OS-6812 implement RFD 118 interpretation of HCKSUM_INET_FULL_V4/6

6 files changed, 84 insertions, 24 deletions

diff --git a/usr/src/man/man9e/mac.9e b/usr/src/man/man9e/mac.9e
index 52984f9791..c22becc131 100644
--- a/usr/src/man/man9e/mac.9e
+++ b/usr/src/man/man9e/mac.9e
@@ -9,9 +9,9 @@
 .\" http://www.illumos.org/license/CDDL.
 .\"
 .\"
-.\" Copyright 2016 Joyent, Inc.
+.\" Copyright 2018 Joyent, Inc.
 .\"
-.Dd March 26, 2017
+.Dd March 23, 2018
 .Dt MAC 9E
 .Os
 .Sh NAME
@@ -550,24 +550,28 @@ The following set of flags may be combined through a bitwise inclusive OR:
 .Bl -tag -width Ds
 .It Sy HCKSUM_INET_PARTIAL
 This indicates that the hardware can calculate a partial checksum for
-both IPv4 and IPv6; however, it requires the pseudo-header checksum be
-calculated for it.
+both IPv4 and IPv6 UDP and TCP packets; however, it requires the pseudo-header
+checksum be calculated for it.
 The pseudo-header checksum will be available for the mblk_t when calling
 .Xr mac_hcksum_get 9F .
-Note this does not imply that the hardware is capable of calculating the
-IPv4 header checksum.
+Note this does not imply that the hardware is capable of calculating
+the partial checksum for other L4 protocols or the IPv4 header checksum.
 That should be indicated with the
 .Sy HCKSUM_IPHDRCKSUM flag.
 .It Sy HCKSUM_INET_FULL_V4
-This indicates that the hardware will fully calculate the L4 checksum
-for outgoing IPv4 packets and does not require a pseudo-header checksum.
+This indicates that the hardware will fully calculate the L4 checksum for
+outgoing IPv4 UDP or TCP packets only, and does not require a pseudo-header
+checksum.
 Note this does not imply that the hardware is capable of calculating the
-IPv4 header checksum.
+checksum for other L4 protocols or the IPv4 header checksum.
 That should be indicated with the
 .Sy HCKSUM_IPHDRCKSUM .
 .It Sy HCKSUM_INET_FULL_V6
-This indicates that the hardware will fully calculate the L4 checksum
-for outgoing IPv6 packets and does not require a pseudo-header checksum.
+This indicates that the hardware will fully calculate the L4 checksum for
+outgoing IPv6 UDP or TCP packets only, and does not require a pseudo-header
+checksum.
+Note this does not imply that the hardware is capable of calculating the
+checksum for any other L4 protocols.
 .It Sy HCKSUM_IPHDRCKSUM
 This indicates that the hardware supports calculating the checksum for
 the IPv4 header itself.
diff --git a/usr/src/uts/common/inet/ip/ip6_output.c b/usr/src/uts/common/inet/ip/ip6_output.c
index b023a2fe6a..dc074454e3 100644
--- a/usr/src/uts/common/inet/ip/ip6_output.c
+++ b/usr/src/uts/common/inet/ip/ip6_output.c
@@ -23,6 +23,7 @@
  * Copyright 2010 Sun Microsystems, Inc.  All rights reserved.
  * Use is subject to license terms.
  * Copyright 2017 OmniTI Computer Consulting, Inc. All rights reserved.
+ * Copyright 2018 Joyent, Inc.
  */
 /* Copyright (c) 1990 Mentat Inc. */
 
@@ -866,8 +867,16 @@ ip_output_cksum_v6(iaflags_t ixaflags, mblk_t *mp, ip6_t *ip6h,
 		    ixa->ixa_raw_cksum_offset);
 		cksum = htons(protocol);
 	} else if (protocol == IPPROTO_ICMPV6) {
-		cksump = IPH_ICMPV6_CHECKSUMP(ip6h, ip_hdr_length);
-		cksum = IP_ICMPV6_CSUM_COMP;	/* Pseudo-header cksum */
+		/*
+		 * Currently we assume no HW support for ICMP checksum calc.
+		 *
+		 * When HW support is advertised for ICMP, we'll want the
+		 * following to be set:
+		 * cksump = IPH_ICMPV6_CHECKSUMP(ip6h, ip_hdr_length);
+		 * cksum = IP_ICMPV6_CSUM_COMP;		Pseudo-header cksum
+		 */
+
+		return (ip_output_sw_cksum_v6(mp, ip6h, ixa));
 	} else {
 	ip_hdr_cksum:
 		/* No IP header checksum for IPv6 */
diff --git a/usr/src/uts/common/inet/ip/ip_output.c b/usr/src/uts/common/inet/ip/ip_output.c
index 5caa043a35..02a59e69c5 100644
--- a/usr/src/uts/common/inet/ip/ip_output.c
+++ b/usr/src/uts/common/inet/ip/ip_output.c
@@ -21,6 +21,7 @@
 
 /*
  * Copyright (c) 2009, 2010, Oracle and/or its affiliates. All rights reserved.
+ * Copyright 2018 Joyent, Inc.
  */
 /* Copyright (c) 1990 Mentat Inc. */
 
@@ -1737,6 +1738,13 @@ ip_output_cksum_v4(iaflags_t ixaflags, mblk_t *mp, ipha_t *ipha,
 #endif
 			sctph->sh_chksum = sctp_cksum(mp, ip_hdr_length);
 		goto ip_hdr_cksum;
+	} else if (protocol == IPPROTO_ICMP) {
+		/*
+		 * Note that we always calculate a SW checksum for ICMP. In the
+		 * future, if HW support for ICMP is advertised, we can change
+		 * this.
+		 */
+		return (ip_output_sw_cksum_v4(mp, ipha, ixa));
 	} else {
 	ip_hdr_cksum:
 		/* Calculate IPv4 header checksum */
diff --git a/usr/src/uts/common/io/qede/qede_gld.c b/usr/src/uts/common/io/qede/qede_gld.c
index 50a4013555..4098409c17 100644
--- a/usr/src/uts/common/io/qede/qede_gld.c
+++ b/usr/src/uts/common/io/qede/qede_gld.c
@@ -33,6 +33,9 @@
 * limitations under the License.
 */
 
+/*
+ * Copyright 2018 Joyent, Inc.
+ */
 
 #include "qede.h"
 
@@ -2009,12 +2012,10 @@ qede_mac_get_capability(void *arg,
 		}
 
 		/*
-		 * Hardware does not support ICMPv6 checksumming. Right now the
-		 * GLDv3 doesn't provide us a way to specify that we don't
-		 * support that. As such, we cannot indicate
-		 * HCKSUM_INET_FULL_V6.
+		 * Hardware does not support ICMPv6 checksumming, but
+		 * HCKSUM_INET_FULL_V4/6 only applies for UDP and TCP.
 		 */
-		*tx_flags = HCKSUM_INET_FULL_V4 |
+		*tx_flags = HCKSUM_INET_FULL_V4 | HCKSUM_INET_FULL_V6 |
 		    HCKSUM_IPHDRCKSUM;
 		ret = B_TRUE;
 		break;
diff --git a/usr/src/uts/common/sys/dlpi.h b/usr/src/uts/common/sys/dlpi.h
index 34f1c17236..d76daffeb7 100644
--- a/usr/src/uts/common/sys/dlpi.h
+++ b/usr/src/uts/common/sys/dlpi.h
@@ -20,6 +20,7 @@
  */
 /*
  * Copyright (c) 1989, 2010, Oracle and/or its affiliates. All rights reserved.
+ * Copyright 2018 Joyent, Inc.
  */
 
 /*	Copyright (c) 1984, 1986, 1987, 1988, 1989 AT&T	*/
@@ -676,11 +677,11 @@ typedef struct {
 #define	HCKSUM_ENABLE		0x01	/* Set to enable hardware checksum */
 					/* capability */
 #define	HCKSUM_INET_PARTIAL	0x02	/* Partial 1's complement checksum */
-					/* ability */
+					/* ability for TCP/UDP packets. */
 #define	HCKSUM_INET_FULL_V4	0x04	/* Full 1's complement checksum */
-					/* ability for IPv4 packets. */
+					/* ability for IPv4 TCP/UDP packets. */
 #define	HCKSUM_INET_FULL_V6	0x08	/* Full 1's complement checksum */
-					/* ability for IPv6 packets. */
+					/* ability for IPv6 TCP/UDP packets. */
 #define	HCKSUM_IPHDRCKSUM	0x10	/* IPv4 Header checksum offload */
 					/* capability */
 #ifdef _KERNEL
diff --git a/usr/src/uts/i86pc/io/viona/viona.c b/usr/src/uts/i86pc/io/viona/viona.c
index 768a85f500..427a45cfad 100644
--- a/usr/src/uts/i86pc/io/viona/viona.c
+++ b/usr/src/uts/i86pc/io/viona/viona.c
@@ -56,6 +56,7 @@
 #include <sys/mac_provider.h>
 #include <sys/mac_client_priv.h>
 #include <sys/vlan.h>
+#include <inet/ip.h>
 
 #include <sys/vmm_drv.h>
 #include <sys/viona_io.h>
@@ -1934,6 +1935,28 @@ viona_desb_release(viona_desb_t *dp)
 	mutex_exit(&ring->vr_lock);
 }
 
+static int
+viona_mb_get_uint8(mblk_t *mp, off_t off, uint8_t *out)
+{
+	size_t mpsize;
+	uint8_t *bp;
+
+	mpsize = msgsize(mp);
+	if (off + sizeof (uint8_t) > mpsize)
+		return (-1);
+
+	mpsize = MBLKL(mp);
+	while (off >= mpsize) {
+		mp = mp->b_cont;
+		off -= mpsize;
+		mpsize = MBLKL(mp);
+	}
+
+	bp = mp->b_rptr + off;
+	*out = *bp;
+	return (0);
+}
+
 static boolean_t
 viona_tx_csum(viona_vring_t *ring, const struct virtio_net_hdr *hdr,
     mblk_t *mp, uint32_t len)
@@ -1942,6 +1965,7 @@ viona_tx_csum(viona_vring_t *ring, const struct virtio_net_hdr *hdr,
 	const struct ether_header *eth;
 	uint_t eth_len = sizeof (struct ether_header);
 	ushort_t ftype;
+	uint8_t ipproto = IPPROTO_NONE; /* NONE is not exactly right, but ok */
 
 	eth = (const struct ether_header *)mp->b_rptr;
 	if (MBLKL(mp) < sizeof (*eth)) {
@@ -1959,11 +1983,22 @@ viona_tx_csum(viona_vring_t *ring, const struct virtio_net_hdr *hdr,
 		ftype = ntohs(veth->ether_type);
 	}
 
+	if (ftype == ETHERTYPE_IP) {
+		const size_t off = offsetof(ipha_t, ipha_protocol) + eth_len;
+
+		(void) viona_mb_get_uint8(mp, off, &ipproto);
+	} else if (ftype == ETHERTYPE_IPV6) {
+		const size_t off = offsetof(ip6_t, ip6_nxt) + eth_len;
+
+		(void) viona_mb_get_uint8(mp, off, &ipproto);
+	}
+
 	/*
 	 * Partial checksum support from the NIC is ideal, since it most
 	 * closely maps to the interface defined by virtio.
 	 */
-	if ((link->l_cap_csum & HCKSUM_INET_PARTIAL) != 0) {
+	if ((link->l_cap_csum & HCKSUM_INET_PARTIAL) != 0 &&
+	    (ipproto == IPPROTO_TCP || ipproto == IPPROTO_UDP)) {
 		uint_t start, stuff, end;
 
 		/*
@@ -1984,7 +2019,8 @@ viona_tx_csum(viona_vring_t *ring, const struct virtio_net_hdr *hdr,
 	 * checksum will need to calculated inline.
 	 */
 	if (ftype == ETHERTYPE_IP) {
-		if ((link->l_cap_csum & HCKSUM_INET_FULL_V4) != 0) {
+		if ((link->l_cap_csum & HCKSUM_INET_FULL_V4) != 0 &&
+		    (ipproto == IPPROTO_TCP || ipproto == IPPROTO_UDP)) {
 			mac_hcksum_set(mp, 0, 0, 0, 0, HCK_FULLCKSUM);
 			return (B_TRUE);
 		}
@@ -1994,7 +2030,8 @@ viona_tx_csum(viona_vring_t *ring, const struct virtio_net_hdr *hdr,
 		VIONA_RING_STAT_INCR(ring, fail_hcksum);
 		return (B_FALSE);
 	} else if (ftype == ETHERTYPE_IPV6) {
-		if ((link->l_cap_csum & HCKSUM_INET_FULL_V6) != 0) {
+		if ((link->l_cap_csum & HCKSUM_INET_FULL_V6) != 0 &&
+		    (ipproto == IPPROTO_TCP || ipproto == IPPROTO_UDP)) {
 			mac_hcksum_set(mp, 0, 0, 0, 0, HCK_FULLCKSUM);
 			return (B_TRUE);
 		}