tcp: Set IPV6_V6ONLY for IPv6 listeners. #1555
Conversation
The current behaviour of listening on an IPv6 address varies depending on the operating system's IP address stack implementation. Some support IPv4-mapped IPv6 addresses (e.g. Linux and newer versions of Windows) so a single IPv6 address would support IPv4-mapped addresses too. Others do not (e.g. OpenBSD). If they do, then some support them by default (e.g. Linux) and some do not (e.g. Windows). This PR attempts to implement the same behaviour accross operating systems. The strategy is as follows: Disable IPv4-mapped IPv6 addresses, hence the socket option IPV6_V6ONLY is always set to true. This allows binding two sockets to the same port and also avoids the problem of comparing mixed addresses which leads issues such as libp2p#1552.
| socket.set_reuse_address(true)?; | ||
| } | ||
| socket.bind(&socket_addr.into())?; | ||
| socket.listen(1024)?; // we may want to make this configurable |
There was a problem hiding this comment.
The default in Rust seems to be 128
https://github.com/rust-lang/rust/blob/c06e4aca19046b07d952b16e9f002bfab38fde6b/src/libstd/sys/unix/ext/net.rs#L823
| socket.listen(1024)?; // we may want to make this configurable | |
| socket.listen(128)?; // we may want to make this configurable |
There was a problem hiding this comment.
I was following MIOs lead here as it is what tokio and async-std use.
| s | ||
| }; | ||
| if cfg!(target_family = "unix") { | ||
| socket.set_reuse_address(true)?; |
There was a problem hiding this comment.
I have no idea of the implications of that change. What are the reasons for that?
There was a problem hiding this comment.
I do not know why it is restricted to Unix, but again this is what MIO does which in turn does what libstd does.
| s | ||
| }; | ||
| if cfg!(target_family = "unix") { | ||
| socket.set_reuse_address(true)?; |
There was a problem hiding this comment.
Does that imply that on non-unix platforms one can not use dualstack nor two listeners (ipv4 and ipv6 on the same port)? (If so, not sure whether it would be important.)
There was a problem hiding this comment.
It depends on the particular platform. On Windows for example IPV6_V6ONLY seems to be true by default and use of SO_REUSEADDR is discouraged if I read this correctly, but frankly I do not know Windows and if someone has ideas how to configure things more optimally on this and other non-unix platforms we can make those changes in subsequent PRs.
Co-Authored-By: Pierre Krieger <pierre.krieger1708@gmail.com>
The current behaviour of listening on an IPv6 address varies depending on the operating system's IP address stack implementation. Some support IPv4-mapped IPv6 addresses (e.g. Linux and newer versions of Windows) so a single IPv6 address would support IPv4-mapped addresses too. Others do not (e.g. OpenBSD). If they do, then some support them by default (e.g. Linux) and some do not (e.g. Windows).
This PR attempts to implement the same behaviour accross operating systems. The strategy is as follows:
Disable IPv4-mapped IPv6 addresses, hence the socket option IPV6_V6ONLY is always set to true.
This allows binding two sockets to the same port and also avoids the problem of comparing mixed addresses which leads to issues such as #1552.