Index: dev/e1000/if_igb.c
===================================================================
--- dev/e1000/if_igb.c	(revision 229933)
+++ dev/e1000/if_igb.c	(working copy)
@@ -172,10 +172,8 @@
 static int	igb_resume(device_t);
 #if __FreeBSD_version >= 800000
 static int	igb_mq_start(struct ifnet *, struct mbuf *);
-static int	igb_mq_start_locked(struct ifnet *,
-		    struct tx_ring *, struct mbuf *);
+static void	igb_mq_start_locked(struct ifnet *, struct tx_ring *);
 static void	igb_qflush(struct ifnet *);
-static void	igb_deferred_mq_start(void *, int);
 #else
 static void	igb_start(struct ifnet *);
 static void	igb_start_locked(struct tx_ring *, struct ifnet *ifp);
@@ -802,7 +800,7 @@
 		for (int i = 0; i < adapter->num_queues; i++, txr++) {
 			IGB_TX_LOCK(txr);
 			if (!drbr_empty(ifp, txr->br))
-				igb_mq_start_locked(ifp, txr, NULL);
+				igb_mq_start_locked(ifp, txr);
 			IGB_TX_UNLOCK(txr);
 		}
 #endif
@@ -911,48 +909,38 @@
 	txr = &adapter->tx_rings[i];
 	que = &adapter->queues[i];
 
-	if (IGB_TX_TRYLOCK(txr)) {
-		err = igb_mq_start_locked(ifp, txr, m);
+	err = drbr_enqueue(ifp, txr->br, m);
+	if (err == 0) {
+		IGB_TX_LOCK(txr);
+		igb_mq_start_locked(ifp, txr);
 		IGB_TX_UNLOCK(txr);
-	} else {
-		err = drbr_enqueue(ifp, txr->br, m);
-		taskqueue_enqueue(que->tq, &txr->txq_task);
-	}
+	} else if (err == EINPROGRESS)
+		err = 0;
 
 	return (err);
 }
 
-static int
-igb_mq_start_locked(struct ifnet *ifp, struct tx_ring *txr, struct mbuf *m)
+static void
+igb_mq_start_locked(struct ifnet *ifp, struct tx_ring *txr)
 {
 	struct adapter  *adapter = txr->adapter;
         struct mbuf     *next;
-        int             err = 0, enq;
+        int             enq;
 
 	IGB_TX_LOCK_ASSERT(txr);
 
 	if ((ifp->if_drv_flags & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
-	    IFF_DRV_RUNNING || adapter->link_active == 0) {
-		if (m != NULL)
-			err = drbr_enqueue(ifp, txr->br, m);
-		return (err);
-	}
+	    IFF_DRV_RUNNING || adapter->link_active == 0)
+		return;
 
 	enq = 0;
-	if (m == NULL) {
-		next = drbr_dequeue(ifp, txr->br);
-	} else if (drbr_needs_enqueue(ifp, txr->br)) {
-		if ((err = drbr_enqueue(ifp, txr->br, m)) != 0)
-			return (err);
-		next = drbr_dequeue(ifp, txr->br);
-	} else
-		next = m;
+	next = drbr_dequeue(ifp, txr->br);
 
 	/* Process the queue */
 	while (next != NULL) {
-		if ((err = igb_xmit(txr, &next)) != 0) {
+		if (igb_xmit(txr, &next) != 0) {
 			if (next != NULL)
-				err = drbr_enqueue(ifp, txr->br, next);
+				drbr_enqueue(ifp, txr->br, next);
 			break;
 		}
 		enq++;
@@ -973,26 +961,9 @@
 		txr->queue_status = IGB_QUEUE_WORKING;
 		txr->watchdog_time = ticks;
 	}
-	return (err);
 }
 
 /*
- * Called from a taskqueue to drain queued transmit packets.
- */
-static void
-igb_deferred_mq_start(void *arg, int pending)
-{
-	struct tx_ring *txr = arg;
-	struct adapter *adapter = txr->adapter;
-	struct ifnet *ifp = adapter->ifp;
-
-	IGB_TX_LOCK(txr);
-	if (!drbr_empty(ifp, txr->br))
-		igb_mq_start_locked(ifp, txr, NULL);
-	IGB_TX_UNLOCK(txr);
-}
-
-/*
 ** Flush all ring buffers
 */
 static void
@@ -1339,7 +1310,7 @@
 			more = TRUE;
 #if __FreeBSD_version >= 800000
 		if (!drbr_empty(ifp, txr->br))
-			igb_mq_start_locked(ifp, txr, NULL);
+			igb_mq_start_locked(ifp, txr);
 #else
 		igb_start_locked(txr, ifp);
 #endif
@@ -1463,7 +1434,7 @@
 	} while (loop-- && more);
 #if __FreeBSD_version >= 800000
 	if (!drbr_empty(ifp, txr->br))
-		igb_mq_start_locked(ifp, txr, NULL);
+		igb_mq_start_locked(ifp, txr);
 #else
 	igb_start_locked(txr, ifp);
 #endif
@@ -2075,6 +2046,16 @@
 			device_printf(dev, "Link: thermal downshift\n");
 		/* This can sleep */
 		if_link_state_change(ifp, LINK_STATE_UP);
+#if __FreeBSD_version < 800000
+		igb_start(ifp);
+#else
+		for (int i = 0; i < adapter->num_queues; i++, txr++) {
+			IGB_TX_LOCK(txr);
+			if (!drbr_empty(ifp, txr->br))
+				igb_mq_start_locked(ifp, txr);
+			IGB_TX_UNLOCK(txr);
+		}
+#endif
 	} else if (!link_check && (adapter->link_active == 1)) {
 		ifp->if_baudrate = adapter->link_speed = 0;
 		adapter->link_duplex = 0;
@@ -2212,7 +2193,6 @@
 {
 	device_t		dev = adapter->dev;
 	struct igb_queue	*que = adapter->queues;
-	struct tx_ring		*txr = adapter->tx_rings;
 	int			error, rid = 0;
 
 	/* Turn off all interrupts */
@@ -2231,10 +2211,6 @@
 		return (ENXIO);
 	}
 
-#if __FreeBSD_version >= 800000
-	TASK_INIT(&txr->txq_task, 0, igb_deferred_mq_start, txr);
-#endif
-
 	/*
 	 * Try allocating a fast interrupt and the associated deferred
 	 * processing contexts.
@@ -2305,10 +2281,7 @@
 		*/
 		if (adapter->num_queues > 1)
 			bus_bind_intr(dev, que->res, i);
-#if __FreeBSD_version >= 800000
-		TASK_INIT(&que->txr->txq_task, 0, igb_deferred_mq_start,
-		    que->txr);
-#endif
+
 		/* Make tasklet for deferred handling */
 		TASK_INIT(&que->que_task, 0, igb_handle_que, que);
 		que->tq = taskqueue_create("igb_que", M_NOWAIT,
@@ -2529,9 +2502,6 @@
 
 	for (int i = 0; i < adapter->num_queues; i++, que++) {
 		if (que->tq != NULL) {
-#if __FreeBSD_version >= 800000
-			taskqueue_drain(que->tq, &que->txr->txq_task);
-#endif
 			taskqueue_drain(que->tq, &que->que_task);
 			taskqueue_free(que->tq);
 		}
Index: dev/e1000/if_igb.h
===================================================================
--- dev/e1000/if_igb.h	(revision 229933)
+++ dev/e1000/if_igb.h	(working copy)
@@ -297,7 +297,6 @@
 	struct buf_ring		*br;
 #endif
 	bus_dma_tag_t		txtag;
-	struct task		txq_task;
 
 	u32			bytes;
 	u32			packets;
Index: dev/e1000/if_em.c
===================================================================
--- dev/e1000/if_em.c	(revision 229933)
+++ dev/e1000/if_em.c	(working copy)
@@ -192,13 +192,13 @@
 static int	em_shutdown(device_t);
 static int	em_suspend(device_t);
 static int	em_resume(device_t);
-static void	em_start(struct ifnet *);
-static void	em_start_locked(struct ifnet *, struct tx_ring *);
 #ifdef EM_MULTIQUEUE
 static int	em_mq_start(struct ifnet *, struct mbuf *);
-static int	em_mq_start_locked(struct ifnet *,
-		    struct tx_ring *, struct mbuf *);
+static void	em_mq_start_locked(struct ifnet *, struct tx_ring *);
 static void	em_qflush(struct ifnet *);
+#else
+static void	em_start(struct ifnet *);
+static void	em_start_locked(struct ifnet *, struct tx_ring *);
 #endif
 static int	em_ioctl(struct ifnet *, u_long, caddr_t);
 static void	em_init(void *);
@@ -804,13 +804,25 @@
 em_resume(device_t dev)
 {
 	struct adapter *adapter = device_get_softc(dev);
+#ifdef EM_MULTIQUEUE
+	struct tx_ring *txr = adapter->tx_rings;
+#endif
 	struct ifnet *ifp = adapter->ifp;
 
 	EM_CORE_LOCK(adapter);
 	em_init_locked(adapter);
 	em_init_manageability(adapter);
 	EM_CORE_UNLOCK(adapter);
+#ifdef EM_MULTIQUEUE
+	for (int i = 0; i < adapter->num_queues; i++, txr++) {
+		EM_TX_LOCK(txr);
+		if (!drbr_empty(ifp, txr->br))
+			em_mq_start_locked(ifp, txr);
+		EM_TX_UNLOCK(txr);
+	}
+#else
 	em_start(ifp);
+#endif
 
 	return bus_generic_resume(dev);
 }
@@ -827,39 +839,29 @@
  **********************************************************************/
 
 #ifdef EM_MULTIQUEUE
-static int
-em_mq_start_locked(struct ifnet *ifp, struct tx_ring *txr, struct mbuf *m)
+static void
+em_mq_start_locked(struct ifnet *ifp, struct tx_ring *txr)
 {
 	struct adapter  *adapter = txr->adapter;
         struct mbuf     *next;
-        int             err = 0, enq = 0;
+        int             enq;
 
 	if ((ifp->if_drv_flags & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
-	    IFF_DRV_RUNNING || adapter->link_active == 0) {
-		if (m != NULL)
-			err = drbr_enqueue(ifp, txr->br, m);
-		return (err);
-	}
+	    IFF_DRV_RUNNING || adapter->link_active == 0)
+		return;
 
         /* Call cleanup if number of TX descriptors low */
 	if (txr->tx_avail <= EM_TX_CLEANUP_THRESHOLD)
 		em_txeof(txr);
 
 	enq = 0;
-	if (m == NULL) {
-		next = drbr_dequeue(ifp, txr->br);
-	} else if (drbr_needs_enqueue(ifp, txr->br)) {
-		if ((err = drbr_enqueue(ifp, txr->br, m)) != 0)
-			return (err);
-		next = drbr_dequeue(ifp, txr->br);
-	} else
-		next = m;
+	next = drbr_dequeue(ifp, txr->br);
 
 	/* Process the queue */
 	while (next != NULL) {
-		if ((err = em_xmit(txr, &next)) != 0) {
+		if (em_xmit(txr, &next) != 0) {
                         if (next != NULL)
-                                err = drbr_enqueue(ifp, txr->br, next);
+                                drbr_enqueue(ifp, txr->br, next);
                         break;
 		}
 		enq++;
@@ -879,7 +881,6 @@
                 txr->queue_status = EM_QUEUE_WORKING;
 		txr->watchdog_time = ticks;
 	}
-	return (err);
 }
 
 /*
@@ -892,11 +893,13 @@
 	struct tx_ring	*txr = adapter->tx_rings;
 	int 		error;
 
-	if (EM_TX_TRYLOCK(txr)) {
-		error = em_mq_start_locked(ifp, txr, m);
+	error = drbr_enqueue(ifp, txr->br, m);
+	if (error == 0) {
+		EM_TX_LOCK(txr);
+		em_mq_start_locked(ifp, txr);
 		EM_TX_UNLOCK(txr);
-	} else 
-		error = drbr_enqueue(ifp, txr->br, m);
+	} else if (error == EINPROGRESS)
+		error = 0;
 
 	return (error);
 }
@@ -920,7 +923,7 @@
 	if_qflush(ifp);
 }
 
-#endif /* EM_MULTIQUEUE */
+#else  /* !EM_MULTIQUEUE */
 
 static void
 em_start_locked(struct ifnet *ifp, struct tx_ring *txr)
@@ -985,6 +988,8 @@
 	return;
 }
 
+#endif /* EM_MULTIQUEUE */
+
 /*********************************************************************
  *  Ioctl entry point
  *
@@ -1418,7 +1423,7 @@
 	em_txeof(txr);
 #ifdef EM_MULTIQUEUE
 	if (!drbr_empty(ifp, txr->br))
-		em_mq_start_locked(ifp, txr, NULL);
+		em_mq_start_locked(ifp, txr);
 #else
 	em_start_locked(ifp, txr);
 #endif
@@ -1491,7 +1496,7 @@
 		em_txeof(txr);
 #ifdef EM_MULTIQUEUE
 		if (!drbr_empty(ifp, txr->br))
-			em_mq_start_locked(ifp, txr, NULL);
+			em_mq_start_locked(ifp, txr);
 #else
 		em_start_locked(ifp, txr);
 #endif
@@ -1603,7 +1608,7 @@
 	em_txeof(txr);
 #ifdef EM_MULTIQUEUE
 	if (!drbr_empty(ifp, txr->br))
-		em_mq_start_locked(ifp, txr, NULL);
+		em_mq_start_locked(ifp, txr);
 #else
 	em_start_locked(ifp, txr);
 #endif
@@ -2285,6 +2290,16 @@
 		adapter->smartspeed = 0;
 		ifp->if_baudrate = adapter->link_speed * 1000000;
 		if_link_state_change(ifp, LINK_STATE_UP);
+#ifdef EM_MULTIQUEUE
+		for (int i = 0; i < adapter->num_queues; i++, txr++) {
+			EM_TX_LOCK(txr);
+			if (!drbr_empty(ifp, txr->br))
+				em_mq_start_locked(ifp, txr);
+			EM_TX_UNLOCK(txr);
+		}
+#else
+		em_start(ifp);
+#endif
 	} else if (!link_check && (adapter->link_active == 1)) {
 		ifp->if_baudrate = adapter->link_speed = 0;
 		adapter->link_duplex = 0;
@@ -2816,7 +2831,13 @@
 	ifp->if_softc = adapter;
 	ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST;
 	ifp->if_ioctl = em_ioctl;
+#ifdef EM_MULTIQUEUE
+	/* Multiqueue tx functions */
+	ifp->if_transmit = em_mq_start;
+	ifp->if_qflush = em_qflush;
+#else
 	ifp->if_start = em_start;
+#endif	
 	IFQ_SET_MAXLEN(&ifp->if_snd, adapter->num_tx_desc - 1);
 	ifp->if_snd.ifq_drv_maxlen = adapter->num_tx_desc - 1;
 	IFQ_SET_READY(&ifp->if_snd);
@@ -2825,11 +2846,6 @@
 
 	ifp->if_capabilities = ifp->if_capenable = 0;
 
-#ifdef EM_MULTIQUEUE
-	/* Multiqueue tx functions */
-	ifp->if_transmit = em_mq_start;
-	ifp->if_qflush = em_qflush;
-#endif	
 
 	ifp->if_capabilities |= IFCAP_HWCSUM | IFCAP_VLAN_HWCSUM;
 	ifp->if_capenable |= IFCAP_HWCSUM | IFCAP_VLAN_HWCSUM;
Index: dev/cxgbe/adapter.h
===================================================================
--- dev/cxgbe/adapter.h	(revision 229933)
+++ dev/cxgbe/adapter.h	(working copy)
@@ -621,7 +621,7 @@
 void t4_intr_err(void *);
 void t4_intr_evt(void *);
 int t4_mgmt_tx(struct adapter *, struct mbuf *);
-int t4_eth_tx(struct ifnet *, struct sge_txq *, struct mbuf *);
+void t4_eth_tx(struct ifnet *, struct sge_txq *);
 void t4_update_fl_bufsize(struct ifnet *);
 
 #endif
Index: dev/cxgbe/t4_sge.c
===================================================================
--- dev/cxgbe/t4_sge.c	(revision 229933)
+++ dev/cxgbe/t4_sge.c	(working copy)
@@ -838,14 +838,14 @@
     sizeof(struct cpl_tx_pkt_core) \
     ) / 8 )
 
-int
-t4_eth_tx(struct ifnet *ifp, struct sge_txq *txq, struct mbuf *m)
+void
+t4_eth_tx(struct ifnet *ifp, struct sge_txq *txq)
 {
 	struct port_info *pi = (void *)ifp->if_softc;
 	struct adapter *sc = pi->adapter;
 	struct sge_eq *eq = &txq->eq;
 	struct buf_ring *br = txq->br;
-	struct mbuf *next;
+	struct mbuf *m, *next;
 	int rc, coalescing, can_reclaim;
 	struct txpkts txpkts;
 	struct sgl sgl;
@@ -862,7 +862,8 @@
 	if (eq->avail < 8)
 		reclaim_tx_descs(txq, 0, 8);
 
-	for (; m; m = next ? next : drbr_dequeue(ifp, br)) {
+	for (m = txq->m ? txq->m : drbr_dequeue(ifp, br); m;
+	     m = next ? next : drbr_dequeue(ifp, br)) {
 
 		if (eq->avail < 8)
 			break;
@@ -977,8 +978,6 @@
 	can_reclaim = reclaimable(eq);
 	if (can_reclaim >= 32)
 		reclaim_tx_descs(txq, can_reclaim, 128);
-
-	return (0);
 }
 
 void
Index: dev/cxgbe/t4_main.c
===================================================================
--- dev/cxgbe/t4_main.c	(revision 229933)
+++ dev/cxgbe/t4_main.c	(working copy)
@@ -118,7 +118,6 @@
 /* ifnet + media interface */
 static void cxgbe_init(void *);
 static int cxgbe_ioctl(struct ifnet *, unsigned long, caddr_t);
-static void cxgbe_start(struct ifnet *);
 static int cxgbe_transmit(struct ifnet *, struct mbuf *);
 static void cxgbe_qflush(struct ifnet *);
 static int cxgbe_media_change(struct ifnet *);
@@ -301,7 +300,6 @@
 #ifdef T4_DEVLOG
 static int sysctl_devlog(SYSCTL_HANDLER_ARGS);
 #endif
-static inline void txq_start(struct ifnet *, struct sge_txq *);
 static uint32_t fconf_to_mode(uint32_t);
 static uint32_t mode_to_fconf(uint32_t);
 static uint32_t fspec_to_fconf(struct t4_filter_specification *);
@@ -774,14 +772,9 @@
 
 	ifp->if_init = cxgbe_init;
 	ifp->if_ioctl = cxgbe_ioctl;
-	ifp->if_start = cxgbe_start;
 	ifp->if_transmit = cxgbe_transmit;
 	ifp->if_qflush = cxgbe_qflush;
 
-	ifp->if_snd.ifq_drv_maxlen = 1024;
-	IFQ_SET_MAXLEN(&ifp->if_snd, ifp->if_snd.ifq_drv_maxlen);
-	IFQ_SET_READY(&ifp->if_snd);
-
 	ifp->if_capabilities = T4_CAP;
 	ifp->if_capenable = T4_CAP_ENABLE;
 	ifp->if_hwassist = CSUM_TCP | CSUM_UDP | CSUM_IP | CSUM_TSO;
@@ -1020,21 +1013,6 @@
 	return (rc);
 }
 
-static void
-cxgbe_start(struct ifnet *ifp)
-{
-	struct port_info *pi = ifp->if_softc;
-	struct sge_txq *txq;
-	int i;
-
-	for_each_txq(pi, i, txq) {
-		if (TXQ_TRYLOCK(txq)) {
-			txq_start(ifp, txq);
-			TXQ_UNLOCK(txq);
-		}
-	}
-}
-
 static int
 cxgbe_transmit(struct ifnet *ifp, struct mbuf *m)
 {
@@ -1055,43 +1033,13 @@
 		txq += (m->m_pkthdr.flowid % pi->ntxq);
 	br = txq->br;
 
-	if (TXQ_TRYLOCK(txq) == 0) {
-		/*
-		 * XXX: make sure that this packet really is sent out.  There is
-		 * a small race where t4_eth_tx may stop draining the drbr and
-		 * goes away, just before we enqueued this mbuf.
-		 */
-
-		return (drbr_enqueue(ifp, br, m));
-	}
-
-	/*
-	 * txq->m is the mbuf that is held up due to a temporary shortage of
-	 * resources and it should be put on the wire first.  Then what's in
-	 * drbr and finally the mbuf that was just passed in to us.
-	 *
-	 * Return code should indicate the fate of the mbuf that was passed in
-	 * this time.
-	 */
-
-	TXQ_LOCK_ASSERT_OWNED(txq);
-	if (drbr_needs_enqueue(ifp, br) || txq->m) {
-
-		/* Queued for transmission. */
-
-		rc = drbr_enqueue(ifp, br, m);
-		m = txq->m ? txq->m : drbr_dequeue(ifp, br);
-		(void) t4_eth_tx(ifp, txq, m);
+	rc = drbr_enqueue(ifp, br, m);
+	if (rc == 0) {
+		TXQ_LOCK(txq);
+		t4_eth_tx(ifp, txq);
 		TXQ_UNLOCK(txq);
-		return (rc);
-	}
-
-	/* Direct transmission. */
-	rc = t4_eth_tx(ifp, txq, m);
-	if (rc != 0 && txq->m)
-		rc = 0;	/* held, will be transmitted soon (hopefully) */
-
-	TXQ_UNLOCK(txq);
+	} else if (rc == EINPROGRESS)
+		rc = 0;
 	return (rc);
 }
 
@@ -2891,20 +2839,6 @@
 }
 #endif
 
-static inline void
-txq_start(struct ifnet *ifp, struct sge_txq *txq)
-{
-	struct buf_ring *br;
-	struct mbuf *m;
-
-	TXQ_LOCK_ASSERT_OWNED(txq);
-
-	br = txq->br;
-	m = txq->m ? txq->m : drbr_dequeue(ifp, br);
-	if (m)
-		t4_eth_tx(ifp, txq, m);
-}
-
 void
 cxgbe_txq_start(void *arg, int count)
 {
@@ -2913,7 +2847,7 @@
 	TXQ_LOCK(txq);
 	if (txq->eq.flags & EQ_CRFLUSHED) {
 		txq->eq.flags &= ~EQ_CRFLUSHED;
-		txq_start(txq->ifp, txq);
+		t4_eth_tx(txq->ifp, txq);
 	} else
 		wakeup_one(txq);	/* txq is going away, wakeup free_txq */
 	TXQ_UNLOCK(txq);
Index: dev/cxgb/cxgb_sge.c
===================================================================
--- dev/cxgb/cxgb_sge.c	(revision 229933)
+++ dev/cxgb/cxgb_sge.c	(working copy)
@@ -1685,60 +1685,33 @@
 	if (txq->db_pending)
 		check_ring_tx_db(pi->adapter, txq, 1);
 
+#if 1
+	/*
+	 * XXX: Should be handled by the ISR noticing TX completions
+	 * and thus more room in the queue?
+	 */
 	if (!TXQ_RING_EMPTY(qs) && callout_pending(&txq->txq_timer) == 0 &&
 	    pi->link_config.link_ok)
 		callout_reset_on(&txq->txq_timer, 1, cxgb_tx_timeout,
 		    qs, txq->txq_timer.c_cpu);
+#endif
 	if (m_head != NULL)
 		m_freem(m_head);
 }
 
-static int
-cxgb_transmit_locked(struct ifnet *ifp, struct sge_qset *qs, struct mbuf *m)
+static void
+cxgb_transmit_locked(struct ifnet *ifp, struct sge_qset *qs)
 {
 	struct port_info *pi = qs->port;
 	struct sge_txq *txq = &qs->txq[TXQ_ETH];
 	struct buf_ring *br = txq->txq_mr;
-	int error, avail;
 
-	avail = txq->size - txq->in_use;
 	TXQ_LOCK_ASSERT(qs);
 
-	/*
-	 * We can only do a direct transmit if the following are true:
-	 * - we aren't coalescing (ring < 3/4 full)
-	 * - the link is up -- checked in caller
-	 * - there are no packets enqueued already
-	 * - there is space in hardware transmit queue 
-	 */
-	if (check_pkt_coalesce(qs) == 0 &&
-	    !TXQ_RING_NEEDS_ENQUEUE(qs) && avail > TX_MAX_DESC) {
-		if (t3_encap(qs, &m)) {
-			if (m != NULL &&
-			    (error = drbr_enqueue(ifp, br, m)) != 0) 
-				return (error);
-		} else {
-			if (txq->db_pending)
-				check_ring_tx_db(pi->adapter, txq, 1);
-
-			/*
-			 * We've bypassed the buf ring so we need to update
-			 * the stats directly
-			 */
-			txq->txq_direct_packets++;
-			txq->txq_direct_bytes += m->m_pkthdr.len;
-		}
-	} else if ((error = drbr_enqueue(ifp, br, m)) != 0)
-		return (error);
-
 	reclaim_completed_tx(qs, cxgb_tx_reclaim_threshold, TXQ_ETH);
 	if (!TXQ_RING_EMPTY(qs) && pi->link_config.link_ok &&
 	    (!check_pkt_coalesce(qs) || (drbr_inuse(ifp, br) >= 7)))
 		cxgb_start_locked(qs);
-	else if (!TXQ_RING_EMPTY(qs) && !callout_pending(&txq->txq_timer))
-		callout_reset_on(&txq->txq_timer, 1, cxgb_tx_timeout,
-		    qs, txq->txq_timer.c_cpu);
-	return (0);
 }
 
 int
@@ -1758,29 +1731,18 @@
 		qidx = (m->m_pkthdr.flowid % pi->nqsets) + pi->first_qset;
 
 	qs = &pi->adapter->sge.qs[qidx];
-	
-	if (TXQ_TRYLOCK(qs)) {
-		/* XXX running */
-		error = cxgb_transmit_locked(ifp, qs, m);
+
+	error = drbr_enqueue(ifp, qs->txq[TXQ_ETH].txq_mr, m);
+	if (error == 0) {
+		TXQ_LOCK(qs);
+		cxgb_transmit_locked(ifp, qs);
 		TXQ_UNLOCK(qs);
-	} else
-		error = drbr_enqueue(ifp, qs->txq[TXQ_ETH].txq_mr, m);
+	} else if (error == EINPROGRESS) {
+		error = 0;
+	}
 	return (error);
 }
-void
-cxgb_start(struct ifnet *ifp)
-{
-	struct port_info *pi = ifp->if_softc;
-	struct sge_qset *qs = &pi->adapter->sge.qs[pi->first_qset];
-	
-	if (!pi->link_config.link_ok)
-		return;
 
-	TXQ_LOCK(qs);
-	cxgb_start_locked(qs);
-	TXQ_UNLOCK(qs);
-}
-
 void
 cxgb_qflush(struct ifnet *ifp)
 {
Index: dev/cxgb/cxgb_main.c
===================================================================
--- dev/cxgb/cxgb_main.c	(revision 229933)
+++ dev/cxgb/cxgb_main.c	(working copy)
@@ -1023,12 +1023,9 @@
 	ifp->if_softc = p;
 	ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST;
 	ifp->if_ioctl = cxgb_ioctl;
-	ifp->if_start = cxgb_start;
+	ifp->if_transmit = cxgb_transmit;
+	ifp->if_qflush = cxgb_qflush;
 
-	ifp->if_snd.ifq_drv_maxlen = max(cxgb_snd_queue_len, ifqmaxlen);
-	IFQ_SET_MAXLEN(&ifp->if_snd, ifp->if_snd.ifq_drv_maxlen);
-	IFQ_SET_READY(&ifp->if_snd);
-
 	ifp->if_capabilities = CXGB_CAP;
 	ifp->if_capenable = CXGB_CAP_ENABLE;
 	ifp->if_hwassist = CSUM_TCP | CSUM_UDP | CSUM_IP | CSUM_TSO;
@@ -1043,8 +1040,6 @@
 	}
 
 	ether_ifattach(ifp, p->hw_addr);
-	ifp->if_transmit = cxgb_transmit;
-	ifp->if_qflush = cxgb_qflush;
 
 #ifdef DEFAULT_JUMBO
 	if (sc->params.nports <= 2)
Index: dev/cxgb/cxgb_adapter.h
===================================================================
--- dev/cxgb/cxgb_adapter.h	(revision 229933)
+++ dev/cxgb/cxgb_adapter.h	(working copy)
@@ -572,5 +572,4 @@
 void cxgb_tx_watchdog(void *arg);
 int cxgb_transmit(struct ifnet *ifp, struct mbuf *m);
 void cxgb_qflush(struct ifnet *ifp);
-void cxgb_start(struct ifnet *ifp);
 #endif
Index: sys/buf_ring.h
===================================================================
--- sys/buf_ring.h	(revision 229933)
+++ sys/buf_ring.h	(working copy)
@@ -49,10 +49,12 @@
 	uint64_t		br_drops;
 	uint64_t		br_prod_bufs;
 	uint64_t		br_prod_bytes;
+	volatile int		br_busy;
 	/*
 	 * Pad out to next L2 cache line
 	 */
-	uint64_t	  	_pad0[11];
+	uint32_t		_pad0[1];
+	uint64_t	  	_pad1[10];
 
 	volatile uint32_t	br_cons_head;
 	volatile uint32_t	br_cons_tail;
@@ -62,7 +64,7 @@
 	/*
 	 * Pad out to next L2 cache line
 	 */
-	uint64_t	  	_pad1[14];
+	uint64_t	  	_pad2[14];
 #ifdef DEBUG_BUFRING
 	struct mtx		*br_lock;
 #endif	
@@ -78,7 +80,7 @@
 {
 	uint32_t prod_head, prod_next;
 	uint32_t cons_tail;
-	int success;
+	int busy, success;
 #ifdef DEBUG_BUFRING
 	int i;
 	for (i = br->br_cons_head; i != br->br_prod_head;
@@ -98,16 +100,16 @@
 			critical_exit();
 			return (ENOBUFS);
 		}
-		
+
 		success = atomic_cmpset_int(&br->br_prod_head, prod_head,
 		    prod_next);
 	} while (success == 0);
 #ifdef DEBUG_BUFRING
 	if (br->br_ring[prod_head] != NULL)
 		panic("dangling value in enqueue");
-#endif	
-	br->br_ring[prod_head] = buf;
-	wmb();
+#endif
+	atomic_store_rel_ptr((uintptr_t *)&br->br_ring[prod_head],
+	    (uintptr_t)buf);
 
 	/*
 	 * If there are other enqueues in progress
@@ -119,8 +121,9 @@
 	br->br_prod_bufs++;
 	br->br_prod_bytes += nbytes;
 	br->br_prod_tail = prod_next;
+	busy = atomic_load_acq_int(&br->br_busy);
 	critical_exit();
-	return (0);
+	return (busy ? EINPROGRESS : 0);
 }
 
 static __inline int
@@ -158,11 +161,10 @@
 		    cons_next);
 	} while (success == 0);		
 
-	buf = br->br_ring[cons_head];
+	buf = (void *)atomic_load_acq_ptr((uintptr_t *)&br->br_ring[cons_head]);
 #ifdef DEBUG_BUFRING
 	br->br_ring[cons_head] = NULL;
 #endif
-	rmb();
 	
 	/*
 	 * If there are other dequeues in progress
@@ -196,8 +198,10 @@
 	cons_next = (cons_head + 1) & br->br_cons_mask;
 	cons_next_next = (cons_head + 2) & br->br_cons_mask;
 	
-	if (cons_head == prod_tail) 
+	if (cons_head == prod_tail) {
+		KASSERT(br->br_busy == 0, ("empty buf_ring still busy"));
 		return (NULL);
+	}
 
 #ifdef PREFETCH_DEFINED	
 	if (cons_next != prod_tail) {		
@@ -206,7 +210,8 @@
 			prefetch(br->br_ring[cons_next_next]);
 	}
 #endif
-	br->br_cons_head = cons_next;
+	br->br_busy = (cons_next != prod_tail);
+	atomic_store_rel_int(&br->br_cons_head, cons_next);
 	buf = br->br_ring[cons_head];
 
 #ifdef DEBUG_BUFRING