_SUCCESS;
+
+	if (vdev->pdev->delay_stats_flag) {
 		dp_rx_compute_delay(vdev, nbuf);
-		return QDF_STATUS_E_FAILURE;
+		return QDF_STATUS_SUCCESS;
 	}
 
+	/*
+	 * Update the per peer delay stats
+	 */
+	dp_rx_update_peer_delay_stats(soc, nbuf);
 	return QDF_STATUS_SUCCESS;
 }
 
@@ -8255,6 +8292,21 @@ static QDF_STATUS dp_get_psoc_param(struct cdp_soc_t *cdp_soc,
 				    enum cdp_psoc_param_type param,
 				    cdp_config_param_type *val)
 {
+	struct dp_soc *soc = (struct dp_soc *)cdp_soc;
+
+	if (!soc)
+		return QDF_STATUS_E_FAILURE;
+
+	switch (param) {
+	case CDP_CFG_PEER_EXT_STATS:
+		val->cdp_psoc_param_pext_stats =
+			wlan_cfg_is_peer_ext_stats_enabled(soc->wlan_cfg_ctx);
+		break;
+	default:
+		dp_warn("Invalid param");
+		break;
+	}
+
 	return QDF_STATUS_SUCCESS;
 }
 
diff --git a/dp/wifi3.0/dp_rx.c b/dp/wifi3.0/dp_rx.c
index 149f6fed8b..71bf680d79 100644
--- a/dp/wifi3.0/dp_rx.c
+++ b/dp/wifi3.0/dp_rx.c
@@ -32,6 +32,7 @@
 #ifdef FEATURE_WDS
 #include "dp_txrx_wds.h"
 #endif
+#include "dp_hist.h"
 
 #ifdef ATH_RX_PRI_SAVE
 #define DP_RX_TID_SAVE(_nbuf, _tid) \
@@ -1189,6 +1190,24 @@ qdf_nbuf_t dp_rx_sg_create(qdf_nbuf_t nbuf)
 	return parent;
 }
 
+#ifdef QCA_PEER_EXT_STATS
+/*
+ * dp_rx_compute_tid_delay - Computer per TID delay stats
+ * @peer: DP soc context
+ * @nbuf: NBuffer
+ *
+ * Return: Void
+ */
+void dp_rx_compute_tid_delay(struct cdp_delay_tid_stats *stats,
+			     qdf_nbuf_t nbuf)
+{
+	struct cdp_delay_rx_stats  *rx_delay = &stats->rx_delay;
+	uint32_t to_stack = qdf_nbuf_get_timedelta_ms(nbuf);
+
+	dp_hist_update_stats(&rx_delay->to_stack_delay, to_stack);
+}
+#endif /* QCA_PEER_EXT_STATS */
+
 /**
  * dp_rx_compute_delay() - Compute and fill in all timestamps
  *				to pass in correct fields
@@ -2283,7 +2302,9 @@ done:
 
 		rx_pdev = vdev->pdev;
 		DP_RX_TID_SAVE(nbuf, tid);
-		if (qdf_unlikely(rx_pdev->delay_stats_flag))
+		if (qdf_unlikely(rx_pdev->delay_stats_flag) ||
+		    qdf_unlikely(wlan_cfg_is_peer_ext_stats_enabled(
+				 soc->wlan_cfg_ctx)))
 			qdf_nbuf_set_timestamp(nbuf);
 
 		ring_id = QDF_NBUF_CB_RX_CTX_ID(nbuf);
diff --git a/dp/wifi3.0/dp_rx.h b/dp/wifi3.0/dp_rx.h
index 6747a86d0f..b09e23bedf 100644
--- a/dp/wifi3.0/dp_rx.h
+++ b/dp/wifi3.0/dp_rx.h
@@ -1233,6 +1233,12 @@ void dp_rx_dump_info_and_assert(struct dp_soc *soc,
 				struct dp_rx_desc *rx_desc);
 
 void dp_rx_compute_delay(struct dp_vdev *vdev, qdf_nbuf_t nbuf);
+
+#ifdef QCA_PEER_EXT_STATS
+void dp_rx_compute_tid_delay(struct cdp_delay_tid_stats *stats,
+			     qdf_nbuf_t nbuf);
+#endif /* QCA_PEER_EXT_STATS */
+
 #ifdef RX_DESC_DEBUG_CHECK
 /**
  * dp_rx_desc_check_magic() - check the magic value in dp_rx_desc
diff --git a/dp/wifi3.0/dp_tx.c b/dp/wifi3.0/dp_tx.c
index 58ad9eb244..7bca13caf2 100644
--- a/dp/wifi3.0/dp_tx.c
+++ b/dp/wifi3.0/dp_tx.c
@@ -40,6 +40,7 @@
 #ifdef ATH_SUPPORT_IQUE
 #include "dp_txrx_me.h"
 #endif
+#include "dp_hist.h"
 
 
 /* TODO Add support in TSO */
@@ -1225,8 +1226,10 @@ static QDF_STATUS dp_tx_hw_enqueue(struct dp_soc *soc, struct dp_vdev *vdev,
 	if (tx_desc->flags & DP_TX_DESC_FLAG_MESH)
 		hal_tx_desc_set_mesh_en(soc->hal_soc, hal_tx_desc_cached, 1);
 
-	if (qdf_unlikely(vdev->pdev->delay_stats_flag))
-		tx_desc->timestamp = qdf_ktime_to_ms(qdf_ktime_get());
+	if (qdf_unlikely(vdev->pdev->delay_stats_flag) ||
+	    qdf_unlikely(wlan_cfg_is_peer_ext_stats_enabled(
+			 soc->wlan_cfg_ctx)))
+		tx_desc->timestamp = qdf_ktime_to_ms(qdf_ktime_real_get());
 
 	dp_verbose_debug("length:%d , type = %d, dma_addr %llx, offset %d desc id %u",
 			 tx_desc->length, type, (uint64_t)tx_desc->dma_addr,
@@ -2986,6 +2989,84 @@ void dp_tx_comp_fill_tx_completion_stats(struct dp_tx_desc_s *tx_desc,
 
 #endif
 
+#ifdef QCA_PEER_EXT_STATS
+/*
+ * dp_tx_compute_tid_delay() - Compute per TID delay
+ * @stats: Per TID delay stats
+ * @tx_desc: Software Tx descriptor
+ *
+ * Compute the software enqueue and hw enqueue delays and
+ * update the respective histograms
+ *
+ * Return: void
+ */
+static void dp_tx_compute_tid_delay(struct cdp_delay_tid_stats *stats,
+				    struct dp_tx_desc_s *tx_desc)
+{
+	struct cdp_delay_tx_stats  *tx_delay = &stats->tx_delay;
+	int64_t current_timestamp, timestamp_ingress, timestamp_hw_enqueue;
+	uint32_t sw_enqueue_delay, fwhw_transmit_delay;
+
+	current_timestamp = qdf_ktime_to_ms(qdf_ktime_real_get());
+	timestamp_ingress = qdf_nbuf_get_timestamp(tx_desc->nbuf);
+	timestamp_hw_enqueue = tx_desc->timestamp;
+	sw_enqueue_delay = (uint32_t)(timestamp_hw_enqueue - timestamp_ingress);
+	fwhw_transmit_delay = (uint32_t)(current_timestamp -
+					 timestamp_hw_enqueue);
+
+	/*
+	 * Update the Tx software enqueue delay and HW enque-Completion delay.
+	 */
+	dp_hist_update_stats(&tx_delay->tx_swq_delay, sw_enqueue_delay);
+	dp_hist_update_stats(&tx_delay->hwtx_delay, fwhw_transmit_delay);
+}
+
+/*
+ * dp_tx_update_peer_ext_stats() - Update the peer extended stats
+ * @peer: DP peer context
+ * @tx_desc: Tx software descriptor
+ * @tid: Transmission ID
+ * @ring_id: Rx CPU context ID/CPU_ID
+ *
+ * Update the peer extended stats. These are enhanced other
+ * delay stats per msdu level.
+ *
+ * Return: void
+ */
+static void dp_tx_update_peer_ext_stats(struct dp_peer *peer,
+					struct dp_tx_desc_s *tx_desc,
+					uint8_t tid, uint8_t ring_id)
+{
+	struct dp_pdev *pdev = peer->vdev->pdev;
+	struct dp_soc *soc = NULL;
+	struct cdp_peer_ext_stats *pext_stats = NULL;
+
+	soc = pdev->soc;
+	if (qdf_likely(!wlan_cfg_is_peer_ext_stats_enabled(soc->wlan_cfg_ctx)))
+		return;
+
+	pext_stats = peer->pext_stats;
+
+	qdf_assert(pext_stats);
+	qdf_assert(ring < CDP_MAX_TXRX_CTX);
+
+	/*
+	 * For non-TID packets use the TID 9
+	 */
+	if (qdf_unlikely(tid >= CDP_MAX_DATA_TIDS))
+		tid = CDP_MAX_DATA_TIDS - 1;
+
+	dp_tx_compute_tid_delay(&pext_stats->delay_stats[tid][ring_id],
+				tx_desc);
+}
+#else
+static inline void dp_tx_update_peer_ext_stats(struct dp_peer *peer,
+					       struct dp_tx_desc_s *tx_desc,
+					       uint8_t tid, uint8_t ring_id)
+{
+}
+#endif
+
 /**
  * dp_tx_compute_delay() - Compute and fill in all timestamps
  *				to pass in correct fields
@@ -3006,7 +3087,7 @@ static void dp_tx_compute_delay(struct dp_vdev *vdev,
 	if (qdf_likely(!vdev->pdev->delay_stats_flag))
 		return;
 
-	current_timestamp = qdf_ktime_to_ms(qdf_ktime_get());
+	current_timestamp = qdf_ktime_to_ms(qdf_ktime_real_get());
 	timestamp_ingress = qdf_nbuf_get_timestamp(tx_desc->nbuf);
 	timestamp_hw_enqueue = tx_desc->timestamp;
 	sw_enqueue_delay = (uint32_t)(timestamp_hw_enqueue - timestamp_ingress);
@@ -3356,7 +3437,7 @@ dp_tx_comp_process_desc(struct dp_soc *soc,
 	 * scatter gather packets
 	 */
 	if (qdf_unlikely(!!desc->pdev->latency_capture_enable)) {
-		time_latency = (qdf_ktime_to_ms(qdf_ktime_get()) -
+		time_latency = (qdf_ktime_to_ms(qdf_ktime_real_get()) -
 				desc->timestamp);
 	}
 	if (!(desc->msdu_ext_desc)) {
@@ -3531,6 +3612,7 @@ void dp_tx_comp_process_tx_status(struct dp_soc *soc,
 	}
 
 	dp_tx_update_peer_stats(tx_desc, ts, peer, ring_id);
+	dp_tx_update_peer_ext_stats(peer, tx_desc, ts->tid, ring_id);
 
 #ifdef QCA_SUPPORT_RDK_STATS
 	if (soc->wlanstats_enabled)
@@ -3723,7 +3805,6 @@ void dp_tx_process_htt_completion(struct dp_tx_desc_s *tx_desc, uint8_t *status,
 		}
 
 		peer = dp_peer_find_by_id(soc, ts.peer_id);
-
 		if (qdf_likely(peer))
 			dp_peer_unref_del_find_by_id(peer);