The final routing for ipv4 packets may be done with the IP address from the message header not that from the address buffer. If the addresses are different FLOWI_FLAG_KNOWN_NH must be set so that a temporary 'struct rtable' entry is created to send the message. However the allocate + free (under RCU) is relatively expensive and can be avoided by a quick check shows the addresses match.
Signed-off-by: David Laight <david.lai...@aculab.com> --- This makes a considerable difference when we are sending a lot of RTP streams from a raw socket. IP_HDRINCL has to be set so that the calculated UDP checksum is right. net/ipv4/raw.c | 30 ++++++++++++++++++++++++++++-- 1 file changed, 28 insertions(+), 2 deletions(-) diff --git a/net/ipv4/raw.c b/net/ipv4/raw.c index 3183413..0a81376 100644 --- a/net/ipv4/raw.c +++ b/net/ipv4/raw.c @@ -495,6 +495,27 @@ static int raw_getfrag(void *from, char *to, int offset, int len, int odd, return ip_generic_getfrag(rfv->msg, to, offset, len, odd, skb); } +static bool raw_msg_addr_matches(struct msghdr *msg, __be32 daddr) +{ + const struct iovec *iov; + __be32 msg_daddr; + + /* Check common case of user buffer with header in the first fragment. + * If we return false the message is still sent. + */ + + if (!iter_is_iovec(&msg->msg_iter)) + return false; + iov = msg->msg_iter.iov; + if (!iov || iov->iov_len < 20) + return false; + + if (get_user(msg_daddr, (__be32 __user *)(iov->iov_base + 16))) + return false; + + return daddr == msg_daddr; +} + static int raw_sendmsg(struct sock *sk, struct msghdr *msg, size_t len) { struct inet_sock *inet = inet_sk(sk); @@ -626,9 +647,14 @@ static int raw_sendmsg(struct sock *sk, struct msghdr *msg, size_t len) flowi4_init_output(&fl4, ipc.oif, ipc.sockc.mark, tos, RT_SCOPE_UNIVERSE, hdrincl ? IPPROTO_RAW : sk->sk_protocol, - inet_sk_flowi_flags(sk) | - (hdrincl ? FLOWI_FLAG_KNOWN_NH : 0), + inet_sk_flowi_flags(sk), daddr, saddr, 0, 0, sk->sk_uid); + /* The final message routing may be done with the destination address + * in the user-supplied ipv4 header. If this differs from 'daddr' then + * a temporary destination table entry has to be created. + */ + if (hdrincl && !raw_msg_addr_matches(msg, daddr)) + fl4.flowi4_flags |= FLOWI_FLAG_KNOWN_NH; if (!hdrincl) { rfv.msg = msg; -- 1.8.1.2 - Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK Registration No: 1397386 (Wales)