Add support for offloading IP/TCP/UDP checksums to NIC hardware which

supports them.

17 files changed, 648 insertions, 176 deletions

diff --git a/sys/alpha/alpha/in_cksum.c b/sys/alpha/alpha/in_cksum.c
index 945021359587..e8b6872c01b2 100644
--- a/sys/alpha/alpha/in_cksum.c
+++ b/sys/alpha/alpha/in_cksum.c
@@ -69,6 +69,7 @@
 	sum = l_util.s[0] + l_util.s[1];				  \
 	ADDCARRY(sum);							  \
     }
+#define INVERT		sum == 0xffff ? sum : ~sum & 0xffff
 
 static const u_int32_t in_masks[] = {
 	/*0 bytes*/ /*1 byte*/	/*2 bytes*/ /*3 bytes*/
@@ -173,6 +174,27 @@ in_cksumdata(buf, len)
 	return sum;
 }
 
+u_short
+in_addword(u_short a, u_short b)
+{
+	u_int64_t sum = a + b;
+
+	ADDCARRY(sum);
+	return (sum);
+}
+
+u_short
+in_pseudo(u_int32_t a, u_int32_t b, u_int32_t c)
+{
+	u_int64_t sum;
+	union q_util q_util;
+	union l_util l_util;
+		    
+	sum = (u_int64_t) a + b + c;
+	REDUCE16;
+	return (sum);
+}
+
 int
 in_cksum(m, len)
 	register struct mbuf *m;
@@ -204,6 +226,50 @@ in_cksum(m, len)
 	return (~sum & 0xffff);
 }
 
+u_short
+in_cksum_skip(m, len, skip)
+	struct mbuf *m;
+	int len;
+	int skip;
+{
+	u_int64_t sum = 0;
+	int mlen = 0;
+	int clen = 0;
+	caddr_t addr;
+	union q_util q_util;
+	union l_util l_util;
+
+        len -= skip;
+        for (; skip && m; m = m->m_next) {
+                if (m->m_len > skip) {
+                        mlen = m->m_len - skip;
+			addr = mtod(m, caddr_t) + skip;
+                        goto skip_start;
+                } else {
+                        skip -= m->m_len;
+                }
+        }
+
+	for (; m && len; m = m->m_next) {
+		if (m->m_len == 0)
+			continue;
+		mlen = m->m_len;
+		addr = mtod(m, caddr_t);
+skip_start:
+		if (len < mlen)
+			mlen = len;
+		if ((clen ^ (long) addr) & 1)
+		    sum += in_cksumdata(addr, mlen) << 8;
+		else
+		    sum += in_cksumdata(addr, mlen);
+
+		clen += mlen;
+		len -= mlen;
+	}
+	REDUCE16;
+	return (INVERT);
+}
+
 u_int in_cksum_hdr(ip)
     const struct ip *ip;
 {
diff --git a/sys/alpha/include/in_cksum.h b/sys/alpha/include/in_cksum.h
index e00477983832..aa61a5fb6eb8 100644
--- a/sys/alpha/include/in_cksum.h
+++ b/sys/alpha/include/in_cksum.h
@@ -73,6 +73,9 @@ typedef	unsigned in_psum_t;
 u_int in_cksum_hdr(const struct ip *ip);
 in_psum_t in_cksum_partial(in_psum_t psum, const u_short *w, int len);
 int	in_cksum_finalize(in_psum_t psum);
+u_short	in_addword(u_short sum, u_short b);
+u_short	in_pseudo(u_int sum, u_int b, u_int c);
+u_short	in_cksum_skip(struct mbuf *m, int len, int skip);
 #endif
 
 #endif /* _MACHINE_IN_CKSUM_H_ */
diff --git a/sys/dev/ti/if_ti.c b/sys/dev/ti/if_ti.c
index 5b00f305d778..de9c3bd33856 100644
--- a/sys/dev/ti/if_ti.c
+++ b/sys/dev/ti/if_ti.c
@@ -122,9 +122,7 @@
 #include <pci/ti_fw.h>
 #include <pci/ti_fw2.h>
 
-#ifdef M_HWCKSUM
-/*#define TI_CSUM_OFFLOAD*/
-#endif
+#define TI_CSUM_FEATURES	(CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_IP_FRAGS)
 
 #if !defined(lint)
 static const char rcsid[] =
@@ -792,11 +790,9 @@ static int ti_newbuf_std(sc, i, m)
 	r = &sc->ti_rdata->ti_rx_std_ring[i];
 	TI_HOSTADDR(r->ti_addr) = vtophys(mtod(m_new, caddr_t));
 	r->ti_type = TI_BDTYPE_RECV_BD;
-#ifdef TI_CSUM_OFFLOAD
-	r->ti_flags = TI_BDFLAG_TCP_UDP_CKSUM|TI_BDFLAG_IP_CKSUM;
-#else
 	r->ti_flags = 0;
-#endif
+	if (sc->arpcom.ac_if.if_hwassist)
+		r->ti_flags |= TI_BDFLAG_TCP_UDP_CKSUM | TI_BDFLAG_IP_CKSUM;
 	r->ti_len = m_new->m_len;
 	r->ti_idx = i;
 
@@ -835,9 +831,8 @@ static int ti_newbuf_mini(sc, i, m)
 	TI_HOSTADDR(r->ti_addr) = vtophys(mtod(m_new, caddr_t));
 	r->ti_type = TI_BDTYPE_RECV_BD;
 	r->ti_flags = TI_BDFLAG_MINI_RING;
-#ifdef TI_CSUM_OFFLOAD
-	r->ti_flags |= TI_BDFLAG_TCP_UDP_CKSUM|TI_BDFLAG_IP_CKSUM;
-#endif
+	if (sc->arpcom.ac_if.if_hwassist)
+		r->ti_flags |= TI_BDFLAG_TCP_UDP_CKSUM | TI_BDFLAG_IP_CKSUM;
 	r->ti_len = m_new->m_len;
 	r->ti_idx = i;
 
@@ -896,9 +891,8 @@ static int ti_newbuf_jumbo(sc, i, m)
 	TI_HOSTADDR(r->ti_addr) = vtophys(mtod(m_new, caddr_t));
 	r->ti_type = TI_BDTYPE_RECV_JUMBO_BD;
 	r->ti_flags = TI_BDFLAG_JUMBO_RING;
-#ifdef TI_CSUM_OFFLOAD
-	r->ti_flags |= TI_BDFLAG_TCP_UDP_CKSUM|TI_BDFLAG_IP_CKSUM;
-#endif
+	if (sc->arpcom.ac_if.if_hwassist)
+		r->ti_flags |= TI_BDFLAG_TCP_UDP_CKSUM | TI_BDFLAG_IP_CKSUM;
 	r->ti_len = m_new->m_len;
 	r->ti_idx = i;
 
@@ -1206,6 +1200,8 @@ static int ti_chipinit(sc)
 	/* Initialize link to down state. */
 	sc->ti_linkstat = TI_EV_CODE_LINK_DOWN;
 
+	sc->arpcom.ac_if.if_hwassist = TI_CSUM_FEATURES;
+
 	/* Set endianness before we access any non-PCI registers. */
 #if BYTE_ORDER == BIG_ENDIAN
 	CSR_WRITE_4(sc, TI_MISC_HOST_CTL,
@@ -1316,11 +1312,10 @@ static int ti_chipinit(sc)
 	 * Only allow 1 DMA channel to be active at a time.
 	 * I don't think this is a good idea, but without it
 	 * the firmware racks up lots of nicDmaReadRingFull
-	 * errors.
+	 * errors.  This is not compatible with hardware checksums.
 	 */
-#ifndef TI_CSUM_OFFLOAD
-	TI_SETBIT(sc, TI_GCR_OPMODE, TI_OPMODE_1_DMA_ACTIVE);
-#endif
+	if (sc->arpcom.ac_if.if_hwassist == 0)
+		TI_SETBIT(sc, TI_GCR_OPMODE, TI_OPMODE_1_DMA_ACTIVE);
 
 	/* Recommended settings from Tigon manual. */
 	CSR_WRITE_4(sc, TI_GCR_DMA_WRITECFG, TI_DMA_STATE_THRESH_8W);
@@ -1399,9 +1394,9 @@ static int ti_gibinit(sc)
 	TI_HOSTADDR(rcb->ti_hostaddr) = vtophys(&sc->ti_rdata->ti_rx_std_ring);
 	rcb->ti_max_len = TI_FRAMELEN;
 	rcb->ti_flags = 0;
-#ifdef TI_CSUM_OFFLOAD
-	rcb->ti_flags |= TI_RCB_FLAG_TCP_UDP_CKSUM|TI_RCB_FLAG_IP_CKSUM;
-#endif
+	if (sc->arpcom.ac_if.if_hwassist)
+		rcb->ti_flags |= TI_RCB_FLAG_TCP_UDP_CKSUM |
+		     TI_RCB_FLAG_IP_CKSUM | TI_RCB_FLAG_NO_PHDR_CKSUM;
 #if NVLAN > 0
 	rcb->ti_flags |= TI_RCB_FLAG_VLAN_ASSIST;
 #endif
@@ -1412,9 +1407,9 @@ static int ti_gibinit(sc)
 	    vtophys(&sc->ti_rdata->ti_rx_jumbo_ring);
 	rcb->ti_max_len = TI_JUMBO_FRAMELEN;
 	rcb->ti_flags = 0;
-#ifdef TI_CSUM_OFFLOAD
-	rcb->ti_flags |= TI_RCB_FLAG_TCP_UDP_CKSUM|TI_RCB_FLAG_IP_CKSUM;
-#endif
+	if (sc->arpcom.ac_if.if_hwassist)
+		rcb->ti_flags |= TI_RCB_FLAG_TCP_UDP_CKSUM |
+		     TI_RCB_FLAG_IP_CKSUM | TI_RCB_FLAG_NO_PHDR_CKSUM;
 #if NVLAN > 0
 	rcb->ti_flags |= TI_RCB_FLAG_VLAN_ASSIST;
 #endif
@@ -1432,9 +1427,9 @@ static int ti_gibinit(sc)
 		rcb->ti_flags = TI_RCB_FLAG_RING_DISABLED;
 	else
 		rcb->ti_flags = 0;
-#ifdef TI_CSUM_OFFLOAD
-	rcb->ti_flags |= TI_RCB_FLAG_TCP_UDP_CKSUM|TI_RCB_FLAG_IP_CKSUM;
-#endif
+	if (sc->arpcom.ac_if.if_hwassist)
+		rcb->ti_flags |= TI_RCB_FLAG_TCP_UDP_CKSUM |
+		     TI_RCB_FLAG_IP_CKSUM | TI_RCB_FLAG_NO_PHDR_CKSUM;
 #if NVLAN > 0
 	rcb->ti_flags |= TI_RCB_FLAG_VLAN_ASSIST;
 #endif
@@ -1474,6 +1469,9 @@ static int ti_gibinit(sc)
 #if NVLAN > 0
 	rcb->ti_flags |= TI_RCB_FLAG_VLAN_ASSIST;
 #endif
+	if (sc->arpcom.ac_if.if_hwassist)
+		rcb->ti_flags |= TI_RCB_FLAG_TCP_UDP_CKSUM |
+		     TI_RCB_FLAG_IP_CKSUM | TI_RCB_FLAG_NO_PHDR_CKSUM;
 	rcb->ti_max_len = TI_TX_RING_CNT;
 	if (sc->ti_hwrev == TI_HWREV_TIGON)
 		TI_HOSTADDR(rcb->ti_hostaddr) = TI_TX_RING_BASE;
@@ -1791,9 +1789,6 @@ static void ti_rxeof(sc)
 		u_int16_t		vlan_tag = 0;
 		int			have_tag = 0;
 #endif
-#ifdef TI_CSUM_OFFLOAD
-		struct ip		*ip;
-#endif
 
 		cur_rx =
 		    &sc->ti_rdata->ti_rx_return_ring[sc->ti_rx_saved_considx];
@@ -1876,12 +1871,13 @@ static void ti_rxeof(sc)
 		/* Remove header from mbuf and pass it on. */
 		m_adj(m, sizeof(struct ether_header));
 
-#ifdef TI_CSUM_OFFLOAD
-		ip = mtod(m, struct ip *);
-		if (!(cur_rx->ti_tcp_udp_cksum ^ 0xFFFF) &&
-		    !(ip->ip_off & htons(IP_MF | IP_OFFMASK | IP_RF)))
-			m->m_flags |= M_HWCKSUM;
-#endif
+		if (ifp->if_hwassist) {
+			m->m_pkthdr.csum_flags |= CSUM_IP_CHECKED |
+			    CSUM_DATA_VALID;
+			if ((cur_rx->ti_ip_cksum ^ 0xffff) == 0)
+				m->m_pkthdr.csum_flags |= CSUM_IP_VALID;
+			m->m_pkthdr.csum_data = cur_rx->ti_tcp_udp_cksum;
+		}
 
 #if NVLAN > 0
 		/*
@@ -2025,6 +2021,7 @@ static int ti_encap(sc, m_head, txidx)
 	struct ti_tx_desc	*f = NULL;
 	struct mbuf		*m;
 	u_int32_t		frag, cur, cnt = 0;
+	u_int16_t		csum_flags = 0;
 #if NVLAN > 0
 	struct ifvlan		*ifv = NULL;
 
@@ -2037,6 +2034,16 @@ static int ti_encap(sc, m_head, txidx)
 	m = m_head;
 	cur = frag = *txidx;
 
+	if (m_head->m_pkthdr.csum_flags) {
+		if (m_head->m_pkthdr.csum_flags & CSUM_IP)
+			csum_flags |= TI_BDFLAG_IP_CKSUM;
+		if (m_head->m_pkthdr.csum_flags & (CSUM_TCP | CSUM_UDP))
+			csum_flags |= TI_BDFLAG_TCP_UDP_CKSUM;
+		if (m_head->m_flags & M_LASTFRAG)
+			csum_flags |= TI_BDFLAG_IP_FRAG_END;
+		else if (m_head->m_flags & M_FRAG)
+			csum_flags |= TI_BDFLAG_IP_FRAG;
+	}
 	/*
  	 * Start packing the mbufs in this chain into
 	 * the fragment pointers. Stop when we run out
@@ -2064,7 +2071,7 @@ static int ti_encap(sc, m_head, txidx)
 				break;
 			TI_HOSTADDR(f->ti_addr) = vtophys(mtod(m, vm_offset_t));
 			f->ti_len = m->m_len;
-			f->ti_flags = 0;
+			f->ti_flags = csum_flags;
 #if NVLAN > 0
 			if (ifv != NULL) {
 				f->ti_flags |= TI_BDFLAG_VLAN_TAG;
@@ -2125,6 +2132,24 @@ static void ti_start(ifp)
 			break;
 
 		/*
+		 * XXX
+		 * safety overkill.  If this is a fragmented packet chain
+		 * with delayed TCP/UDP checksums, then only encapsulate
+		 * it if we have enough descriptors to handle the entire
+		 * chain at once.
+		 * (paranoia -- may not actually be needed)
+		 */
+		if (m_head->m_flags & M_FIRSTFRAG &&
+		    m_head->m_pkthdr.csum_flags & (CSUM_DELAY_DATA)) {
+			if ((TI_TX_RING_CNT - sc->ti_txcnt) <
+			    m_head->m_pkthdr.csum_data + 16) {
+				IF_PREPEND(&ifp->if_snd, m_head);
+				ifp->if_flags |= IFF_OACTIVE;
+				break;
+			}
+		}
+
+		/*
 		 * Pack the data into the transmit ring. If we
 		 * don't have room, set the OACTIVE flag and wait
 		 * for the NIC to drain the ring.
diff --git a/sys/i386/i386/in_cksum.c b/sys/i386/i386/in_cksum.c
index ada4f7116b06..856215a7eef8 100644
--- a/sys/i386/i386/in_cksum.c
+++ b/sys/i386/i386/in_cksum.c
@@ -57,6 +57,7 @@
 #undef	ADDCARRY
 #define ADDCARRY(x)     if ((x) > 0xffff) (x) -= 0xffff
 #define REDUCE          {sum = (sum & 0xffff) + (sum >> 16); ADDCARRY(sum);}
+#define INVERT		sum == 0xffff ? sum : ~sum & 0xffff
 
 /*
  * Thanks to gcc we don't have to guess
@@ -246,6 +247,194 @@ in_cksum(m, len)
 	return (~sum & 0xffff);
 }
 
+u_short
+in_cksum_skip(m, len, skip)
+	struct mbuf *m;
+	int len;
+	int skip;
+{
+	register u_short *w;
+	register unsigned sum = 0;
+	register int mlen = 0;
+	int byte_swapped = 0;
+	union { char	c[2]; u_short	s; } su;
+
+	len -= skip;
+	for (; skip && m; m = m->m_next) {
+		if (m->m_len > skip) {
+			mlen = m->m_len - skip;
+			w = (u_short *)(mtod(m, u_char *) + skip);
+			goto skip_start;
+		} else {
+			skip -= m->m_len;
+		}
+	}
+
+	for (;m && len; m = m->m_next) {
+		if (m->m_len == 0)
+			continue;
+		w = mtod(m, u_short *);
+		if (mlen == -1) {
+			/*
+			 * The first byte of this mbuf is the continuation
+			 * of a word spanning between this mbuf and the
+			 * last mbuf.
+			 */
+
+			/* su.c[0] is already saved when scanning previous
+			 * mbuf.  sum was REDUCEd when we found mlen == -1
+			 */
+			su.c[1] = *(u_char *)w;
+			sum += su.s;
+			w = (u_short *)((char *)w + 1);
+			mlen = m->m_len - 1;
+			len--;
+		} else
+			mlen = m->m_len;
+skip_start:
+		if (len < mlen)
+			mlen = len;
+		len -= mlen;
+		/*
+		 * Force to long boundary so we do longword aligned
+		 * memory operations
+		 */
+		if (3 & (int) w) {
+			REDUCE;
+			if ((1 & (int) w) && (mlen > 0)) {
+				sum <<= 8;
+				su.c[0] = *(char *)w;
+				w = (u_short *)((char *)w + 1);
+				mlen--;
+				byte_swapped = 1;
+			}
+			if ((2 & (int) w) && (mlen >= 2)) {
+				sum += *w++;
+				mlen -= 2;
+			}
+		}
+		/*
+		 * Advance to a 486 cache line boundary.
+		 */
+		if (4 & (int) w && mlen >= 4) {
+			ADD(0);
+			MOP;
+			w += 2;
+			mlen -= 4;
+		}
+		if (8 & (int) w && mlen >= 8) {
+			ADD(0);
+			ADDC(4);
+			MOP;
+			w += 4;
+			mlen -= 8;
+		}
+		/*
+		 * Do as much of the checksum as possible 32 bits at at time.
+		 * In fact, this loop is unrolled to make overhead from
+		 * branches &c small.
+		 */
+		mlen -= 1;
+		while ((mlen -= 32) >= 0) {
+			u_char junk;
+			/*
+			 * Add with carry 16 words and fold in the last
+			 * carry by adding a 0 with carry.
+			 *
+			 * The early ADD(16) and the LOAD(32) are to load
+			 * the next 2 cache lines in advance on 486's.  The
+			 * 486 has a penalty of 2 clock cycles for loading
+			 * a cache line, plus whatever time the external
+			 * memory takes to load the first word(s) addressed.
+			 * These penalties are unavoidable.  Subsequent
+			 * accesses to a cache line being loaded (and to
+			 * other external memory?) are delayed until the
+			 * whole load finishes.  These penalties are mostly
+			 * avoided by not accessing external memory for
+			 * 8 cycles after the ADD(16) and 12 cycles after
+			 * the LOAD(32).  The loop terminates when mlen
+			 * is initially 33 (not 32) to guaranteed that
+			 * the LOAD(32) is within bounds.
+			 */
+			ADD(16);
+			ADDC(0);
+			ADDC(4);
+			ADDC(8);
+			ADDC(12);
+			LOAD(32);
+			ADDC(20);
+			ADDC(24);
+			ADDC(28);
+			MOP;
+			w += 16;
+		}
+		mlen += 32 + 1;
+		if (mlen >= 32) {
+			ADD(16);
+			ADDC(0);
+			ADDC(4);
+			ADDC(8);
+			ADDC(12);
+			ADDC(20);
+			ADDC(24);
+			ADDC(28);
+			MOP;
+			w += 16;
+			mlen -= 32;
+		}
+		if (mlen >= 16) {
+			ADD(0);
+			ADDC(4);
+			ADDC(8);
+			ADDC(12);
+			MOP;
+			w += 8;
+			mlen -= 16;
+		}
+		if (mlen >= 8) {
+			ADD(0);
+			ADDC(4);
+			MOP;
+			w += 4;
+			mlen -= 8;
+		}
+		if (mlen == 0 && byte_swapped == 0)
+			continue;       /* worth 1% maybe ?? */
+		REDUCE;
+		while ((mlen -= 2) >= 0) {
+			sum += *w++;
+		}
+		if (byte_swapped) {
+			sum <<= 8;
+			byte_swapped = 0;
+			if (mlen == -1) {
+				su.c[1] = *(char *)w;
+				sum += su.s;
+				mlen = 0;
+			} else
+				mlen = -1;
+		} else if (mlen == -1)
+			/*
+			 * This mbuf has odd number of bytes.
+			 * There could be a word split betwen
+			 * this mbuf and the next mbuf.
+			 * Save the last byte (to prepend to next mbuf).
+			 */
+			su.c[0] = *(char *)w;
+	}
+
+	if (len)
+		printf("cksum: out of data\n");
+	if (mlen == -1) {
+		/* The last mbuf has odd # of bytes. Follow the
+		   standard (the odd byte is shifted left by 8 bits) */
+		su.c[1] = 0;
+		sum += su.s;
+	}
+	REDUCE;
+	return (INVERT);
+}
+
 /*
  * This is the exact same algorithm as above with a few exceptions:
  * (1) it is designed to operate on buffers, not mbufs
diff --git a/sys/i386/include/in_cksum.h b/sys/i386/include/in_cksum.h
index a4419591c9de..3c971ef914ff 100644
--- a/sys/i386/include/in_cksum.h
+++ b/sys/i386/include/in_cksum.h
@@ -81,6 +81,30 @@ in_cksum_update(struct ip *ip)
 	ip->ip_sum = htons(__tmpsum + (__tmpsum >> 16));
 }
 
+static __inline u_short
+in_addword(u_short sum, u_short b)
+{
+		    
+	__asm("addw %2, %0" : "=r" (sum) : "0" (sum), "r" (b));
+	__asm("adcw $0, %0" : "=r" (sum) : "0" (sum));
+
+	return (sum);
+}
+
+static __inline u_short
+in_pseudo(u_int sum, u_int b, u_int c)
+{
+		    
+	__asm("addl %2, %0" : "=r" (sum) : "0" (sum), "r" (b));
+	__asm("adcl %2, %0" : "=r" (sum) : "0" (sum), "r" (c));
+	__asm("adcl $0, %0" : "=r" (sum) : "0" (sum));
+
+	sum = (sum & 0xffff) + (sum >> 16);
+	if (sum > 0xffff)
+		sum -= 0xffff;
+	return (sum);
+}
+
 #else
 u_int in_cksum_hdr __P((const struct ip *));
 #define	in_cksum_update(ip) \
@@ -94,6 +118,7 @@ u_int in_cksum_hdr __P((const struct ip *));
 
 typedef	unsigned in_psum_t;
 #ifdef _KERNEL
+u_short	in_cksum_skip(struct mbuf *m, int len, int skip); 
 in_psum_t in_cksum_partial(in_psum_t psum, const u_short *w, int len);
 int	in_cksum_finalize(in_psum_t psum);
 #endif /* _KERNEL */
diff --git a/sys/net/if.h b/sys/net/if.h
index 0e0ff819e039..8afeac026853 100644
--- a/sys/net/if.h
+++ b/sys/net/if.h
@@ -72,8 +72,8 @@ struct if_data {
 	u_long	ifi_omcasts;		/* packets sent via multicast */
 	u_long	ifi_iqdrops;		/* dropped on input, this interface */
 	u_long	ifi_noproto;		/* destined for unsupported protocol */
-	u_long	ifi_recvtiming;		/* usec spent receiving when timing */
-	u_long	ifi_xmittiming;		/* usec spent xmitting when timing */
+	u_long	ifi_hwassist;		/* HW offload capabilities */
+	u_long	ifi_unused;		/* XXX was ifi_xmittiming */
 	struct	timeval ifi_lastchange;	/* time of last administrative change */
 };
 
diff --git a/sys/net/if_var.h b/sys/net/if_var.h
index 80de848d11aa..a59e0160246f 100644
--- a/sys/net/if_var.h
+++ b/sys/net/if_var.h
@@ -150,6 +150,7 @@ typedef void if_init_f_t __P((void *));
 #define	if_hdrlen	if_data.ifi_hdrlen
 #define	if_metric	if_data.ifi_metric
 #define	if_baudrate	if_data.ifi_baudrate
+#define	if_hwassist	if_data.ifi_hwassist
 #define	if_ipackets	if_data.ifi_ipackets
 #define	if_ierrors	if_data.ifi_ierrors
 #define	if_opackets	if_data.ifi_opackets
diff --git a/sys/netinet/ip_input.c b/sys/netinet/ip_input.c
index 5d1bff38d53d..0d9273d89ee1 100644
--- a/sys/netinet/ip_input.c
+++ b/sys/netinet/ip_input.c
@@ -324,10 +324,14 @@ ip_input(struct mbuf *m)
 		}
 		ip = mtod(m, struct ip *);
 	}
-	if (hlen == sizeof(struct ip)) {
-		sum = in_cksum_hdr(ip);
+	if (m->m_pkthdr.csum_flags & CSUM_IP_CHECKED) {
+		sum = !(m->m_pkthdr.csum_flags & CSUM_IP_VALID);
 	} else {
-		sum = in_cksum(m, hlen);
+		if (hlen == sizeof(struct ip)) {
+			sum = in_cksum_hdr(ip);
+		} else {
+			sum = in_cksum(m, hlen);
+		}
 	}
 	if (sum) {
 		ipstat.ips_badsum++;
@@ -841,6 +845,9 @@ ip_reass(m, fp, where)
 	 * our data already.  If so, drop the data from the incoming
 	 * segment.  If it provides all of our data, drop us, otherwise
 	 * stick new segment in the proper place.
+	 *
+	 * If some of the data is dropped from the the preceding
+	 * segment, then it's checksum is invalidated.
 	 */
 	if (p) {
 		i = GETIP(p)->ip_off + GETIP(p)->ip_len - ip->ip_off;
@@ -848,6 +855,7 @@ ip_reass(m, fp, where)
 			if (i >= ip->ip_len)
 				goto dropfrag;
 			m_adj(m, i);
+			m->m_pkthdr.csum_flags = 0;
 			ip->ip_off += i;
 			ip->ip_len -= i;
 		}
@@ -870,6 +878,7 @@ ip_reass(m, fp, where)
 			GETIP(q)->ip_len -= i;
 			GETIP(q)->ip_off += i;
 			m_adj(q, i);
+			q->m_pkthdr.csum_flags = 0;
 			break;
 		}
 		nq = q->m_nextpkt;
@@ -927,6 +936,8 @@ inserted:
 		nq = q->m_nextpkt;
 		q->m_nextpkt = NULL;
 		m_cat(m, q);
+		m->m_pkthdr.csum_flags &= q->m_pkthdr.csum_flags;
+		m->m_pkthdr.csum_data += q->m_pkthdr.csum_data;
 	}
 
 #ifdef IPDIVERT
diff --git a/sys/netinet/ip_output.c b/sys/netinet/ip_output.c
index ee361ce401c9..b44765f909a3 100644
--- a/sys/netinet/ip_output.c
+++ b/sys/netinet/ip_output.c
@@ -96,6 +96,7 @@ static MALLOC_DEFINE(M_IPMOPTS, "ip_moptions", "internet multicast options");
 
 u_short ip_id;
 
+static void	in_delayed_cksum(struct mbuf *m);
 static struct mbuf *ip_insertoptions __P((struct mbuf *, struct mbuf *, int *));
 static void	ip_mloopback
 	__P((struct ifnet *, struct mbuf *, struct sockaddr_in *, int));
@@ -132,7 +133,7 @@ ip_output(m0, opt, ro, flags, imo)
 	int len, off, error = 0;
 	struct sockaddr_in *dst;
 	struct in_ifaddr *ia;
-	int isbroadcast;
+	int isbroadcast, sw_csum;
 #ifdef IPSEC
 	struct route iproute;
 	struct socket *so = NULL;
@@ -692,6 +693,15 @@ pass:
 		state.ro = ro;
 	state.dst = (struct sockaddr *)dst;
 
+	/*
+	 * XXX
+	 * delayed checksums are not currently compatible with IPsec
+	 */
+	if (m->m_pkthdr.csum_flags & CSUM_DELAY_DATA) {
+		in_delayed_cksum(m);
+		m->m_pkthdr.csum_flags &= ~CSUM_DELAY_DATA;
+	}
+
 	error = ipsec4_output(&state, sp, flags);
 
 	m = state.m;
@@ -754,17 +764,29 @@ pass:
 skip_ipsec:
 #endif /*IPSEC*/
 
+	sw_csum = m->m_pkthdr.csum_flags | CSUM_IP;
+	m->m_pkthdr.csum_flags = sw_csum & ifp->if_hwassist;
+	sw_csum &= ~ifp->if_hwassist;
+	if (sw_csum & CSUM_DELAY_DATA) {
+		in_delayed_cksum(m);
+		sw_csum &= ~CSUM_DELAY_DATA;
+	}
+
 	/*
-	 * If small enough for interface, can just send directly.
+	 * If small enough for interface, or the interface will take
+	 * care of the fragmentation for us, can just send directly.
 	 */
-	if ((u_short)ip->ip_len <= ifp->if_mtu) {
+	if ((u_short)ip->ip_len <= ifp->if_mtu ||
+	    ifp->if_hwassist & CSUM_FRAGMENT) {
 		ip->ip_len = htons((u_short)ip->ip_len);
 		ip->ip_off = htons((u_short)ip->ip_off);
 		ip->ip_sum = 0;
-		if (ip->ip_vhl == IP_VHL_BORING) {
-			ip->ip_sum = in_cksum_hdr(ip);
-		} else {
-			ip->ip_sum = in_cksum(m, hlen);
+		if (sw_csum & CSUM_DELAY_IP) {
+			if (ip->ip_vhl == IP_VHL_BORING) {
+				ip->ip_sum = in_cksum_hdr(ip);
+			} else {
+				ip->ip_sum = in_cksum(m, hlen);
+			}
 		}
 		error = (*ifp->if_output)(ifp, m,
 				(struct sockaddr *)dst, ro->ro_rt);
@@ -797,9 +819,20 @@ skip_ipsec:
 		goto bad;
 	}
 
+	/*
+	 * if the interface will not calculate checksums on
+	 * fragmented packets, then do it here.
+	 */
+	if (m->m_pkthdr.csum_flags & CSUM_DELAY_DATA &&
+	    (ifp->if_hwassist & CSUM_IP_FRAGS) == 0) {
+		in_delayed_cksum(m);
+		m->m_pkthdr.csum_flags &= ~CSUM_DELAY_DATA;
+	}
+
     {
 	int mhlen, firstlen = len;
 	struct mbuf **mnext = &m->m_nextpkt;
+	int nfrags = 1;
 
 	/*
 	 * Loop through length of segment after first fragment,
@@ -814,7 +847,7 @@ skip_ipsec:
 			ipstat.ips_odropped++;
 			goto sendorfree;
 		}
-		m->m_flags |= (m0->m_flags & M_MCAST);
+		m->m_flags |= (m0->m_flags & M_MCAST) | M_FRAG;
 		m->m_data += max_linkhdr;
 		mhip = mtod(m, struct ip *);
 		*mhip = *ip;
@@ -840,17 +873,27 @@ skip_ipsec:
 		}
 		m->m_pkthdr.len = mhlen + len;
 		m->m_pkthdr.rcvif = (struct ifnet *)0;
+		m->m_pkthdr.csum_flags = m0->m_pkthdr.csum_flags;
 		mhip->ip_off = htons((u_short)mhip->ip_off);
 		mhip->ip_sum = 0;
-		if (mhip->ip_vhl == IP_VHL_BORING) {
-			mhip->ip_sum = in_cksum_hdr(mhip);
-		} else {
-			mhip->ip_sum = in_cksum(m, mhlen);
+		if (sw_csum & CSUM_DELAY_IP) {
+			if (mhip->ip_vhl == IP_VHL_BORING) {
+				mhip->ip_sum = in_cksum_hdr(mhip);
+			} else {
+				mhip->ip_sum = in_cksum(m, mhlen);
+			}
 		}
 		*mnext = m;
 		mnext = &m->m_nextpkt;
-		ipstat.ips_ofragments++;
+		nfrags++;
 	}
+	ipstat.ips_ofragments += nfrags;
+
+	/* set first/last markers for fragment chain */
+	m->m_flags |= M_LASTFRAG;
+	m0->m_flags |= M_FIRSTFRAG | M_FRAG;
+	m0->m_pkthdr.csum_data = nfrags;
+
 	/*
 	 * Update first fragment by trimming what's been copied out
 	 * and updating header, then send each fragment (in order).
@@ -861,10 +904,12 @@ skip_ipsec:
 	ip->ip_len = htons((u_short)m->m_pkthdr.len);
 	ip->ip_off = htons((u_short)(ip->ip_off | IP_MF));
 	ip->ip_sum = 0;
-	if (ip->ip_vhl == IP_VHL_BORING) {
-		ip->ip_sum = in_cksum_hdr(ip);
-	} else {
-		ip->ip_sum = in_cksum(m, hlen);
+	if (sw_csum & CSUM_DELAY_IP) {
+		if (ip->ip_vhl == IP_VHL_BORING) {
+			ip->ip_sum = in_cksum_hdr(ip);
+		} else {
+			ip->ip_sum = in_cksum(m, hlen);
+		}
 	}
 sendorfree:
 	for (m = m0; m; m = m0) {
@@ -898,6 +943,31 @@ bad:
 	goto done;
 }
 
+static void
+in_delayed_cksum(struct mbuf *m)
+{
+	struct ip *ip;
+	u_short csum, offset;
+
+	ip = mtod(m, struct ip *);
+	offset = IP_VHL_HL(ip->ip_vhl) << 2 ;
+	csum = in_cksum_skip(m, ip->ip_len, offset);
+	offset += m->m_pkthdr.csum_data;	/* checksum offset */
+
+	if (offset + sizeof(u_short) > m->m_len) {
+		printf("delayed m_pullup, m->len: %d  off: %d  p: %d\n",
+		    m->m_len, offset, ip->ip_p);
+		/*
+		 * XXX
+		 * this shouldn't happen, but if it does, the
+		 * correct behavior may be to insert the checksum
+		 * in the existing chain instead of rearranging it.
+		 */
+		m = m_pullup(m, offset + sizeof(u_short));
+	}
+	*(u_short *)(m->m_data + offset) = csum;
+}
+
 /*
  * Insert IP options into preformed packet.
  * Adjust IP destination as required for IP source routing,
diff --git a/sys/netinet/tcp_input.c b/sys/netinet/tcp_input.c
index 1d2d8c38e58f..dc8ba3c1b906 100644
--- a/sys/netinet/tcp_input.c
+++ b/sys/netinet/tcp_input.c
@@ -100,6 +100,8 @@ struct tcphdr tcp_savetcp;
 #include <netkey/key.h>
 #endif /*IPSEC*/
 
+#include <machine/in_cksum.h>
+
 MALLOC_DEFINE(M_TSEGQ, "tseg_qent", "TCP segment queue entry");
 
 static int	tcprexmtthresh = 3;
@@ -425,17 +427,27 @@ tcp_input(m, off0, proto)
 	}
 	ip = mtod(m, struct ip *);
 	ipov = (struct ipovly *)ip;
-
-	/*
-	 * Checksum extended TCP header and data.
-	 */
-	tlen = ip->ip_len;
-	len = sizeof (struct ip) + tlen;
-	bzero(ipov->ih_x1, sizeof(ipov->ih_x1));
-	ipov->ih_len = (u_short)tlen;
-	HTONS(ipov->ih_len);
 	th = (struct tcphdr *)((caddr_t)ip + off0);
-	th->th_sum = in_cksum(m, len);
+	tlen = ip->ip_len;
+
+	if (m->m_pkthdr.csum_flags & CSUM_DATA_VALID) {
+		if (m->m_pkthdr.csum_flags & CSUM_PSEUDO_HDR)
+                	th->th_sum = m->m_pkthdr.csum_data;
+		else
+	                th->th_sum = in_pseudo(ip->ip_src.s_addr,
+			    ip->ip_dst.s_addr, htonl(m->m_pkthdr.csum_data +
+			    ip->ip_len + IPPROTO_TCP));
+		th->th_sum ^= 0xffff;
+	} else {
+		/*
+		 * Checksum extended TCP header and data.
+		 */
+		len = sizeof (struct ip) + tlen;
+		bzero(ipov->ih_x1, sizeof(ipov->ih_x1));
+		ipov->ih_len = (u_short)tlen;
+		HTONS(ipov->ih_len);
+		th->th_sum = in_cksum(m, len);
+	}
 	if (th->th_sum) {
 		tcpstat.tcps_rcvbadsum++;
 		goto drop;
diff --git a/sys/netinet/tcp_output.c b/sys/netinet/tcp_output.c
index ae85e84ef327..238d25a6b15d 100644
--- a/sys/netinet/tcp_output.c
+++ b/sys/netinet/tcp_output.c
@@ -80,6 +80,8 @@
 #include <netinet6/ipsec.h>
 #endif /*IPSEC*/
 
+#include <machine/in_cksum.h>
+
 #ifdef notyet
 extern struct mbuf *m_copypack();
 #endif
@@ -645,6 +647,7 @@ send:
 	ip = mtod(m, struct ip *);
 	ipov = (struct ipovly *)ip;
 	th = (struct tcphdr *)(ip + 1);
+	/* this picks up the pseudo header (w/o the length) */
 	bcopy((caddr_t)tp->t_template->tt_ipgen, (caddr_t)ip,
 	      sizeof(struct ip));
 	bcopy((caddr_t)&tp->t_template->tt_t, (caddr_t)th,
@@ -722,15 +725,15 @@ send:
 	else
 #endif /* INET6 */
       {
+	m->m_pkthdr.csum_flags = CSUM_TCP;
+	m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
 	if (len + optlen)
-		ipov->ih_len = htons((u_short)(sizeof (struct tcphdr) +
-		    optlen + len));
-	th->th_sum = in_cksum(m, (int)(hdrlen + len));
-#ifdef INET6
-	/* Re-initialization for later version check */
-	ip->ip_v = IPVERSION;
-	
-#endif /* INET6 */
+		th->th_sum = in_addword(th->th_sum, 
+		    htons((u_short)(optlen + len)));
+
+	/* IP version must be set here for ipv4/ipv6 checking later */
+	KASSERT(ip->ip_v == IPVERSION,
+	    ("%s: IP version incorrect: %d", __FUNCTION__, ip->ip_v));
       }
 
 	/*
diff --git a/sys/netinet/tcp_reass.c b/sys/netinet/tcp_reass.c
index 1d2d8c38e58f..dc8ba3c1b906 100644
--- a/sys/netinet/tcp_reass.c
+++ b/sys/netinet/tcp_reass.c
@@ -100,6 +100,8 @@ struct tcphdr tcp_savetcp;
 #include <netkey/key.h>
 #endif /*IPSEC*/
 
+#include <machine/in_cksum.h>
+
 MALLOC_DEFINE(M_TSEGQ, "tseg_qent", "TCP segment queue entry");
 
 static int	tcprexmtthresh = 3;
@@ -425,17 +427,27 @@ tcp_input(m, off0, proto)
 	}
 	ip = mtod(m, struct ip *);
 	ipov = (struct ipovly *)ip;
-
-	/*
-	 * Checksum extended TCP header and data.
-	 */
-	tlen = ip->ip_len;
-	len = sizeof (struct ip) + tlen;
-	bzero(ipov->ih_x1, sizeof(ipov->ih_x1));
-	ipov->ih_len = (u_short)tlen;
-	HTONS(ipov->ih_len);
 	th = (struct tcphdr *)((caddr_t)ip + off0);
-	th->th_sum = in_cksum(m, len);
+	tlen = ip->ip_len;
+
+	if (m->m_pkthdr.csum_flags & CSUM_DATA_VALID) {
+		if (m->m_pkthdr.csum_flags & CSUM_PSEUDO_HDR)
+                	th->th_sum = m->m_pkthdr.csum_data;
+		else
+	                th->th_sum = in_pseudo(ip->ip_src.s_addr,
+			    ip->ip_dst.s_addr, htonl(m->m_pkthdr.csum_data +
+			    ip->ip_len + IPPROTO_TCP));
+		th->th_sum ^= 0xffff;
+	} else {
+		/*
+		 * Checksum extended TCP header and data.
+		 */
+		len = sizeof (struct ip) + tlen;
+		bzero(ipov->ih_x1, sizeof(ipov->ih_x1));
+		ipov->ih_len = (u_short)tlen;
+		HTONS(ipov->ih_len);
+		th->th_sum = in_cksum(m, len);
+	}
 	if (th->th_sum) {
 		tcpstat.tcps_rcvbadsum++;
 		goto drop;
diff --git a/sys/netinet/tcp_subr.c b/sys/netinet/tcp_subr.c
index ee62998c63ef..d4c552f03cce 100644
--- a/sys/netinet/tcp_subr.c
+++ b/sys/netinet/tcp_subr.c
@@ -39,6 +39,7 @@
 #include "opt_ipsec.h"
 #include "opt_tcpdebug.h"
 
+#include <stddef.h>
 #include <sys/param.h>
 #include <sys/systm.h>
 #include <sys/callout.h>
@@ -93,6 +94,8 @@
 #include <netinet6/ipsec.h>
 #endif /*IPSEC*/
 
+#include <machine/in_cksum.h>
+
 int 	tcp_mssdflt = TCP_MSS;
 SYSCTL_INT(_net_inet_tcp, TCPCTL_MSSDFLT, mssdflt, CTLFLAG_RW, 
     &tcp_mssdflt , 0, "Default TCP Maximum Segment Size");
@@ -242,17 +245,19 @@ tcp_template(tp)
 		ip6->ip6_plen = sizeof(struct tcphdr);
 		ip6->ip6_src = inp->in6p_laddr;
 		ip6->ip6_dst = inp->in6p_faddr;
+		n->tt_t.th_sum = 0;
 	} else
 #endif
       {
-	register struct ipovly *ipov;
-
-	ipov = (struct ipovly *)n->tt_ipgen;
-	bzero(ipov->ih_x1, sizeof(ipov->ih_x1));
-	ipov->ih_pr = IPPROTO_TCP;
-	ipov->ih_len = htons(sizeof (struct tcpiphdr) - sizeof (struct ip));
-	ipov->ih_src = inp->inp_laddr;
-	ipov->ih_dst = inp->inp_faddr;
+	struct ip *ip = (struct ip *)n->tt_ipgen;
+
+	bzero(ip, sizeof(struct ip));		/* XXX overkill? */
+	ip->ip_vhl = IP_VHL_BORING;
+	ip->ip_p = IPPROTO_TCP;
+	ip->ip_src = inp->inp_laddr;
+	ip->ip_dst = inp->inp_faddr;
+	n->tt_t.th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
+	    htons(sizeof(struct tcphdr) + IPPROTO_TCP));
       }
 	n->tt_t.th_sport = inp->inp_lport;
 	n->tt_t.th_dport = inp->inp_fport;
@@ -262,7 +267,6 @@ tcp_template(tp)
 	n->tt_t.th_off = 5;
 	n->tt_t.th_flags = 0;
 	n->tt_t.th_win = 0;
-	n->tt_t.th_sum = 0;
 	n->tt_t.th_urp = 0;
 	return (n);
 }
@@ -296,7 +300,6 @@ tcp_respond(tp, ipgen, th, m, ack, seq, flags)
 	struct route *ro = 0;
 	struct route sro;
 	struct ip *ip;
-	struct ipovly *ipov;
 	struct tcphdr *nth;
 #ifdef INET6
 	struct route_in6 *ro6 = 0;
@@ -311,7 +314,6 @@ tcp_respond(tp, ipgen, th, m, ack, seq, flags)
 	ip6 = ipgen;
 #endif /* INET6 */
 	ip = ipgen;
-	ipov = ipgen;
 
 	if (tp) {
 		if (!(flags & TH_RST)) {
@@ -358,7 +360,6 @@ tcp_respond(tp, ipgen, th, m, ack, seq, flags)
 	      {
 		bcopy((caddr_t)ip, mtod(m, caddr_t), sizeof(struct ip));
 		ip = mtod(m, struct ip *);
-		ipov = mtod(m, struct ipovly *);
 		nth = (struct tcphdr *)(ip + 1);
 	      }
 		bcopy((caddr_t)th, (caddr_t)nth, sizeof(struct tcphdr));
@@ -400,8 +401,9 @@ tcp_respond(tp, ipgen, th, m, ack, seq, flags)
 	} else
 #endif
       {
-	ipov->ih_len = htons((u_short)(sizeof (struct tcphdr) + tlen));
 	tlen += sizeof (struct tcpiphdr);
+	ip->ip_len = tlen;
+	ip->ip_ttl = ip_defttl;
       }
 	m->m_len = tlen;
 	m->m_pkthdr.len = tlen;
@@ -416,7 +418,6 @@ tcp_respond(tp, ipgen, th, m, ack, seq, flags)
 	else
 		nth->th_win = htons((u_short)win);
 	nth->th_urp = 0;
-	nth->th_sum = 0;
 #ifdef INET6
 	if (isipv6) {
 		nth->th_sum = in6_cksum(m, IPPROTO_TCP,
@@ -429,14 +430,10 @@ tcp_respond(tp, ipgen, th, m, ack, seq, flags)
 	} else
 #endif /* INET6 */
       {
-	bzero(ipov->ih_x1, sizeof(ipov->ih_x1));
-	nth->th_sum = in_cksum(m, tlen);
-#ifdef INET6
-	/* Re-initialization for later version check */
-	ip->ip_vhl = IP_MAKE_VHL(IPVERSION, 0);
-#endif /* INET6 */
-	ip->ip_len = tlen;
-	ip->ip_ttl = ip_defttl;
+        nth->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
+	    htons((u_short)(tlen - sizeof(struct ip) + ip->ip_p)));
+        m->m_pkthdr.csum_flags = CSUM_TCP;
+        m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
       }
 #ifdef TCPDEBUG
 	if (tp == NULL || (tp->t_inpcb->inp_socket->so_options & SO_DEBUG))
diff --git a/sys/netinet/tcp_timewait.c b/sys/netinet/tcp_timewait.c
index ee62998c63ef..d4c552f03cce 100644
--- a/sys/netinet/tcp_timewait.c
+++ b/sys/netinet/tcp_timewait.c
@@ -39,6 +39,7 @@
 #include "opt_ipsec.h"
 #include "opt_tcpdebug.h"
 
+#include <stddef.h>
 #include <sys/param.h>
 #include <sys/systm.h>
 #include <sys/callout.h>
@@ -93,6 +94,8 @@
 #include <netinet6/ipsec.h>
 #endif /*IPSEC*/
 
+#include <machine/in_cksum.h>
+
 int 	tcp_mssdflt = TCP_MSS;
 SYSCTL_INT(_net_inet_tcp, TCPCTL_MSSDFLT, mssdflt, CTLFLAG_RW, 
     &tcp_mssdflt , 0, "Default TCP Maximum Segment Size");
@@ -242,17 +245,19 @@ tcp_template(tp)
 		ip6->ip6_plen = sizeof(struct tcphdr);
 		ip6->ip6_src = inp->in6p_laddr;
 		ip6->ip6_dst = inp->in6p_faddr;
+		n->tt_t.th_sum = 0;
 	} else
 #endif
       {
-	register struct ipovly *ipov;
-
-	ipov = (struct ipovly *)n->tt_ipgen;
-	bzero(ipov->ih_x1, sizeof(ipov->ih_x1));
-	ipov->ih_pr = IPPROTO_TCP;
-	ipov->ih_len = htons(sizeof (struct tcpiphdr) - sizeof (struct ip));
-	ipov->ih_src = inp->inp_laddr;
-	ipov->ih_dst = inp->inp_faddr;
+	struct ip *ip = (struct ip *)n->tt_ipgen;
+
+	bzero(ip, sizeof(struct ip));		/* XXX overkill? */
+	ip->ip_vhl = IP_VHL_BORING;
+	ip->ip_p = IPPROTO_TCP;
+	ip->ip_src = inp->inp_laddr;
+	ip->ip_dst = inp->inp_faddr;
+	n->tt_t.th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
+	    htons(sizeof(struct tcphdr) + IPPROTO_TCP));
       }
 	n->tt_t.th_sport = inp->inp_lport;
 	n->tt_t.th_dport = inp->inp_fport;
@@ -262,7 +267,6 @@ tcp_template(tp)
 	n->tt_t.th_off = 5;
 	n->tt_t.th_flags = 0;
 	n->tt_t.th_win = 0;
-	n->tt_t.th_sum = 0;
 	n->tt_t.th_urp = 0;
 	return (n);
 }
@@ -296,7 +300,6 @@ tcp_respond(tp, ipgen, th, m, ack, seq, flags)
 	struct route *ro = 0;
 	struct route sro;
 	struct ip *ip;
-	struct ipovly *ipov;
 	struct tcphdr *nth;
 #ifdef INET6
 	struct route_in6 *ro6 = 0;
@@ -311,7 +314,6 @@ tcp_respond(tp, ipgen, th, m, ack, seq, flags)
 	ip6 = ipgen;
 #endif /* INET6 */
 	ip = ipgen;
-	ipov = ipgen;
 
 	if (tp) {
 		if (!(flags & TH_RST)) {
@@ -358,7 +360,6 @@ tcp_respond(tp, ipgen, th, m, ack, seq, flags)
 	      {
 		bcopy((caddr_t)ip, mtod(m, caddr_t), sizeof(struct ip));
 		ip = mtod(m, struct ip *);
-		ipov = mtod(m, struct ipovly *);
 		nth = (struct tcphdr *)(ip + 1);
 	      }
 		bcopy((caddr_t)th, (caddr_t)nth, sizeof(struct tcphdr));
@@ -400,8 +401,9 @@ tcp_respond(tp, ipgen, th, m, ack, seq, flags)
 	} else
 #endif
       {
-	ipov->ih_len = htons((u_short)(sizeof (struct tcphdr) + tlen));
 	tlen += sizeof (struct tcpiphdr);
+	ip->ip_len = tlen;
+	ip->ip_ttl = ip_defttl;
       }
 	m->m_len = tlen;
 	m->m_pkthdr.len = tlen;
@@ -416,7 +418,6 @@ tcp_respond(tp, ipgen, th, m, ack, seq, flags)
 	else
 		nth->th_win = htons((u_short)win);
 	nth->th_urp = 0;
-	nth->th_sum = 0;
 #ifdef INET6
 	if (isipv6) {
 		nth->th_sum = in6_cksum(m, IPPROTO_TCP,
@@ -429,14 +430,10 @@ tcp_respond(tp, ipgen, th, m, ack, seq, flags)
 	} else
 #endif /* INET6 */
       {
-	bzero(ipov->ih_x1, sizeof(ipov->ih_x1));
-	nth->th_sum = in_cksum(m, tlen);
-#ifdef INET6
-	/* Re-initialization for later version check */
-	ip->ip_vhl = IP_MAKE_VHL(IPVERSION, 0);
-#endif /* INET6 */
-	ip->ip_len = tlen;
-	ip->ip_ttl = ip_defttl;
+        nth->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
+	    htons((u_short)(tlen - sizeof(struct ip) + ip->ip_p)));
+        m->m_pkthdr.csum_flags = CSUM_TCP;
+        m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
       }
 #ifdef TCPDEBUG
 	if (tp == NULL || (tp->t_inpcb->inp_socket->so_options & SO_DEBUG))
diff --git a/sys/netinet/udp_usrreq.c b/sys/netinet/udp_usrreq.c
index 2ed9a363ebcc..b8700de91569 100644
--- a/sys/netinet/udp_usrreq.c
+++ b/sys/netinet/udp_usrreq.c
@@ -37,6 +37,7 @@
 #include "opt_ipsec.h"
 #include "opt_inet6.h"
 
+#include <stddef.h>
 #include <sys/param.h>
 #include <sys/systm.h>
 #include <sys/kernel.h>
@@ -76,6 +77,8 @@
 #include <netinet6/ipsec.h>
 #endif /*IPSEC*/
 
+#include <machine/in_cksum.h>
+
 /*
  * UDP protocol implementation.
  * Per RFC 768, August, 1980.
@@ -208,9 +211,19 @@ udp_input(m, off, proto)
 	 * Checksum extended UDP header and data.
 	 */
 	if (uh->uh_sum) {
-		bzero(((struct ipovly *)ip)->ih_x1, 9);
-		((struct ipovly *)ip)->ih_len = uh->uh_ulen;
-		uh->uh_sum = in_cksum(m, len + sizeof (struct ip));
+		if (m->m_pkthdr.csum_flags & CSUM_DATA_VALID) {
+			if (m->m_pkthdr.csum_flags & CSUM_PSEUDO_HDR)
+				uh->uh_sum = m->m_pkthdr.csum_data;
+			else
+	                	uh->uh_sum = in_pseudo(ip->ip_src.s_addr,
+				    ip->ip_dst.s_addr, htonl(ip->ip_len +
+				    m->m_pkthdr.csum_data + IPPROTO_UDP));
+			uh->uh_sum ^= 0xffff;
+		} else {
+			bzero(((struct ipovly *)ip)->ih_x1, 9);
+			((struct ipovly *)ip)->ih_len = uh->uh_ulen;
+			uh->uh_sum = in_cksum(m, len + sizeof (struct ip));
+		}
 		if (uh->uh_sum) {
 			udpstat.udps_badsum++;
 			m_freem(m);
@@ -679,22 +692,24 @@ udp_output(inp, m, addr, control, p)
 	 * and addresses and length put into network format.
 	 */
 	ui = mtod(m, struct udpiphdr *);
-	bzero(ui->ui_x1, sizeof(ui->ui_x1));
+	bzero(ui->ui_x1, sizeof(ui->ui_x1));	/* XXX still needed? */
 	ui->ui_pr = IPPROTO_UDP;
-	ui->ui_len = htons((u_short)len + sizeof (struct udphdr));
 	ui->ui_src = inp->inp_laddr;
 	ui->ui_dst = inp->inp_faddr;
 	ui->ui_sport = inp->inp_lport;
 	ui->ui_dport = inp->inp_fport;
-	ui->ui_ulen = ui->ui_len;
+	ui->ui_ulen = htons((u_short)len + sizeof(struct udphdr));
 
 	/*
-	 * Stuff checksum and output datagram.
+	 * Set up checksum and output datagram.
 	 */
-	ui->ui_sum = 0;
 	if (udpcksum) {
-	    if ((ui->ui_sum = in_cksum(m, sizeof (struct udpiphdr) + len)) == 0)
-		ui->ui_sum = 0xffff;
+        	ui->ui_sum = in_pseudo(ui->ui_src.s_addr, ui->ui_dst.s_addr,
+		    htons((u_short)len + sizeof(struct udphdr) + IPPROTO_UDP));
+		m->m_pkthdr.csum_flags = CSUM_UDP;
+		m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
+	} else {
+		ui->ui_sum = 0;
 	}
 	((struct ip *)ui)->ip_len = sizeof (struct udpiphdr) + len;
 	((struct ip *)ui)->ip_ttl = inp->inp_ip_ttl;	/* XXX */
diff --git a/sys/pci/if_ti.c b/sys/pci/if_ti.c
index 5b00f305d778..de9c3bd33856 100644
--- a/sys/pci/if_ti.c
+++ b/sys/pci/if_ti.c
@@ -122,9 +122,7 @@
 #include <pci/ti_fw.h>
 #include <pci/ti_fw2.h>
 
-#ifdef M_HWCKSUM
-/*#define TI_CSUM_OFFLOAD*/
-#endif
+#define TI_CSUM_FEATURES	(CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_IP_FRAGS)
 
 #if !defined(lint)
 static const char rcsid[] =
@@ -792,11 +790,9 @@ static int ti_newbuf_std(sc, i, m)
 	r = &sc->ti_rdata->ti_rx_std_ring[i];
 	TI_HOSTADDR(r->ti_addr) = vtophys(mtod(m_new, caddr_t));
 	r->ti_type = TI_BDTYPE_RECV_BD;
-#ifdef TI_CSUM_OFFLOAD
-	r->ti_flags = TI_BDFLAG_TCP_UDP_CKSUM|TI_BDFLAG_IP_CKSUM;
-#else
 	r->ti_flags = 0;
-#endif
+	if (sc->arpcom.ac_if.if_hwassist)
+		r->ti_flags |= TI_BDFLAG_TCP_UDP_CKSUM | TI_BDFLAG_IP_CKSUM;
 	r->ti_len = m_new->m_len;
 	r->ti_idx = i;
 
@@ -835,9 +831,8 @@ static int ti_newbuf_mini(sc, i, m)
 	TI_HOSTADDR(r->ti_addr) = vtophys(mtod(m_new, caddr_t));
 	r->ti_type = TI_BDTYPE_RECV_BD;
 	r->ti_flags = TI_BDFLAG_MINI_RING;
-#ifdef TI_CSUM_OFFLOAD
-	r->ti_flags |= TI_BDFLAG_TCP_UDP_CKSUM|TI_BDFLAG_IP_CKSUM;
-#endif
+	if (sc->arpcom.ac_if.if_hwassist)
+		r->ti_flags |= TI_BDFLAG_TCP_UDP_CKSUM | TI_BDFLAG_IP_CKSUM;
 	r->ti_len = m_new->m_len;
 	r->ti_idx = i;
 
@@ -896,9 +891,8 @@ static int ti_newbuf_jumbo(sc, i, m)
 	TI_HOSTADDR(r->ti_addr) = vtophys(mtod(m_new, caddr_t));
 	r->ti_type = TI_BDTYPE_RECV_JUMBO_BD;
 	r->ti_flags = TI_BDFLAG_JUMBO_RING;
-#ifdef TI_CSUM_OFFLOAD
-	r->ti_flags |= TI_BDFLAG_TCP_UDP_CKSUM|TI_BDFLAG_IP_CKSUM;
-#endif
+	if (sc->arpcom.ac_if.if_hwassist)
+		r->ti_flags |= TI_BDFLAG_TCP_UDP_CKSUM | TI_BDFLAG_IP_CKSUM;
 	r->ti_len = m_new->m_len;
 	r->ti_idx = i;
 
@@ -1206,6 +1200,8 @@ static int ti_chipinit(sc)
 	/* Initialize link to down state. */
 	sc->ti_linkstat = TI_EV_CODE_LINK_DOWN;
 
+	sc->arpcom.ac_if.if_hwassist = TI_CSUM_FEATURES;
+
 	/* Set endianness before we access any non-PCI registers. */
 #if BYTE_ORDER == BIG_ENDIAN
 	CSR_WRITE_4(sc, TI_MISC_HOST_CTL,
@@ -1316,11 +1312,10 @@ static int ti_chipinit(sc)
 	 * Only allow 1 DMA channel to be active at a time.
 	 * I don't think this is a good idea, but without it
 	 * the firmware racks up lots of nicDmaReadRingFull
-	 * errors.
+	 * errors.  This is not compatible with hardware checksums.
 	 */
-#ifndef TI_CSUM_OFFLOAD
-	TI_SETBIT(sc, TI_GCR_OPMODE, TI_OPMODE_1_DMA_ACTIVE);
-#endif
+	if (sc->arpcom.ac_if.if_hwassist == 0)
+		TI_SETBIT(sc, TI_GCR_OPMODE, TI_OPMODE_1_DMA_ACTIVE);
 
 	/* Recommended settings from Tigon manual. */
 	CSR_WRITE_4(sc, TI_GCR_DMA_WRITECFG, TI_DMA_STATE_THRESH_8W);
@@ -1399,9 +1394,9 @@ static int ti_gibinit(sc)
 	TI_HOSTADDR(rcb->ti_hostaddr) = vtophys(&sc->ti_rdata->ti_rx_std_ring);
 	rcb->ti_max_len = TI_FRAMELEN;
 	rcb->ti_flags = 0;
-#ifdef TI_CSUM_OFFLOAD
-	rcb->ti_flags |= TI_RCB_FLAG_TCP_UDP_CKSUM|TI_RCB_FLAG_IP_CKSUM;
-#endif
+	if (sc->arpcom.ac_if.if_hwassist)
+		rcb->ti_flags |= TI_RCB_FLAG_TCP_UDP_CKSUM |
+		     TI_RCB_FLAG_IP_CKSUM | TI_RCB_FLAG_NO_PHDR_CKSUM;
 #if NVLAN > 0
 	rcb->ti_flags |= TI_RCB_FLAG_VLAN_ASSIST;
 #endif
@@ -1412,9 +1407,9 @@ static int ti_gibinit(sc)
 	    vtophys(&sc->ti_rdata->ti_rx_jumbo_ring);
 	rcb->ti_max_len = TI_JUMBO_FRAMELEN;
 	rcb->ti_flags = 0;
-#ifdef TI_CSUM_OFFLOAD
-	rcb->ti_flags |= TI_RCB_FLAG_TCP_UDP_CKSUM|TI_RCB_FLAG_IP_CKSUM;
-#endif
+	if (sc->arpcom.ac_if.if_hwassist)
+		rcb->ti_flags |= TI_RCB_FLAG_TCP_UDP_CKSUM |
+		     TI_RCB_FLAG_IP_CKSUM | TI_RCB_FLAG_NO_PHDR_CKSUM;
 #if NVLAN > 0
 	rcb->ti_flags |= TI_RCB_FLAG_VLAN_ASSIST;
 #endif
@@ -1432,9 +1427,9 @@ static int ti_gibinit(sc)
 		rcb->ti_flags = TI_RCB_FLAG_RING_DISABLED;
 	else
 		rcb->ti_flags = 0;
-#ifdef TI_CSUM_OFFLOAD
-	rcb->ti_flags |= TI_RCB_FLAG_TCP_UDP_CKSUM|TI_RCB_FLAG_IP_CKSUM;
-#endif
+	if (sc->arpcom.ac_if.if_hwassist)
+		rcb->ti_flags |= TI_RCB_FLAG_TCP_UDP_CKSUM |
+		     TI_RCB_FLAG_IP_CKSUM | TI_RCB_FLAG_NO_PHDR_CKSUM;
 #if NVLAN > 0
 	rcb->ti_flags |= TI_RCB_FLAG_VLAN_ASSIST;
 #endif
@@ -1474,6 +1469,9 @@ static int ti_gibinit(sc)
 #if NVLAN > 0
 	rcb->ti_flags |= TI_RCB_FLAG_VLAN_ASSIST;
 #endif
+	if (sc->arpcom.ac_if.if_hwassist)
+		rcb->ti_flags |= TI_RCB_FLAG_TCP_UDP_CKSUM |
+		     TI_RCB_FLAG_IP_CKSUM | TI_RCB_FLAG_NO_PHDR_CKSUM;
 	rcb->ti_max_len = TI_TX_RING_CNT;
 	if (sc->ti_hwrev == TI_HWREV_TIGON)
 		TI_HOSTADDR(rcb->ti_hostaddr) = TI_TX_RING_BASE;
@@ -1791,9 +1789,6 @@ static void ti_rxeof(sc)
 		u_int16_t		vlan_tag = 0;
 		int			have_tag = 0;
 #endif
-#ifdef TI_CSUM_OFFLOAD
-		struct ip		*ip;
-#endif
 
 		cur_rx =
 		    &sc->ti_rdata->ti_rx_return_ring[sc->ti_rx_saved_considx];
@@ -1876,12 +1871,13 @@ static void ti_rxeof(sc)
 		/* Remove header from mbuf and pass it on. */
 		m_adj(m, sizeof(struct ether_header));
 
-#ifdef TI_CSUM_OFFLOAD
-		ip = mtod(m, struct ip *);
-		if (!(cur_rx->ti_tcp_udp_cksum ^ 0xFFFF) &&
-		    !(ip->ip_off & htons(IP_MF | IP_OFFMASK | IP_RF)))
-			m->m_flags |= M_HWCKSUM;
-#endif
+		if (ifp->if_hwassist) {
+			m->m_pkthdr.csum_flags |= CSUM_IP_CHECKED |
+			    CSUM_DATA_VALID;
+			if ((cur_rx->ti_ip_cksum ^ 0xffff) == 0)
+				m->m_pkthdr.csum_flags |= CSUM_IP_VALID;
+			m->m_pkthdr.csum_data = cur_rx->ti_tcp_udp_cksum;
+		}
 
 #if NVLAN > 0
 		/*
@@ -2025,6 +2021,7 @@ static int ti_encap(sc, m_head, txidx)
 	struct ti_tx_desc	*f = NULL;
 	struct mbuf		*m;
 	u_int32_t		frag, cur, cnt = 0;
+	u_int16_t		csum_flags = 0;
 #if NVLAN > 0
 	struct ifvlan		*ifv = NULL;
 
@@ -2037,6 +2034,16 @@ static int ti_encap(sc, m_head, txidx)
 	m = m_head;
 	cur = frag = *txidx;
 
+	if (m_head->m_pkthdr.csum_flags) {
+		if (m_head->m_pkthdr.csum_flags & CSUM_IP)
+			csum_flags |= TI_BDFLAG_IP_CKSUM;
+		if (m_head->m_pkthdr.csum_flags & (CSUM_TCP | CSUM_UDP))
+			csum_flags |= TI_BDFLAG_TCP_UDP_CKSUM;
+		if (m_head->m_flags & M_LASTFRAG)
+			csum_flags |= TI_BDFLAG_IP_FRAG_END;
+		else if (m_head->m_flags & M_FRAG)
+			csum_flags |= TI_BDFLAG_IP_FRAG;
+	}
 	/*
  	 * Start packing the mbufs in this chain into
 	 * the fragment pointers. Stop when we run out
@@ -2064,7 +2071,7 @@ static int ti_encap(sc, m_head, txidx)
 				break;
 			TI_HOSTADDR(f->ti_addr) = vtophys(mtod(m, vm_offset_t));
 			f->ti_len = m->m_len;
-			f->ti_flags = 0;
+			f->ti_flags = csum_flags;
 #if NVLAN > 0
 			if (ifv != NULL) {
 				f->ti_flags |= TI_BDFLAG_VLAN_TAG;
@@ -2125,6 +2132,24 @@ static void ti_start(ifp)
 			break;
 
 		/*
+		 * XXX
+		 * safety overkill.  If this is a fragmented packet chain
+		 * with delayed TCP/UDP checksums, then only encapsulate
+		 * it if we have enough descriptors to handle the entire
+		 * chain at once.
+		 * (paranoia -- may not actually be needed)
+		 */
+		if (m_head->m_flags & M_FIRSTFRAG &&
+		    m_head->m_pkthdr.csum_flags & (CSUM_DELAY_DATA)) {
+			if ((TI_TX_RING_CNT - sc->ti_txcnt) <
+			    m_head->m_pkthdr.csum_data + 16) {
+				IF_PREPEND(&ifp->if_snd, m_head);
+				ifp->if_flags |= IFF_OACTIVE;
+				break;
+			}
+		}
+
+		/*
 		 * Pack the data into the transmit ring. If we
 		 * don't have room, set the OACTIVE flag and wait
 		 * for the NIC to drain the ring.
diff --git a/sys/sys/mbuf.h b/sys/sys/mbuf.h
index bd9bd5cd65e5..54549466962c 100644
--- a/sys/sys/mbuf.h
+++ b/sys/sys/mbuf.h
@@ -80,6 +80,9 @@ struct pkthdr {
 	int	len;			/* total packet length */
 	/* variables for ip and tcp reassembly */
 	caddr_t	header;			/* pointer to packet header */
+	/* variables for hardware checksum */
+	int	csum_flags;		/* flags regarding checksum */
+	int	csum_data;		/* data field used by csum routines */
 };
 
 /* description of external storage mapped into mbuf, valid if M_EXT set */
@@ -131,11 +134,28 @@ struct mbuf {
 #define	M_BCAST		0x0100	/* send/received as link-level broadcast */
 #define	M_MCAST		0x0200	/* send/received as link-level multicast */
 #define	M_FRAG		0x0400	/* packet is a fragment of a larger packet */
+#define	M_FIRSTFRAG	0x0800	/* packet is first fragment */
+#define	M_LASTFRAG	0x1000	/* packet is last fragment */
 
 /* flags copied when copying m_pkthdr */
 #define	M_COPYFLAGS	(M_PKTHDR|M_EOR|M_PROTO1|M_PROTO1|M_PROTO2|M_PROTO3 | \
 			    M_PROTO4|M_PROTO5|M_BCAST|M_MCAST|M_FRAG)
 
+/* flags indicating hw checksum support and sw checksum requirements */
+#define CSUM_IP			0x0001		/* will csum IP */
+#define CSUM_TCP		0x0002		/* will csum TCP */
+#define CSUM_UDP		0x0004		/* will csum UDP */
+#define CSUM_IP_FRAGS		0x0008		/* will csum IP fragments */
+#define CSUM_FRAGMENT		0x0010		/* will do IP fragmentation */
+
+#define CSUM_IP_CHECKED		0x0100		/* did csum IP */
+#define CSUM_IP_VALID		0x0200		/*   ... the csum is valid */
+#define CSUM_DATA_VALID		0x0400		/* csum_data field is valid */
+#define CSUM_PSEUDO_HDR		0x0800		/* csum_data has pseudo hdr */
+
+#define CSUM_DELAY_DATA		(CSUM_TCP | CSUM_UDP)
+#define CSUM_DELAY_IP		(CSUM_IP)	/* XXX add ipv6 here too? */
+
 /* mbuf types */
 #define	MT_FREE		0	/* should be on free list */
 #define	MT_DATA		1	/* dynamic (data) allocation */
@@ -301,6 +321,7 @@ union mcluster {
 		_mm->m_data = _mm->m_pktdat;				\
 		_mm->m_flags = M_PKTHDR;				\
 		_mm->m_pkthdr.rcvif = NULL;				\
+		_mm->m_pkthdr.csum_flags = 0;				\
 		(m) = _mm;						\
 		splx(_ms);						\
 	} else {							\