rfc: create sockets in uv_*_init() functions

[bnoordhuis]

Create sockets immediately in uv_tcp_init(), uv_udp_init(), etc. Don't defer until connect() or bind() time.

Benefits:

1. Less complex, don't have to remember socket options[1].

Drawbacks:

1. Changes API considerably. Caller now needs to specify the socket type upfront (e.g. IPv4 vs. IPv6). Then again, that means we can remove the uv_tcp_bind() / uv_tcp_bind4() dichotomy.
2. Pushes burden of async error reporting in case of e.g. EMFILE to user.

I'm tentatively scheduling this for v0.8.

/cc @piscisaureus

[1] Currently broken in various ways. uv_tcp_keepalive() doesn't remember the delay, uv_udp_set_ttl() doesn't work at all, etc.

[piscisaureus]

@bnoordhuis

I think that would be better. Remarks:

• This would just move the responsibility for deferred socket creation to node.
• I don't understand your comment about EMFILE.
• I'd like other to weigh in as well. (@ry, if you don't mind...)

[saghul]

This crossed my mind during the weekend, so I'll just dump some thoughts here, for whomever wants to read them :-)

• I also think this is a good idea
• Specifying (and saving) the family should be enough. Acceptable values: AF_INET, AF_INET6 and AF_UNSPEC (read on)
• With the current approach, when uv_accept is called the fd has already been accepted, so we shouln't initialize the socket when initializing the client handle. AF_UNSPEC could be used in that case, which wouldn't bind the socket in init.
• The current behavior could be kept if AF_UNSPEC is passed to init.
• Open cannot be used if the socket was already created, so AF_UNSPEC would need to be used in init.
• How do we set the family to the right value if open is used? Maybe leaving it to unespecified it ok?

[indutny]

+1, transitional AF_UNSPEC sounds fine.

[saghul]

Actually, it's not really transitional. We need to avoid binding the socket whendoping uv_accept. Do you have any other ideas on how to handle that case?

[indutny]

Why should we avoid doing it?

[saghul]

Well, imagine we always create the socket in uv_tcp_init. When you initialize the handle for uv_accept()-ing a connection, it would be a wasted effort because now you need to close it and replace it with what accept() returned.

[indutny]

Oh, I see... yeah, this is a bit of problem, indeed. Perhaps we could add API to queue handles ahead of time, but I'm not sure if the resulting code would be much better than the current one.