When we receive a TCP connection, we get the peer address from the accept()
call. In the case of a listening socket with an unspecified address (:: or
0.0.0.0) the local address of the accept()ed socket could vary. We don't
get that from the accept() - we must explicitly call getsockname() to get
it.
Currently we avoid the latency of that extra syscall, and therefore don't
populate the initiating 'oaddr' field of a flow created by an incoming TCP
socket connection. This more or less works, because we rarely need that
local address, but it does cause some oddities:
* For migration we need the local address to recreate the socket on the
destination, so we *do* call getsockname() in vhost-user mode
* It limits our options in terms of forwarding connections flexibly based
on the address to which they're received
* It differs from UDP, where we explicitly use the IP_PKTINFO cmsg to
populate oaddr.
* It means (some) flow debug messages will contain wildcards instead of
real local addresses
In theory we can elide this call when accept()ing from a socket bound to
a specific address instead of a wildcard. However to do that will need
revisions to the data structures we use to keep track of listening sockets.
The lack of this information is making it hard to implement some fixes we
want. So, pay the price of the extra syscall to get this information, with
the hope that we can later optimise it away for some cases.
Signed-off-by: David Gibson