udp: generate gso with UDP_SEGMENT

Support generic segmentation offload for udp datagrams. Callers can
concatenate and send at once the payload of multiple datagrams with
the same destination.

To set segment size, the caller sets socket option UDP_SEGMENT to the
length of each discrete payload. This value must be smaller than or
equal to the relevant MTU.

A follow-up patch adds cmsg UDP_SEGMENT to specify segment size on a
per send call basis.

Total byte length may then exceed MTU. If not an exact multiple of
segment size, the last segment will be shorter.

The implementation adds a gso_size field to the udp socket, ip(v6)
cmsg cookie and inet_cork structure to be able to set the value at
setsockopt or cmsg time and to work with both lockless and corked
paths.

Initial benchmark numbers show UDP GSO about as expensive as TCP GSO.

    tcp tso
     3197 MB/s 54232 msg/s 54232 calls/s
         6,457,754,262      cycles

    tcp gso
     1765 MB/s 29939 msg/s 29939 calls/s
        11,203,021,806      cycles

    tcp without tso/gso *
      739 MB/s 12548 msg/s 12548 calls/s
        11,205,483,630      cycles

    udp
      876 MB/s 14873 msg/s 624666 calls/s
        11,205,777,429      cycles

    udp gso
     2139 MB/s 36282 msg/s 36282 calls/s
        11,204,374,561      cycles

   [*] after reverting commit 0a6b2a1d
       ("tcp: switch to GSO being always on")

Measured total system cycles ('-a') for one core while pinning both
the network receive path and benchmark process to that core:

  perf stat -a -C 12 -e cycles \
    ./udpgso_bench_tx -C 12 -4 -D "$DST" -l 4

Note the reduction in calls/s with GSO. Bytes per syscall drops
increases from 1470 to 61818.
Signed-off-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

include/linux/udp.h

@@ -55,6 +55,7 @@ struct udp_sock {
 	 * when the socket is uncorked.
 	 */
 	__u16		 len;		/* total length of pending frames */
+	__u16		 gso_size;
 	/*
 	 * Fields specific to UDP-Lite.
 	 */
@@ -87,6 +88,8 @@ struct udp_sock {
 	int		forward_deficit;
 };
 
+#define UDP_MAX_SEGMENTS	(1 << 6UL)
+
 static inline struct udp_sock *udp_sk(const struct sock *sk)
 {
 	return (struct udp_sock *)sk;

include/net/inet_sock.h

@@ -147,6 +147,7 @@ struct inet_cork {
 	__u8			ttl;
 	__s16			tos;
 	char			priority;
+	__u16			gso_size;
 };
 
 struct inet_cork_full {

include/net/ip.h

@@ -76,6 +76,7 @@ struct ipcm_cookie {
 	__u8			ttl;
 	__s16			tos;
 	char			priority;
+	__u16			gso_size;
 };
 
 #define IPCB(skb) ((struct inet_skb_parm*)((skb)->cb))

include/net/ipv6.h

@@ -298,6 +298,7 @@ struct ipcm6_cookie {
 	__s16 tclass;
 	__s8  dontfrag;
 	struct ipv6_txoptions *opt;
+	__u16 gso_size;
 };
 
 static inline struct ipv6_txoptions *txopt_get(const struct ipv6_pinfo *np)

include/uapi/linux/udp.h

@@ -32,6 +32,7 @@ struct udphdr {
 #define UDP_ENCAP	100	/* Set the socket to accept encapsulated packets */
 #define UDP_NO_CHECK6_TX 101	/* Disable sending checksum for UDP6X */
 #define UDP_NO_CHECK6_RX 102	/* Disable accpeting checksum for UDP6 */
+#define UDP_SEGMENT	103	/* Set GSO segmentation size */
 
 /* UDP encapsulation types */
 #define UDP_ENCAP_ESPINUDP_NON_IKE	1 /* draft-ietf-ipsec-nat-t-ike-00/01 */

net/ipv4/ip_output.c

@@ -882,7 +882,8 @@ static int __ip_append_data(struct sock *sk,
 	skb = skb_peek_tail(queue);
 
 	exthdrlen = !skb ? rt->dst.header_len : 0;
-	mtu = cork->fragsize;
+	mtu = cork->gso_size ? IP_MAX_MTU : cork->fragsize;
+
 	if (cork->tx_flags & SKBTX_ANY_SW_TSTAMP &&
 	    sk->sk_tsflags & SOF_TIMESTAMPING_OPT_ID)
 		tskey = sk->sk_tskey++;
@@ -906,7 +907,7 @@ static int __ip_append_data(struct sock *sk,
 	if (transhdrlen &&
 	    length + fragheaderlen <= mtu &&
 	    rt->dst.dev->features & (NETIF_F_HW_CSUM | NETIF_F_IP_CSUM) &&
-	    !(flags & MSG_MORE) &&
+	    (!(flags & MSG_MORE) || cork->gso_size) &&
 	    !exthdrlen)
 		csummode = CHECKSUM_PARTIAL;
 
@@ -1135,6 +1136,8 @@ static int ip_setup_cork(struct sock *sk, struct inet_cork *cork,
 	*rtp = NULL;
 	cork->fragsize = ip_sk_use_pmtu(sk) ?
 			 dst_mtu(&rt->dst) : rt->dst.dev->mtu;
+
+	cork->gso_size = sk->sk_type == SOCK_DGRAM ? ipc->gso_size : 0;
 	cork->dst = &rt->dst;
 	cork->length = 0;
 	cork->ttl = ipc->ttl;
@@ -1214,7 +1217,7 @@ ssize_t	ip_append_page(struct sock *sk, struct flowi4 *fl4, struct page *page,
 		return -EOPNOTSUPP;
 
 	hh_len = LL_RESERVED_SPACE(rt->dst.dev);
-	mtu = cork->fragsize;
+	mtu = cork->gso_size ? IP_MAX_MTU : cork->fragsize;
 
 	fragheaderlen = sizeof(struct iphdr) + (opt ? opt->optlen : 0);
 	maxfraglen = ((mtu - fragheaderlen) & ~7) + fragheaderlen;

net/ipv4/udp.c

@@ -757,7 +757,8 @@ void udp_set_csum(bool nocheck, struct sk_buff *skb,
 }
 EXPORT_SYMBOL(udp_set_csum);
 
-static int udp_send_skb(struct sk_buff *skb, struct flowi4 *fl4)
+static int udp_send_skb(struct sk_buff *skb, struct flowi4 *fl4,
+			struct inet_cork *cork)
 {
 	struct sock *sk = skb->sk;
 	struct inet_sock *inet = inet_sk(sk);
@@ -777,6 +778,21 @@ static int udp_send_skb(struct sk_buff *skb, struct flowi4 *fl4)
 	uh->len = htons(len);
 	uh->check = 0;
 
+	if (cork->gso_size) {
+		const int hlen = skb_network_header_len(skb) +
+				 sizeof(struct udphdr);
+
+		if (hlen + cork->gso_size > cork->fragsize)
+			return -EINVAL;
+		if (skb->len > cork->gso_size * UDP_MAX_SEGMENTS)
+			return -EINVAL;
+		if (skb->ip_summed != CHECKSUM_PARTIAL || is_udplite)
+			return -EIO;
+
+		skb_shinfo(skb)->gso_size = cork->gso_size;
+		skb_shinfo(skb)->gso_type = SKB_GSO_UDP_L4;
+	}
+
 	if (is_udplite)  				 /*     UDP-Lite      */
 		csum = udplite_csum(skb);
 
@@ -828,7 +844,7 @@ int udp_push_pending_frames(struct sock *sk)
 	if (!skb)
 		goto out;
 
-	err = udp_send_skb(skb, fl4);
+	err = udp_send_skb(skb, fl4, &inet->cork.base);
 
 out:
 	up->len = 0;
@@ -922,6 +938,7 @@ int udp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len)
 	ipc.sockc.tsflags = sk->sk_tsflags;
 	ipc.addr = inet->inet_saddr;
 	ipc.oif = sk->sk_bound_dev_if;
+	ipc.gso_size = up->gso_size;
 
 	if (msg->msg_controllen) {
 		err = ip_cmsg_send(sk, msg, &ipc, sk->sk_family == AF_INET6);
@@ -1037,7 +1054,7 @@ int udp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len)
 				  &cork, msg->msg_flags);
 		err = PTR_ERR(skb);
 		if (!IS_ERR_OR_NULL(skb))
-			err = udp_send_skb(skb, fl4);
+			err = udp_send_skb(skb, fl4, &cork);
 		goto out;
 	}
 
@@ -2367,6 +2384,12 @@ int udp_lib_setsockopt(struct sock *sk, int level, int optname,
 		up->no_check6_rx = valbool;
 		break;
 
+	case UDP_SEGMENT:
+		if (val < 0 || val > USHRT_MAX)
+			return -EINVAL;
+		up->gso_size = val;
+		break;
+
 	/*
 	 * 	UDP-Lite's partial checksum coverage (RFC 3828).
 	 */
@@ -2457,6 +2480,10 @@ int udp_lib_getsockopt(struct sock *sk, int level, int optname,
 		val = up->no_check6_rx;
 		break;
 
+	case UDP_SEGMENT:
+		val = up->gso_size;
+		break;
+
 	/* The following two cannot be changed on UDP sockets, the return is
 	 * always 0 (which corresponds to the full checksum coverage of UDP). */
 	case UDPLITE_SEND_CSCOV:

net/ipv6/ip6_output.c

@@ -1240,6 +1240,8 @@ static int ip6_setup_cork(struct sock *sk, struct inet_cork_full *cork,
 	if (mtu < IPV6_MIN_MTU)
 		return -EINVAL;
 	cork->base.fragsize = mtu;
+	cork->base.gso_size = sk->sk_type == SOCK_DGRAM ? ipc6->gso_size : 0;
+
 	if (dst_allfrag(xfrm_dst_path(&rt->dst)))
 		cork->base.flags |= IPCORK_ALLFRAG;
 	cork->base.length = 0;
@@ -1281,7 +1283,7 @@ static int __ip6_append_data(struct sock *sk,
 		dst_exthdrlen = rt->dst.header_len - rt->rt6i_nfheader_len;
 	}
 
-	mtu = cork->fragsize;
+	mtu = cork->gso_size ? IP6_MAX_MTU : cork->fragsize;
 	orig_mtu = mtu;
 
 	hh_len = LL_RESERVED_SPACE(rt->dst.dev);
@@ -1329,7 +1331,7 @@ static int __ip6_append_data(struct sock *sk,
 	if (transhdrlen && sk->sk_protocol == IPPROTO_UDP &&
 	    headersize == sizeof(struct ipv6hdr) &&
 	    length <= mtu - headersize &&
-	    !(flags & MSG_MORE) &&
+	    (!(flags & MSG_MORE) || cork->gso_size) &&
 	    rt->dst.dev->features & (NETIF_F_IPV6_CSUM | NETIF_F_HW_CSUM))
 		csummode = CHECKSUM_PARTIAL;
 

net/ipv6/udp.c

@@ -1023,7 +1023,8 @@ static void udp6_hwcsum_outgoing(struct sock *sk, struct sk_buff *skb,
  *	Sending
  */
 
-static int udp_v6_send_skb(struct sk_buff *skb, struct flowi6 *fl6)
+static int udp_v6_send_skb(struct sk_buff *skb, struct flowi6 *fl6,
+			   struct inet_cork *cork)
 {
 	struct sock *sk = skb->sk;
 	struct udphdr *uh;
@@ -1042,6 +1043,21 @@ static int udp_v6_send_skb(struct sk_buff *skb, struct flowi6 *fl6)
 	uh->len = htons(len);
 	uh->check = 0;
 
+	if (cork->gso_size) {
+		const int hlen = skb_network_header_len(skb) +
+				 sizeof(struct udphdr);
+
+		if (hlen + cork->gso_size > cork->fragsize)
+			return -EINVAL;
+		if (skb->len > cork->gso_size * UDP_MAX_SEGMENTS)
+			return -EINVAL;
+		if (skb->ip_summed != CHECKSUM_PARTIAL || is_udplite)
+			return -EIO;
+
+		skb_shinfo(skb)->gso_size = cork->gso_size;
+		skb_shinfo(skb)->gso_type = SKB_GSO_UDP_L4;
+	}
+
 	if (is_udplite)
 		csum = udplite_csum(skb);
 	else if (udp_sk(sk)->no_check6_tx) {   /* UDP csum disabled */
@@ -1093,7 +1109,7 @@ static int udp_v6_push_pending_frames(struct sock *sk)
 	if (!skb)
 		goto out;
 
-	err = udp_v6_send_skb(skb, &fl6);
+	err = udp_v6_send_skb(skb, &fl6, &inet_sk(sk)->cork.base);
 
 out:
 	up->len = 0;
@@ -1127,6 +1143,7 @@ int udpv6_sendmsg(struct sock *sk, struct msghdr *msg, size_t len)
 	ipc6.hlimit = -1;
 	ipc6.tclass = -1;
 	ipc6.dontfrag = -1;
+	ipc6.gso_size = up->gso_size;
 	sockc.tsflags = sk->sk_tsflags;
 
 	/* destination address check */
@@ -1333,7 +1350,7 @@ int udpv6_sendmsg(struct sock *sk, struct msghdr *msg, size_t len)
 				   msg->msg_flags, &cork, &sockc);
 		err = PTR_ERR(skb);
 		if (!IS_ERR_OR_NULL(skb))
-			err = udp_v6_send_skb(skb, &fl6);
+			err = udp_v6_send_skb(skb, &fl6, &cork.base);
 		goto out;
 	}