set the don't fragment (DF) bit on Windows

[tobyxdd]

No description provided.

[codecov]

Codecov Report

Merging #3155 (9903019) into master (0a06396) will decrease coverage by 0.05%.
The diff coverage is 53.85%.

@@            Coverage Diff             @@
##           master    #3155      +/-   ##
==========================================
- Coverage   85.41%   85.36%   -0.05%     
==========================================
  Files         131      131              
  Lines        9698     9710      +12     
==========================================
+ Hits         8283     8288       +5     
- Misses       1045     1049       +4     
- Partials      370      373       +3     

Impacted Files	Coverage Δ
conn_windows.go	57.14% <53.85%> (-5.36%)	⬇️
internal/utils/rand.go	62.50% <0.00%> (-12.50%)	⬇️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 0a06396...9903019. Read the comment docs.

[marten-seemann]

Is there any way we can test this? I realize testing that DF is actually set might be difficult, but at least we should be able to pass a net.UDPConn to newConn and make sure it doesn’t return an error.

[tobyxdd]

Not sure what happened to the failed test here but it doesn't look related to my changes

[marten-seemann]

I haven't checked Wireshark to see if the DF bit is actually set. @tobyxdd, have you?

[tobyxdd]

Yes I did

[tobyxdd]

without vs with on Windows

[marten-seemann]

Thank you @tobyxdd!

[marten-seemann]

@tobyxdd Have you checked if this also works on IPv6? I was just working on an equivalent PR for OSX, and it looks like there are different syscalls for IPv4 and IPv6.

[marten-seemann]

Or am I misunderstanding something here? IPv6 shouldn't allow any packet fragmentation anyway, but then why is there a unix.IPV6_DONTFRAG (see https://pkg.go.dev/golang.org/x/sys/unix).

[tobyxdd]

Just realized that IPv6 does have fragmentation (as an extension header), but only at the source, not routers.

https://en.wikipedia.org/wiki/IPv6_packet#Fragment
https://tools.ietf.org/html/rfc3542
https://blog.apnic.net/2016/05/19/fragmenting-ipv6/

[tobyxdd]

So yes I guess we do need to set DF for IPv6 as well (or get local MTU from the system?)

[marten-seemann]

So yes I guess we do need to set DF for IPv6 as well (or get local MTU from the system?)

I'm not sure. It depends on what happens when we try to send a too large packet. We'd need some way to feed that information back into the MTU discovery code. Getting the local MTU from the system and initializing the MTU discoverer with that seems like the safer option, but I'm not sure it's needed in practice, as our maximum packet size is below 1500 bytes anyway.