diff --git a/sys/dev/mlx5/mlx5_en/mlx5_en_rx.c b/sys/dev/mlx5/mlx5_en/mlx5_en_rx.c
--- a/sys/dev/mlx5/mlx5_en/mlx5_en_rx.c
+++ b/sys/dev/mlx5/mlx5_en/mlx5_en_rx.c
@@ -498,15 +498,38 @@
 	}
 }
 
+static void
+mlx5e_input_mbs(struct mlx5e_rq *rq, struct mbuf **mbs, int c)
+{
+	int i;
+
+	mtx_unlock(&rq->mtx);
+
+	for (i = 0; i < c; i++) {
+#if !defined(HAVE_TCP_LRO_RX)
+		tcp_lro_queue_mbuf(&rq->lro, mbs[i]);
+#else
+		if (mb->m_pkthdr.csum_flags == 0 ||
+		    (if_getcapenable(rq->ifp) & IFCAP_LRO) == 0 ||
+		    rq->lro.lro_cnt == 0 ||
+		    tcp_lro_rx(&rq->lro, mbs[i], 0) != 0) {
+			if_input(rq->ifp, mbs[i]);
+		}
+#endif
+	}
+
+	mtx_lock(&rq->mtx);
+}
+
 static int
-mlx5e_poll_rx_cq(struct mlx5e_rq *rq, int budget)
+mlx5e_poll_rx_cq(struct mlx5e_rq *rq, struct mbuf **mbs, int budget, int *cp)
 {
 	struct pfil_head *pfil;
-	int i, rv;
+	int i, c, rv;
 
 	CURVNET_SET_QUIET(if_getvnet(rq->ifp));
 	pfil = rq->channel->priv->pfil;
-	for (i = 0; i < budget; i++) {
+	for (i = 0, c = 0; i < budget; i++) {
 		struct mlx5e_rx_wqe *wqe;
 		struct mlx5_cqe64 *cqe;
 		struct mbuf *mb;
@@ -595,16 +618,7 @@
 		mb->m_pkthdr.numa_domain = if_getnumadomain(rq->ifp);
 #endif
 
-#if !defined(HAVE_TCP_LRO_RX)
-		tcp_lro_queue_mbuf(&rq->lro, mb);
-#else
-		if (mb->m_pkthdr.csum_flags == 0 ||
-		    (if_getcapenable(rq->ifp) & IFCAP_LRO) == 0 ||
-		    rq->lro.lro_cnt == 0 ||
-		    tcp_lro_rx(&rq->lro, mb, 0) != 0) {
-			if_input(rq->ifp, mb);
-		}
-#endif
+		mbs[c++] = mb;
 wq_ll_pop:
 		mlx5_wq_ll_pop(&rq->wq, wqe_counter_be,
 		    &wqe->next.next_wqe_index);
@@ -615,6 +629,7 @@
 
 	/* ensure cq space is freed before enabling more cqes */
 	atomic_thread_fence_rel();
+	*cp = c;
 	return (i);
 }
 
@@ -623,7 +638,8 @@
 {
 	struct mlx5e_channel *c = container_of(mcq, struct mlx5e_channel, rq.cq.mcq);
 	struct mlx5e_rq *rq = container_of(mcq, struct mlx5e_rq, cq.mcq);
-	int i = 0;
+	struct mbuf *mbs[MLX5E_RX_BUDGET_MAX];
+	int i = 0, mbs_count, wqe_processed;
 
 #ifdef HAVE_PER_CQ_EVENT_PACKET
 #if (MHLEN < 15)
@@ -657,9 +673,11 @@
 	 * Polling the entire CQ without posting new WQEs results in
 	 * lack of receive WQEs during heavy traffic scenarios.
 	 */
-	while (1) {
-		if (mlx5e_poll_rx_cq(rq, MLX5E_RX_BUDGET_MAX) !=
-		    MLX5E_RX_BUDGET_MAX)
+	for (;;) {
+		wqe_processed = mlx5e_poll_rx_cq(rq, mbs, MLX5E_RX_BUDGET_MAX,
+		    &mbs_count);
+		mlx5e_input_mbs(rq, mbs, mbs_count);
+		if (wqe_processed != MLX5E_RX_BUDGET_MAX)
 			break;
 		i += MLX5E_RX_BUDGET_MAX;
 		if (i >= MLX5E_BUDGET_MAX)
@@ -670,9 +688,10 @@
 	/* check for dynamic interrupt moderation callback */
 	if (rq->dim.mode != NET_DIM_CQ_PERIOD_MODE_DISABLED)
 		net_dim(&rq->dim, rq->stats.packets, rq->stats.bytes);
-	mlx5e_cq_arm(&rq->cq, MLX5_GET_DOORBELL_LOCK(&rq->channel->priv->doorbell_lock));
-	tcp_lro_flush_all(&rq->lro);
+	mlx5e_cq_arm(&rq->cq, MLX5_GET_DOORBELL_LOCK(
+	    &rq->channel->priv->doorbell_lock));
 	mtx_unlock(&rq->mtx);
+	tcp_lro_flush_all(&rq->lro);
 
 	for (int j = 0; j != MLX5E_MAX_TX_NUM_TC; j++) {
 		mtx_lock(&c->sq[j].lock);