diff --git a/sys/dev/cxgbe/tom/t4_cpl_io.c b/sys/dev/cxgbe/tom/t4_cpl_io.c
--- a/sys/dev/cxgbe/tom/t4_cpl_io.c
+++ b/sys/dev/cxgbe/tom/t4_cpl_io.c
@@ -1753,17 +1753,18 @@
 		if (changed) {
 			if (toep->ddp.flags & DDP_SC_REQ)
 				toep->ddp.flags ^= DDP_ON | DDP_SC_REQ;
-			else {
-				KASSERT(cpl->ddp_off == 1,
-				    ("%s: DDP switched on by itself.",
-				    __func__));
-
+			else if (cpl->ddp_off == 1) {
 				/* Fell out of DDP mode */
 				toep->ddp.flags &= ~DDP_ON;
 				CTR1(KTR_CXGBE, "%s: fell out of DDP mode",
 				    __func__);
 
 				insert_ddp_data(toep, ddp_placed);
+			} else {
+				/*
+				 * Data was received while still
+				 * ULP_MODE_NONE, just fall through.
+				 */
 			}
 		}
 
diff --git a/sys/dev/cxgbe/tom/t4_ddp.c b/sys/dev/cxgbe/tom/t4_ddp.c
--- a/sys/dev/cxgbe/tom/t4_ddp.c
+++ b/sys/dev/cxgbe/tom/t4_ddp.c
@@ -192,7 +192,7 @@
 		free_pageset(td, db->ps);
 }
 
-void
+static void
 ddp_init_toep(struct toepcb *toep)
 {
 
@@ -810,6 +810,78 @@
 	return (0);
 }
 
+static bool
+set_ddp_ulp_mode(struct toepcb *toep)
+{
+	struct adapter *sc = toep->vi->adapter;
+	struct wrqe *wr;
+	struct work_request_hdr *wrh;
+	struct ulp_txpkt *ulpmc;
+	int fields, len;
+
+	if (!sc->tt.ddp)
+		return (false);
+
+	fields = 0;
+
+	/* Overlay region including W_TCB_RX_DDP_FLAGS */
+	fields += 3;
+
+	/* W_TCB_ULP_TYPE */
+	fields++;
+
+#ifdef USE_DDP_RX_FLOW_CONTROL
+	/* W_TCB_T_FLAGS */
+	fields++;
+#endif
+
+	len = sizeof(*wrh) + fields * roundup2(LEN__SET_TCB_FIELD_ULP, 16);
+	KASSERT(len <= SGE_MAX_WR_LEN,
+	    ("%s: WR with %d TCB field updates too large", __func__, fields));
+
+	wr = alloc_wrqe(len, toep->ctrlq);
+	if (wr == NULL)
+		return (false);
+
+	CTR(KTR_CXGBE, "%s: tid %u", __func__, toep->tid);
+
+	wrh = wrtod(wr);
+	INIT_ULPTX_WRH(wrh, len, 1, 0);	/* atomic */
+	ulpmc = (struct ulp_txpkt *)(wrh + 1);
+
+	/*
+	 * Words 26/27 are zero except for the DDP_OFF flag in
+	 * W_TCB_RX_DDP_FLAGS (27).
+	 */
+	ulpmc = mk_set_tcb_field_ulp(ulpmc, toep, 26,
+	    0xffffffffffffffff, (uint64_t)V_TF_DDP_OFF(1) << 32);
+
+	/* Words 28/29 are zero. */
+	ulpmc = mk_set_tcb_field_ulp(ulpmc, toep, 28,
+	    0xffffffffffffffff, 0);
+
+	/* Words 30/31 are zero. */
+	ulpmc = mk_set_tcb_field_ulp(ulpmc, toep, 30,
+	    0xffffffffffffffff, 0);
+
+	/* Set the ULP mode to ULP_MODE_TCPDDP. */
+	toep->params.ulp_mode = ULP_MODE_TCPDDP;
+	ulpmc = mk_set_tcb_field_ulp(ulpmc, toep, W_TCB_ULP_TYPE,
+	    V_TCB_ULP_TYPE(M_TCB_ULP_TYPE),
+	    V_TCB_ULP_TYPE(ULP_MODE_TCPDDP));
+
+#ifdef USE_DDP_RX_FLOW_CONTROL
+	/* Set TF_RX_FLOW_CONTROL_DDP. */
+	ulpmc = mk_set_tcb_field_ulp(ulpmc, toep, W_TCB_T_FLAGS,
+	    V_TF_RX_FLOW_CONTROL_DDP(1), V_TF_RX_FLOW_CONTROL_DDP(1));
+#endif
+
+	ddp_init_toep(toep);
+
+	t4_wrq_tx(sc, wr);
+	return (true);
+}
+
 static void
 enable_ddp(struct adapter *sc, struct toepcb *toep)
 {
@@ -2203,7 +2275,8 @@
 int
 t4_aio_queue_ddp(struct socket *so, struct kaiocb *job)
 {
-	struct tcpcb *tp = sototcpcb(so);
+	struct inpcb *inp = sotoinpcb(so);
+	struct tcpcb *tp = intotcpcb(inp);
 	struct toepcb *toep = tp->t_toe;
 
 
@@ -2211,6 +2284,15 @@
 	if (job->uaiocb.aio_lio_opcode != LIO_READ)
 		return (EOPNOTSUPP);
 
+	INP_WLOCK(inp);
+	if (__predict_false(ulp_mode(toep) == ULP_MODE_NONE)) {
+		if (!set_ddp_ulp_mode(toep)) {
+			INP_WUNLOCK(inp);
+			return (EOPNOTSUPP);
+		}
+	}
+	INP_WUNLOCK(inp);
+
 	DDP_LOCK(toep);
 
 	/*
diff --git a/sys/dev/cxgbe/tom/t4_tom.h b/sys/dev/cxgbe/tom/t4_tom.h
--- a/sys/dev/cxgbe/tom/t4_tom.h
+++ b/sys/dev/cxgbe/tom/t4_tom.h
@@ -505,7 +505,6 @@
 void t4_ddp_mod_load(void);
 void t4_ddp_mod_unload(void);
 void ddp_assert_empty(struct toepcb *);
-void ddp_init_toep(struct toepcb *);
 void ddp_uninit_toep(struct toepcb *);
 void ddp_queue_toep(struct toepcb *);
 void release_ddp_resources(struct toepcb *toep);
diff --git a/sys/dev/cxgbe/tom/t4_tom.c b/sys/dev/cxgbe/tom/t4_tom.c
--- a/sys/dev/cxgbe/tom/t4_tom.c
+++ b/sys/dev/cxgbe/tom/t4_tom.c
@@ -179,8 +179,7 @@
 	toep->ctrlq = &sc->sge.ctrlq[pi->port_id];
 
 	tls_init_toep(toep);
-	if (ulp_mode(toep) == ULP_MODE_TCPDDP)
-		ddp_init_toep(toep);
+	MPASS(ulp_mode(toep) != ULP_MODE_TCPDDP);
 
 	toep->flags |= TPF_INITIALIZED;
 
@@ -1216,10 +1215,7 @@
 		opt2 |= V_RX_COALESCE(M_RX_COALESCE);
 
 	opt2 |= V_RX_FC_DDP(0) | V_RX_FC_DISABLE(0);
-#ifdef USE_DDP_RX_FLOW_CONTROL
-	if (cp->ulp_mode == ULP_MODE_TCPDDP)
-		opt2 |= F_RX_FC_DDP;
-#endif
+	MPASS(cp->ulp_mode != ULP_MODE_TCPDDP);
 
 	return (htobe32(opt2));
 }
@@ -1327,11 +1323,7 @@
 		cp->tx_align = 0;
 
 	/* ULP mode. */
-	if (s->ddp > 0 ||
-	    (s->ddp < 0 && sc->tt.ddp && (so_options_get(so) & SO_NO_DDP) == 0))
-		cp->ulp_mode = ULP_MODE_TCPDDP;
-	else
-		cp->ulp_mode = ULP_MODE_NONE;
+	cp->ulp_mode = ULP_MODE_NONE;
 
 	/* Rx coalescing. */
 	if (s->rx_coalesce >= 0)
@@ -1972,7 +1964,8 @@
 	if (SOLISTENING(so))
 		return (EINVAL);
 
-	if (ulp_mode(toep) == ULP_MODE_TCPDDP) {
+	if (ulp_mode(toep) == ULP_MODE_TCPDDP ||
+	    ulp_mode(toep) == ULP_MODE_NONE) {
 		error = t4_aio_queue_ddp(so, job);
 		if (error != EOPNOTSUPP)
 			return (error);