MFC r271946 and r272595:

Improve transmit sending offload, TSO, algorithm in general. This
change allows all HCAs from Mellanox Technologies to function properly
when TSO is enabled. See r271946 and r272595 for more details about
this commit.

Sponsored by:	Mellanox Technologies

16 files changed, 263 insertions, 46 deletions

diff --git a/sys/dev/oce/oce_if.c b/sys/dev/oce/oce_if.c
index 48f85e6..273089f 100644
--- a/sys/dev/oce/oce_if.c
+++ b/sys/dev/oce/oce_if.c
@@ -1735,7 +1735,9 @@ oce_attach_ifp(POCE_SOFTC sc)
 	if_initbaudrate(sc->ifp, IF_Gbps(10));
 
 #if __FreeBSD_version >= 1000000
-	sc->ifp->if_hw_tsomax = OCE_MAX_TSO_SIZE;
+	sc->ifp->if_hw_tsomax = 65536 - (ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN);
+	sc->ifp->if_hw_tsomaxsegcount = OCE_MAX_TX_ELEMENTS;
+	sc->ifp->if_hw_tsomaxsegsize = 4096;
 #endif
 
 	ether_ifattach(sc->ifp, sc->macaddr.mac_addr);
diff --git a/sys/dev/oce/oce_if.h b/sys/dev/oce/oce_if.h
index 259a9ed..dc96170 100644
--- a/sys/dev/oce/oce_if.h
+++ b/sys/dev/oce/oce_if.h
@@ -150,7 +150,6 @@ extern int mp_ncpus;			/* system's total active cpu cores */
 #define OCE_MAX_TX_ELEMENTS		29
 #define OCE_MAX_TX_DESC			1024
 #define OCE_MAX_TX_SIZE			65535
-#define OCE_MAX_TSO_SIZE		(65535 - ETHER_HDR_LEN)
 #define OCE_MAX_RX_SIZE			4096
 #define OCE_MAX_RQ_POSTS		255
 #define OCE_DEFAULT_PROMISCUOUS		0
diff --git a/sys/dev/vmware/vmxnet3/if_vmx.c b/sys/dev/vmware/vmxnet3/if_vmx.c
index 5767892..f00a108 100644
--- a/sys/dev/vmware/vmxnet3/if_vmx.c
+++ b/sys/dev/vmware/vmxnet3/if_vmx.c
@@ -1720,7 +1720,9 @@ vmxnet3_setup_interface(struct vmxnet3_softc *sc)
 	ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST;
 	ifp->if_init = vmxnet3_init;
 	ifp->if_ioctl = vmxnet3_ioctl;
-	ifp->if_hw_tsomax = VMXNET3_TSO_MAXSIZE;
+	ifp->if_hw_tsomax = 65536 - (ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN);
+	ifp->if_hw_tsomaxsegcount = VMXNET3_TX_MAXSEGS;
+	ifp->if_hw_tsomaxsegsize = VMXNET3_TX_MAXSEGSIZE;
 
 #ifdef VMXNET3_LEGACY_TX
 	ifp->if_start = vmxnet3_start;
diff --git a/sys/dev/vmware/vmxnet3/if_vmxvar.h b/sys/dev/vmware/vmxnet3/if_vmxvar.h
index 6c79721..8613516 100644
--- a/sys/dev/vmware/vmxnet3/if_vmxvar.h
+++ b/sys/dev/vmware/vmxnet3/if_vmxvar.h
@@ -277,8 +277,6 @@ struct vmxnet3_softc {
  */
 #define VMXNET3_TX_MAXSEGS		32
 #define VMXNET3_TX_MAXSIZE		(VMXNET3_TX_MAXSEGS * MCLBYTES)
-#define VMXNET3_TSO_MAXSIZE \
-    (VMXNET3_TX_MAXSIZE - sizeof(struct ether_vlan_header))
 
 /*
  * Maximum support Tx segments size. The length field in the
diff --git a/sys/dev/xen/netfront/netfront.c b/sys/dev/xen/netfront/netfront.c
index 0890e8d..3c7636f 100644
--- a/sys/dev/xen/netfront/netfront.c
+++ b/sys/dev/xen/netfront/netfront.c
@@ -134,7 +134,6 @@ static const int MODPARM_rx_flip = 0;
  * to mirror the Linux MAX_SKB_FRAGS constant.
  */
 #define	MAX_TX_REQ_FRAGS (65536 / PAGE_SIZE + 2)
-#define	NF_TSO_MAXBURST ((IP_MAXPACKET / PAGE_SIZE) * MCLBYTES)
 
 #define RX_COPY_THRESHOLD 256
 
@@ -2102,7 +2101,9 @@ create_netdev(device_t dev)
 	
     	ifp->if_hwassist = XN_CSUM_FEATURES;
     	ifp->if_capabilities = IFCAP_HWCSUM;
-	ifp->if_hw_tsomax = NF_TSO_MAXBURST;
+	ifp->if_hw_tsomax = 65536 - (ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN);
+	ifp->if_hw_tsomaxsegcount = MAX_TX_REQ_FRAGS;
+	ifp->if_hw_tsomaxsegsize = PAGE_SIZE;
 	
     	ether_ifattach(ifp, np->mac);
     	callout_init(&np->xn_stat_ch, CALLOUT_MPSAFE);
diff --git a/sys/kern/uipc_sockbuf.c b/sys/kern/uipc_sockbuf.c
index b66ea60..0c67347 100644
--- a/sys/kern/uipc_sockbuf.c
+++ b/sys/kern/uipc_sockbuf.c
@@ -1015,6 +1015,37 @@ sbsndptr(struct sockbuf *sb, u_int off, u_int len, u_int *moff)
 }
 
 /*
+ * Return the first mbuf and the mbuf data offset for the provided
+ * send offset without changing the "sb_sndptroff" field.
+ */
+struct mbuf *
+sbsndmbuf(struct sockbuf *sb, u_int off, u_int *moff)
+{
+	struct mbuf *m;
+
+	KASSERT(sb->sb_mb != NULL, ("%s: sb_mb is NULL", __func__));
+
+	/*
+	 * If the "off" is below the stored offset, which happens on
+	 * retransmits, just use "sb_mb":
+	 */
+	if (sb->sb_sndptr == NULL || sb->sb_sndptroff > off) {
+		m = sb->sb_mb;
+	} else {
+		m = sb->sb_sndptr;
+		off -= sb->sb_sndptroff;
+	}
+	while (off > 0 && m != NULL) {
+		if (off < m->m_len)
+			break;
+		off -= m->m_len;
+		m = m->m_next;
+	}
+	*moff = off;
+	return (m);
+}
+
+/*
  * Drop a record off the front of a sockbuf and move the next record to the
  * front.
  */
diff --git a/sys/net/if.c b/sys/net/if.c
index a658384..2e5d0ab 100644
--- a/sys/net/if.c
+++ b/sys/net/if.c
@@ -583,6 +583,57 @@ if_attach(struct ifnet *ifp)
 	if_attach_internal(ifp, 0);
 }
 
+/*
+ * Compute the least common TSO limit.
+ */
+void
+if_hw_tsomax_common(struct ifnet *ifp, struct ifnet_hw_tsomax *pmax)
+{
+	/*
+	 * 1) If there is no limit currently, take the limit from
+	 * the network adapter.
+	 *
+	 * 2) If the network adapter has a limit below the current
+	 * limit, apply it.
+	 */
+	if (pmax->tsomaxbytes == 0 || (ifp->if_hw_tsomax != 0 &&
+	    ifp->if_hw_tsomax < pmax->tsomaxbytes)) {
+		pmax->tsomaxbytes = ifp->if_hw_tsomax;
+	}
+	if (pmax->tsomaxsegcount == 0 || (ifp->if_hw_tsomaxsegcount != 0 &&
+	    ifp->if_hw_tsomaxsegcount < pmax->tsomaxsegcount)) {
+		pmax->tsomaxsegcount = ifp->if_hw_tsomaxsegcount;
+	}
+	if (pmax->tsomaxsegsize == 0 || (ifp->if_hw_tsomaxsegsize != 0 &&
+	    ifp->if_hw_tsomaxsegsize < pmax->tsomaxsegsize)) {
+		pmax->tsomaxsegsize = ifp->if_hw_tsomaxsegsize;
+	}
+}
+
+/*
+ * Update TSO limit of a network adapter.
+ *
+ * Returns zero if no change. Else non-zero.
+ */
+int
+if_hw_tsomax_update(struct ifnet *ifp, struct ifnet_hw_tsomax *pmax)
+{
+	int retval = 0;
+	if (ifp->if_hw_tsomax != pmax->tsomaxbytes) {
+		ifp->if_hw_tsomax = pmax->tsomaxbytes;
+		retval++;
+	}
+	if (ifp->if_hw_tsomaxsegsize != pmax->tsomaxsegsize) {
+		ifp->if_hw_tsomaxsegsize = pmax->tsomaxsegsize;
+		retval++;
+	}
+	if (ifp->if_hw_tsomaxsegcount != pmax->tsomaxsegcount) {
+		ifp->if_hw_tsomaxsegcount = pmax->tsomaxsegcount;
+		retval++;
+	}
+	return (retval);
+}
+
 static void
 if_attach_internal(struct ifnet *ifp, int vmove)
 {
@@ -657,13 +708,36 @@ if_attach_internal(struct ifnet *ifp, int vmove)
 		ifp->if_broadcastaddr = NULL;
 
 #if defined(INET) || defined(INET6)
-		/* Initialize to max value. */
-		if (ifp->if_hw_tsomax == 0)
-			ifp->if_hw_tsomax = min(IP_MAXPACKET, 32 * MCLBYTES -
+		/* Use defaults for TSO, if nothing is set */
+		if (ifp->if_hw_tsomax == 0 &&
+		    ifp->if_hw_tsomaxsegcount == 0 &&
+		    ifp->if_hw_tsomaxsegsize == 0) {
+			/*
+			 * The TSO defaults needs to be such that an
+			 * NFS mbuf list of 35 mbufs totalling just
+			 * below 64K works and that a chain of mbufs
+			 * can be defragged into at most 32 segments:
+			 */
+			ifp->if_hw_tsomax = min(IP_MAXPACKET, (32 * MCLBYTES) -
 			    (ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN));
-		KASSERT(ifp->if_hw_tsomax <= IP_MAXPACKET &&
-		    ifp->if_hw_tsomax >= IP_MAXPACKET / 8,
-		    ("%s: tsomax outside of range", __func__));
+			ifp->if_hw_tsomaxsegcount = 35;
+			ifp->if_hw_tsomaxsegsize = 2048;	/* 2K */
+
+			/* XXX some drivers set IFCAP_TSO after ethernet attach */
+			if (ifp->if_capabilities & IFCAP_TSO) {
+				if_printf(ifp, "Using defaults for TSO: %u/%u/%u\n",
+				    ifp->if_hw_tsomax,
+				    ifp->if_hw_tsomaxsegcount,
+				    ifp->if_hw_tsomaxsegsize);
+			}
+		}
+		/*
+		 * If the "if_hw_tsomax" limit is set, check if it is
+		 * too small:
+		 */
+		KASSERT(ifp->if_hw_tsomax == 0 ||
+		    ifp->if_hw_tsomax >= (IP_MAXPACKET / 8),
+		    ("%s: if_hw_tsomax is outside of range", __func__));
 #endif
 	}
 #ifdef VIMAGE
diff --git a/sys/net/if_lagg.c b/sys/net/if_lagg.c
index 58b7207..7218e29 100644
--- a/sys/net/if_lagg.c
+++ b/sys/net/if_lagg.c
@@ -448,23 +448,18 @@ lagg_capabilities(struct lagg_softc *sc)
 	struct lagg_port *lp;
 	int cap = ~0, ena = ~0;
 	u_long hwa = ~0UL;
-#if defined(INET) || defined(INET6)
-	u_int hw_tsomax = IP_MAXPACKET;	/* Initialize to the maximum value. */
-#else
-	u_int hw_tsomax = ~0;	/* if_hw_tsomax is only for INET/INET6, but.. */
-#endif
+	struct ifnet_hw_tsomax hw_tsomax;
 
 	LAGG_WLOCK_ASSERT(sc);
 
+	memset(&hw_tsomax, 0, sizeof(hw_tsomax));
+
 	/* Get capabilities from the lagg ports */
 	SLIST_FOREACH(lp, &sc->sc_ports, lp_entries) {
 		cap &= lp->lp_ifp->if_capabilities;
 		ena &= lp->lp_ifp->if_capenable;
 		hwa &= lp->lp_ifp->if_hwassist;
-		/* Set to the minimum value of the lagg ports. */
-		if (lp->lp_ifp->if_hw_tsomax < hw_tsomax &&
-		    lp->lp_ifp->if_hw_tsomax > 0)
-			hw_tsomax = lp->lp_ifp->if_hw_tsomax;
+		if_hw_tsomax_common(lp->lp_ifp, &hw_tsomax);
 	}
 	cap = (cap == ~0 ? 0 : cap);
 	ena = (ena == ~0 ? 0 : ena);
@@ -473,11 +468,10 @@ lagg_capabilities(struct lagg_softc *sc)
 	if (sc->sc_ifp->if_capabilities != cap ||
 	    sc->sc_ifp->if_capenable != ena ||
 	    sc->sc_ifp->if_hwassist != hwa ||
-	    sc->sc_ifp->if_hw_tsomax != hw_tsomax) {
+	    if_hw_tsomax_update(sc->sc_ifp, &hw_tsomax) != 0) {
 		sc->sc_ifp->if_capabilities = cap;
 		sc->sc_ifp->if_capenable = ena;
 		sc->sc_ifp->if_hwassist = hwa;
-		sc->sc_ifp->if_hw_tsomax = hw_tsomax;
 		getmicrotime(&sc->sc_ifp->if_lastchange);
 
 		if (sc->sc_ifflags & IFF_DEBUG)
diff --git a/sys/net/if_var.h b/sys/net/if_var.h
index 677a688..9312728 100644
--- a/sys/net/if_var.h
+++ b/sys/net/if_var.h
@@ -116,6 +116,12 @@ struct	ifqueue {
 	struct	mtx ifq_mtx;
 };
 
+struct ifnet_hw_tsomax {
+	u_int	tsomaxbytes;	/* TSO total burst length limit in bytes */
+	u_int	tsomaxsegcount;	/* TSO maximum segment count */
+	u_int	tsomaxsegsize;	/* TSO maximum segment size in bytes */
+};
+
 /*
  * Structure defining a network interface.
  *
@@ -204,10 +210,11 @@ struct ifnet {
 	u_int	if_fib;			/* interface FIB */
 	u_char	if_alloctype;		/* if_type at time of allocation */
 
-	u_int	if_hw_tsomax;		/* tso burst length limit, the minimum
-					 * is (IP_MAXPACKET / 8).
-					 * XXXAO: Have to find a better place
-					 * for it eventually. */
+	u_int	if_hw_tsomax;		/* TSO total burst length
+					 * limit in bytes. A value of
+					 * zero means no limit. Have
+					 * to find a better place for
+					 * it eventually. */
 
 	/*
 	 * Spare fields are added so that we can modify sensitive data
@@ -215,7 +222,14 @@ struct ifnet {
 	 * be used with care where binary compatibility is required.
 	 */
 	char	if_cspare[3];
-	int	if_ispare[4];
+	int	if_ispare[2];
+
+	/*
+	 * TSO fields for segment limits. If a field is zero below,
+	 * there is no limit:
+	 */
+	u_int	if_hw_tsomaxsegcount;	/* TSO maximum segment count */
+	u_int	if_hw_tsomaxsegsize;	/* TSO maximum segment size in bytes */
 	void	*if_pspare[8];		/* 1 netmap, 7 TDB */
 };
 
@@ -967,6 +981,10 @@ int    ether_poll_register(poll_handler_t *h, struct ifnet *ifp);
 int    ether_poll_deregister(struct ifnet *ifp);
 #endif /* DEVICE_POLLING */
 
+/* TSO */
+void if_hw_tsomax_common(struct ifnet *, struct ifnet_hw_tsomax *);
+int if_hw_tsomax_update(struct ifnet *, struct ifnet_hw_tsomax *);
+
 #endif /* _KERNEL */
 
 #endif /* !_NET_IF_VAR_H_ */
diff --git a/sys/net/if_vlan.c b/sys/net/if_vlan.c
index 9d547a6..6a43e37 100644
--- a/sys/net/if_vlan.c
+++ b/sys/net/if_vlan.c
@@ -1483,6 +1483,7 @@ vlan_capabilities(struct ifvlan *ifv)
 {
 	struct ifnet *p = PARENT(ifv);
 	struct ifnet *ifp = ifv->ifv_ifp;
+	struct ifnet_hw_tsomax hw_tsomax;
 
 	TRUNK_LOCK_ASSERT(TRUNK(ifv));
 
@@ -1509,8 +1510,9 @@ vlan_capabilities(struct ifvlan *ifv)
 	 * propagate the hardware-assisted flag. TSO on VLANs
 	 * does not necessarily require hardware VLAN tagging.
 	 */
-	if (p->if_hw_tsomax > 0)
-		ifp->if_hw_tsomax = p->if_hw_tsomax;
+	memset(&hw_tsomax, 0, sizeof(hw_tsomax));
+	if_hw_tsomax_common(p, &hw_tsomax);
+	if_hw_tsomax_update(ifp, &hw_tsomax);
 	if (p->if_capabilities & IFCAP_VLAN_HWTSO)
 		ifp->if_capabilities |= p->if_capabilities & IFCAP_TSO;
 	if (p->if_capenable & IFCAP_VLAN_HWTSO) {
diff --git a/sys/netinet/tcp_input.c b/sys/netinet/tcp_input.c
index d8d550a..92b0d58 100644
--- a/sys/netinet/tcp_input.c
+++ b/sys/netinet/tcp_input.c
@@ -3623,6 +3623,8 @@ tcp_mss(struct tcpcb *tp, int offer)
 	if (cap.ifcap & CSUM_TSO) {
 		tp->t_flags |= TF_TSO;
 		tp->t_tsomax = cap.tsomax;
+		tp->t_tsomaxsegcount = cap.tsomaxsegcount;
+		tp->t_tsomaxsegsize = cap.tsomaxsegsize;
 	}
 }
 
diff --git a/sys/netinet/tcp_output.c b/sys/netinet/tcp_output.c
index 8abb16d..c0e2230 100644
--- a/sys/netinet/tcp_output.c
+++ b/sys/netinet/tcp_output.c
@@ -774,28 +774,112 @@ send:
 		flags &= ~TH_FIN;
 
 		if (tso) {
+			u_int if_hw_tsomax;
+			u_int if_hw_tsomaxsegcount;
+			u_int if_hw_tsomaxsegsize;
+			struct mbuf *mb;
+			u_int moff;
+			int max_len;
+
+			/* extract TSO information */
+			if_hw_tsomax = tp->t_tsomax;
+			if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
+			if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
+
+			/*
+			 * Limit a TSO burst to prevent it from
+			 * overflowing or exceeding the maximum length
+			 * allowed by the network interface:
+			 */
 			KASSERT(ipoptlen == 0,
 			    ("%s: TSO can't do IP options", __func__));
 
 			/*
-			 * Limit a burst to t_tsomax minus IP,
-			 * TCP and options length to keep ip->ip_len
-			 * from overflowing or exceeding the maximum
-			 * length allowed by the network interface.
+			 * Check if we should limit by maximum payload
+			 * length:
 			 */
-			if (len > tp->t_tsomax - hdrlen) {
-				len = tp->t_tsomax - hdrlen;
-				sendalot = 1;
+			if (if_hw_tsomax != 0) {
+				/* compute maximum TSO length */
+				max_len = (if_hw_tsomax - hdrlen);
+				if (max_len <= 0) {
+					len = 0;
+				} else if (len > (u_int)max_len) {
+					sendalot = 1;
+					len = (u_int)max_len;
+				}
+			}
+
+			/*
+			 * Check if we should limit by maximum segment
+			 * size and count:
+			 */
+			if (if_hw_tsomaxsegcount != 0 &&
+			    if_hw_tsomaxsegsize != 0) {
+				max_len = 0;
+				mb = sbsndmbuf(&so->so_snd, off, &moff);
+
+				while (mb != NULL && (u_int)max_len < len) {
+					u_int mlen;
+					u_int frags;
+
+					/*
+					 * Get length of mbuf fragment
+					 * and how many hardware frags,
+					 * rounded up, it would use:
+					 */
+					mlen = (mb->m_len - moff);
+					frags = howmany(mlen,
+					    if_hw_tsomaxsegsize);
+
+					/* Handle special case: Zero Length Mbuf */
+					if (frags == 0)
+						frags = 1;
+
+					/*
+					 * Check if the fragment limit
+					 * will be reached or exceeded:
+					 */
+					if (frags >= if_hw_tsomaxsegcount) {
+						max_len += min(mlen,
+						    if_hw_tsomaxsegcount *
+						    if_hw_tsomaxsegsize);
+						break;
+					}
+					max_len += mlen;
+					if_hw_tsomaxsegcount -= frags;
+					moff = 0;
+					mb = mb->m_next;
+				}
+				if (max_len <= 0) {
+					len = 0;
+				} else if (len > (u_int)max_len) {
+					sendalot = 1;
+					len = (u_int)max_len;
+				}
 			}
 
 			/*
 			 * Prevent the last segment from being
-			 * fractional unless the send sockbuf can
-			 * be emptied.
+			 * fractional unless the send sockbuf can be
+			 * emptied:
+			 */
+			max_len = (tp->t_maxopd - optlen);
+			if ((off + len) < so->so_snd.sb_cc) {
+				moff = len % (u_int)max_len;
+				if (moff != 0) {
+					len -= moff;
+					sendalot = 1;
+				}
+			}
+
+			/*
+			 * In case there are too many small fragments
+			 * don't use TSO:
 			 */
-			if (sendalot && off + len < so->so_snd.sb_cc) {
-				len -= len % (tp->t_maxopd - optlen);
+			if (len <= (u_int)max_len) {
+				len = (u_int)max_len;
 				sendalot = 1;
+				tso = 0;
 			}
 
 			/*
diff --git a/sys/netinet/tcp_subr.c b/sys/netinet/tcp_subr.c
index 0c15314..c492de7 100644
--- a/sys/netinet/tcp_subr.c
+++ b/sys/netinet/tcp_subr.c
@@ -1808,6 +1808,8 @@ tcp_maxmtu(struct in_conninfo *inc, struct tcp_ifcap *cap)
 			    ifp->if_hwassist & CSUM_TSO) {
 				cap->ifcap |= CSUM_TSO;
 				cap->tsomax = ifp->if_hw_tsomax;
+				cap->tsomaxsegcount = ifp->if_hw_tsomaxsegcount;
+				cap->tsomaxsegsize = ifp->if_hw_tsomaxsegsize;
 			}
 		}
 		RTFREE(sro.ro_rt);
@@ -1847,6 +1849,8 @@ tcp_maxmtu6(struct in_conninfo *inc, struct tcp_ifcap *cap)
 			    ifp->if_hwassist & CSUM_TSO) {
 				cap->ifcap |= CSUM_TSO;
 				cap->tsomax = ifp->if_hw_tsomax;
+				cap->tsomaxsegcount = ifp->if_hw_tsomaxsegcount;
+				cap->tsomaxsegsize = ifp->if_hw_tsomaxsegsize;
 			}
 		}
 		RTFREE(sro6.ro_rt);
diff --git a/sys/netinet/tcp_var.h b/sys/netinet/tcp_var.h
index 10b1096..72bb7ae 100644
--- a/sys/netinet/tcp_var.h
+++ b/sys/netinet/tcp_var.h
@@ -208,13 +208,15 @@ struct tcpcb {
 	u_int	t_keepintvl;		/* interval between keepalives */
 	u_int	t_keepcnt;		/* number of keepalives before close */
 
-	u_int	t_tsomax;		/* tso burst length limit */
+	u_int	t_tsomax;		/* TSO total burst length limit in bytes */
 	u_int	t_pmtud_saved_maxopd;	/* pre-blackhole MSS */
 	u_int	t_flags2;		/* More tcpcb flags storage */
 
 	uint32_t t_ispare[6];		/* 5 UTO, 1 TBD */
 	void	*t_pspare2[4];		/* 1 TCP_SIGNATURE, 3 TBD */
-	uint64_t _pad[6];		/* 6 TBD (1-2 CC/RTT?) */
+	uint64_t _pad[5];		/* 5 TBD (1-2 CC/RTT?) */
+	uint32_t t_tsomaxsegcount;	/* TSO maximum segment count */
+	uint32_t t_tsomaxsegsize;	/* TSO maximum segment size in bytes */
 };
 
 /*
@@ -342,6 +344,8 @@ struct hc_metrics_lite {	/* must stay in sync with hc_metrics */
 struct tcp_ifcap {
 	int	ifcap;
 	u_int	tsomax;
+	u_int	tsomaxsegcount;
+	u_int	tsomaxsegsize;
 };
 
 #ifndef _NETINET_IN_PCB_H_
diff --git a/sys/ofed/drivers/net/mlx4/en_netdev.c b/sys/ofed/drivers/net/mlx4/en_netdev.c
index cd89a22..32c9948 100644
--- a/sys/ofed/drivers/net/mlx4/en_netdev.c
+++ b/sys/ofed/drivers/net/mlx4/en_netdev.c
@@ -2106,12 +2106,12 @@ int mlx4_en_init_netdev(struct mlx4_en_dev *mdev, int port,
 
 	if (mdev->LSO_support)
 		dev->if_capabilities |= IFCAP_TSO4 | IFCAP_TSO6 | IFCAP_VLAN_HWTSO;
-#if 0
+
 	/* set TSO limits so that we don't have to drop TX packets */
 	dev->if_hw_tsomax = 65536 - (ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN);
 	dev->if_hw_tsomaxsegcount = 16;
 	dev->if_hw_tsomaxsegsize = 65536;       /* XXX can do up to 4GByte */
-#endif
+
 	dev->if_capenable = dev->if_capabilities;
 
 	dev->if_hwassist = 0;
diff --git a/sys/sys/sockbuf.h b/sys/sys/sockbuf.h
index 877b530..ef80e9c 100644
--- a/sys/sys/sockbuf.h
+++ b/sys/sys/sockbuf.h
@@ -158,6 +158,8 @@ int	sbreserve_locked(struct sockbuf *sb, u_long cc, struct socket *so,
 	    struct thread *td);
 struct mbuf *
 	sbsndptr(struct sockbuf *sb, u_int off, u_int len, u_int *moff);
+struct mbuf *
+	sbsndmbuf(struct sockbuf *sb, u_int off, u_int *moff);
 void	sbtoxsockbuf(struct sockbuf *sb, struct xsockbuf *xsb);
 int	sbwait(struct sockbuf *sb);
 int	sblock(struct sockbuf *sb, int flags);