Creating TCPSocket fails with getaddrinfo: Non-recoverable failure in name resolution (Socket::Error)

[charles-l]

May be related to #2568?

Using the same code in that issue:

require "socket"
p TCPSocket.new("google.com", 80)
p TCPSocket.new("google.com", 80)

The resulting output is (for Crystal - it works with ruby code, so It's probably something related to libevent):

getaddrinfo: Non-recoverable failure in name resolution (Socket::Error)
[4316279] *CallStack::unwind:Array(Pointer(Void)) +87
[4316170] *CallStack#initialize<CallStack>:Array(Pointer(Void)) +10
[4316122] *CallStack::new:CallStack +42
[4368809] *Socket::Error +41
[4368745] *Socket::Error +105
[4363191] *TCPSocket#initialize<TCPSocket, String, Int32, Nil, Nil>:Nil +1655
[4361512] *TCPSocket::new<String, Int32, Nil, Nil>:TCPSocket +184
[4361323] *TCPSocket::new<String, Int32>:TCPSocket +27
[4293247] ???
[4301177] main +41
[140049360811841] __libc_start_main +241
[4291273] _start +41
[0] ???

I'm on ArchLinux, running Crystal version 0.17.3, and libevent version 2.0.22-2.

[asterite]

We should maybe revert the commit that introduces this "feature" (anti-feature, I'd say, because it's broken). Problem is, dns_timeout would be ignored for now, and of course things will be a bit slower. I'll check with @waj when he comes back, I prefer to have something that works but is slower than something fast but broken. Of course the best thing to do is to fix this (don't know how, maybe not using libevent), but that will take more time.

[ysbaddaden]

If we don't use libevent dns, then we're up to blocking Crystal altogether whenever we make a request... not just the current fiber, but everything, which can take from some hundred milliseconds to 5 seconds if unlucky :(

I looked around and there aren't any correct solution. Even getaddrinfo_a, (only available by glibc) uses signals instead of a couple of functions (start, poll).

[asterite]

I guess then we'd need to do what @waj said here

[charles-l]

So, is this something that's only broken on Linux or is it an issue elsewhere as well? I'll have to take a closer look at the IPSocket code, but I'm not sure what's causing this, since the example libnotify code for a DNS lookup works in C.

[jhass]

Should we close all these tickets and create a more general meta ticket about implementing a proper DNS/mDNS resolver?

[asterite]

@jhass Please! And thank you for cleaning up some issues :-)

[jhass]

Closing in favor of #2660