[Feature]: Support for SO_BINDTODEVICE?

[shaunco]

Description of new feature

While listening for UDP packets, it is common to listen on "0.0.0.0:123" or just ":123", which allows for the receipt of both unicast and broadcast UDP packets, however there are times where an application might want to listen on just a single network interface. Obtaining that interface's IP and then listening on "1.2.3.4:123" will only receive unicast packets, and thus the preferred mechanism is to still listen on ":123" (for v4+v6) or "0.0.0.0:123" (for v4), but to use SO_BINDTODEVICE to bind the socket to a single interface.

Once bound, tools like ss show it as 0.0.0.0%eth0:123 rather than just 0.0.0.0:123.

Scenarios for new feature

On network appliances with multiple interfaces, those interfaces are often attached to different networks - such as a network gateway that is attached to a LAN network on eth0 and the WAN network on eth1. If that device wants to listen for unicast+broadcast UDP on just the LAN interface, there is currently no way to do this with gnet (at least not that I could find).

Breaking changes or not?

No

Code snippets (optional)

For most linux/freebsd based systems:

unix.SetsockoptString(fd, unix.SOL_SOCKET, unix.SO_BINDTODEVICE, a.interfaceName)

For darwin:

iface, _ := net.InterfaceByName(devName)
unix.SetsockoptInt(int(fd), unix.IPPROTO_IP, unix.IP_RECVIF, iface.Index)
unix.SetsockoptInt(int(fd), unix.IPPROTO_IP, unix.IP_BOUND_IF, iface.Index)

For windows, there is example calls at https://stackoverflow.com/a/73041705 , but they'd need to be converted to go.

Alternatives for new feature

Could have separate listeners for unicast and broadcast traffic, but it gets messy.

Could listen to everything and then filter in OnTraffic, but it is much more efficient to have the kernel just bind to a single interface.

Additional context (optional)

WithMulticastInterfaceIndex() does something similar for multicast traffic.

[panjf2000]

Sounds reasonable, I'll see what we can do here.

[panjf2000]

I've done some investigation on this. SO_BINDTODEVICE is linux-specific and it seems that there are no equivalents on other platforms. IP_BOUND_IF seems like a prospective substitute on macOS but it's osx-specific and it doesn't exist on other BSD-like systems.

As for IP_RECVIF on freebsd, it doesn't look like an exact equivalent to SO_BINDTODEVICE according to its man pages.

Therefore, if we're going to do this, it looks like we're only able to do it on linux and mac.

References

• SO_BINDTODEVICE undeclared on freebsd 12
• IP(4) for freebsd
• OS X equivalent of SO_BINDTODEVICE
• Use Linux SO_BINDTODEVICE and mac IP_BOUND_IF to bind socket to a network interface
• socket bind vs SO_BINDTODEVICE

[shaunco]

Thanks for digging into this. Would that make the functions just NOPs on other platforms or does it imply platform-specific APIs?

[panjf2000]

Thanks for digging into this. Would that make the functions just NOPs on other platforms or does it imply platform-specific APIs?

If we're going to do this, API should be present on all platforms that gnet supports and returns an ErrUnsupportedOp.

[github-actions]

This issue is marked as stale because it has been open for 30 days with no activity.

You should take one of the following actions:

• Manually close this issue if it is no longer relevant
• Comment if you have more information to share

This issue will be automatically closed in 7 days if no further activity occurs.

[shaunco]

No more information to share, but still very interested in this functionality.

[github-actions]

This issue is marked as stale because it has been open for 30 days with no activity.

You should take one of the following actions:

• Manually close this issue if it is no longer relevant
• Comment if you have more information to share

This issue will be automatically closed in 7 days if no further activity occurs.

[panjf2000]

I plan on doing this next month if nothing else gets in the way.

[panjf2000]

Updated:
The root cause should have been the loopback interface which doesn't normally receive broadcast packets. I might need to test it on "eth0".

---

After some tests, I found that setting SO_BINDTODEVICE to an interface had the same behavior as binding the specific IP address: no broadcast packet would be received. That is to say, your scenario can never happen by setting SO_BINDTODEVICE, why would you say it can? Did you run any tests to support your assumption? @shaunco