Improve transmit sending offload, TSO, algorithm in general.

The current TSO limitation feature only takes the total number of
bytes in an mbuf chain into account and does not limit by the number
of mbufs in a chain. Some kinds of hardware is limited by two
factors. One is the fragment length and the second is the fragment
count. Both of these limits need to be taken into account when doing
TSO. Else some kinds of hardware might have to drop completely valid
mbuf chains because they cannot loaded into the given hardware's DMA
engine. The new way of doing TSO limitation has been made backwards
compatible as input from other FreeBSD developers and will use
defaults for values not set.

Reviewed by:	adrian, rmacklem
Sponsored by:	Mellanox Technologies
MFC after:	1 week

sys/dev/oce/oce_if.c

@@ -1731,7 +1731,9 @@ oce_attach_ifp(POCE_SOFTC sc)
 	sc->ifp->if_baudrate = 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);

sys/dev/oce/oce_if.h

@@ -152,7 +152,6 @@ 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

sys/dev/vmware/vmxnet3/if_vmx.c

@@ -1722,7 +1722,9 @@ 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;

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

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);

sys/kern/uipc_sockbuf.c

@@ -1014,6 +1014,37 @@ sbsndptr(struct sockbuf *sb, u_int off, u_int len, u_int *moff)
 	return (ret);
 }
 
+/*
+ * 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.

sys/net/if.c

@@ -584,6 +584,57 @@ if_attach(struct ifnet *ifp)
 	if_attach_internal(ifp, 0);
 }
 
+/*
+ * Compute the least common TSO limit.
+ */
+void
+if_hw_tsomax_common(if_t 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(if_t 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)
 {
@@ -659,13 +710,36 @@ 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

sys/net/if_lagg.c

@@ -454,23 +454,18 @@ 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);
@@ -479,11 +474,10 @@ 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)

sys/net/if_var.h

@@ -119,6 +119,12 @@ typedef void (*if_qflush_fn_t)(if_t);
 typedef int (*if_transmit_fn_t)(if_t, struct mbuf *);
 typedef	uint64_t (*if_get_counter_t)(if_t, ift_counter);
 
+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.
  *
@@ -222,10 +228,11 @@ struct ifnet {
 	if_get_counter_t if_get_counter; /* get counter values */
 
 	/* Stuff that's only temporary and doesn't belong here. */
-	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. */
 	/*
 	 * Old, racy and expensive statistics, should not be used in
 	 * new drivers.
@@ -243,6 +250,10 @@ struct ifnet {
 	uint64_t	if_oqdrops;	/* dropped on output */
 	uint64_t	if_noproto;	/* destined for unsupported protocol */
 
+	/* 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 */
+
 	/*
 	 * Spare fields to be added before branching a stable branch, so
 	 * that structure can be enhanced without changing the kernel
@@ -596,5 +607,9 @@ struct mbuf* drbr_dequeue_drv(if_t ifp, struct buf_ring *br);
 int drbr_needs_enqueue_drv(if_t ifp, struct buf_ring *br);
 int drbr_enqueue_drv(if_t ifp, struct buf_ring *br, struct mbuf *m);
 
+/* TSO */
+void if_hw_tsomax_common(if_t ifp, struct ifnet_hw_tsomax *);
+int if_hw_tsomax_update(if_t ifp, struct ifnet_hw_tsomax *);
+
 #endif /* _KERNEL */
 #endif /* !_NET_IF_VAR_H_ */

sys/net/if_vlan.c

@@ -1536,6 +1536,7 @@ 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));
 
@@ -1562,8 +1563,9 @@ 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) {

sys/netinet/tcp_input.c

@@ -3591,6 +3591,8 @@ 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;
 	}
 }
 

sys/netinet/tcp_output.c

@@ -767,28 +767,113 @@ tcp_output(struct tcpcb *tp)
 		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 cur_length;
+					u_int cur_frags;
+
+					/*
+					 * Get length of mbuf fragment
+					 * and how many hardware
+					 * frags, rounded up, it would
+					 * use:
+					 */
+					cur_length = (mb->m_len - moff);
+					cur_frags = (cur_length + if_hw_tsomaxsegsize -
+					    1) / if_hw_tsomaxsegsize;
+
+					/* Handle special case: Zero Length Mbuf */
+					if (cur_frags == 0)
+						cur_frags = 1;
+
+					/*
+					 * Check if the fragment limit
+					 * will be reached or
+					 * exceeded:
+					 */
+					if (cur_frags >= if_hw_tsomaxsegcount) {
+						max_len += min(cur_length,
+						    if_hw_tsomaxsegcount *
+						    if_hw_tsomaxsegsize);
+						break;
+					}
+					max_len += cur_length;
+					if_hw_tsomaxsegcount -= cur_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:
 			 */
-			if (sendalot && off + len < so->so_snd.sb_cc) {
-				len -= len % (tp->t_maxopd - optlen);
+			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 (len <= (u_int)max_len) {
+				len = (u_int)max_len;
 				sendalot = 1;
+				tso = 0;
 			}
 
 			/*

sys/netinet/tcp_subr.c

@@ -1819,6 +1819,8 @@ 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);
@@ -1858,6 +1860,8 @@ 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);

sys/netinet/tcp_var.h

@@ -199,9 +199,12 @@ 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 */
+
+	uint32_t t_ispare[6];		/* 5 UTO, 1 TBD */
+	uint32_t t_tsomaxsegcount;	/* TSO maximum segment count */
+	uint32_t t_tsomaxsegsize;	/* TSO maximum segment size in bytes */
 
-	uint32_t t_ispare[8];		/* 5 UTO, 3 TBD */
 	void	*t_pspare2[4];		/* 1 TCP_SIGNATURE, 3 TBD */
 	uint64_t _pad[6];		/* 6 TBD (1-2 CC/RTT?) */
 };
@@ -324,6 +327,8 @@ 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_

sys/ofed/drivers/net/mlx4/en_netdev.c

@@ -673,6 +673,11 @@ int mlx4_en_do_start_port(struct net_device *dev)
 	else
 		priv->rx_csum = 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 */
+
 	err = mlx4_wol_read(priv->mdev->dev, &config, priv->port);
 	if (err) {
 		en_err(priv, "Failed to get WoL info, unable to modify\n");

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);