On Thu, 2017-03-30 at 16:36 -0700, Cong Wang wrote:

On Tue, Mar 28, 2017 at 5:52 PM, Eric Dumazet [off-list ref] wrote:

quoted

On Tue, 2017-03-28 at 16:11 -0700, Eric Dumazet wrote:

quoted

Yes, this looks better.

Although you probably need to change a bit later this part :

if (!inet->inet_saddr)
      inet->inet_saddr = fl4->saddr;  /* Update source address */

I came up with the following tested patch for IPv4

diff --git a/net/ipv4/datagram.c b/net/ipv4/datagram.c
index f915abff1350a86af8d5bb89725b751c061b0fb5..1454b6191e0d38ffae0ae260578858285bc5f77b 100644
--- a/net/ipv4/datagram.c
+++ b/net/ipv4/datagram.c

@@ -40,7 +40,7 @@ int __ip4_datagram_connect(struct sock *sk, struct sockaddr *uaddr, int addr_len
        sk_dst_reset(sk);

        oif = sk->sk_bound_dev_if;
-       saddr = inet->inet_saddr;
+       saddr = (sk->sk_userlocks & SOCK_BINDADDR_LOCK) ? inet->inet_saddr : 0;
        if (ipv4_is_multicast(usin->sin_addr.s_addr)) {
                if (!oif)
                        oif = inet->mc_index;

@@ -64,9 +64,8 @@ int __ip4_datagram_connect(struct sock *sk, struct sockaddr *uaddr, int addr_len
                err = -EACCES;
                goto out;
        }
-       if (!inet->inet_saddr)
+       if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK)) {
                inet->inet_saddr = fl4->saddr;  /* Update source address */
-       if (!inet->inet_rcv_saddr) {
                inet->inet_rcv_saddr = fl4->saddr;
                if (sk->sk_prot->rehash)
                        sk->sk_prot->rehash(sk);

Why do we need this here? If you mean bind() INADDR_ANY is bound,
then it is totally a different problem?

Proper delivery of RX packets will need to find the socket, and this
needs the 2-tuple (source address, source port) info for UDP.

So after a connect(), we need to rehash

BTW, I am still not sure about what POSIX says about the connect()
behavior, I can only find this [1]:

"
If the initiating socket is not connection-mode, then connect() shall set the
socket's peer address, and no connection is made. For SOCK_DGRAM
sockets, the peer address identifies where all datagrams are sent on
subsequent send() functions, and limits the remote sender for subsequent
recv() functions.
"

It doesn't say anything about source address. But the man page [2] says:

"
When
       connect(2) is called on an unbound socket, the socket is
       automatically bound to a random free port or to a usable shared port
       with the local address set to INADDR_ANY.
"

Seems the last part is inaccurate, kernel actually picks a source address
from route instead of just using INADDR_ANY for connect(2).

So, for me, I think the following behaviors make sense for UDP:

1) When a bind() is called before connect()'s, aka:

bind();
connect(addr1); // should not change source addr

It depends. bind() can be only allocating the source port.

If bind(INADDR_ANY) was used, then we need to determine source addr at
connect() time.

Point of connect() is that future send() wont have to guess the 4-tuple
infos. But also that incoming packets will find this precise socket
thanks to a higher score.

And tools like "ss -aun" should display the 4-tuple after a successful
connect()

connect(addr2); // should fail is the source addr can not reach peer addr

2) No bind() before connect()'s, aka:

connect(addr1); // Free to bind a source addr
connect(addr2); // Free to bind a new source addr and change peer addr

Exactly. My patch does this.

Thoughts?

1. http://pubs.opengroup.org/onlinepubs/9699919799/functions/connect.html
2. http://man7.org/linux/man-pages/man7/ip.7.html