net: dns lookup should retry on TCP for truncated response

[gopherbot]

by [email protected]:

* What steps will reproduce the problem?
1. Setup a DNS server with SRV records too large to fit into a single UDP datagram.
2. Query said record via net.LookupSRV()
3. run "tcpdump port 53" with appropriate options to verify that Go tries only
via UDP, receives a truncated packet (DNS Flag "Truncated") and then doesn't
retry via TCP.

* What is the expected output?

net.LookupSRV() shouldn't return an error, but return appropriate record data. Doing the
equivalent query via Dig succeeds, as it sends the exact same UDP query at first, but
then tries TCP after seeing a truncated answer.

* What do you see instead?

net.LookupSRV() fails with an error: "lookup <...record-name...>. on
[::1]:53: no such host"

* Which compiler are you using (5g, 6g, 8g, gccgo)?

Shouldn't matter, see below.

* Which operating system are you using?

Ubuntu Linux 12.04.2 LTS

* Which version are you using?  (run 'go version')

Shouldn't matter, since only UDP is hardcoded in lookup() in HEAD (2013-06-11):

http://golang.org/src/pkg/net/dnsclient_unix.go#L90

See also "Truncation Flag" in
http://www.tcpipguide.com/free/t_DNSMessageHeaderandQuestionSectionFormat.htm

Note that LookupIP(), etc. seem to be using different lookup implementations, so they
might not be affected by the same problem.

[gopherbot]

Comment 1 by [email protected]:

Correction: several other net.Lookup*() functions (net.LookupIP(), net.LookupMX(), etc.)
seem to all use lookup() in the backend as well, so they should also be affected by this
problem.

[davecheney]

Comment 2:

Are you using the net package with cgo enabled (the default) or disabled ?
If cgo is being used then shouldn't this be handled directly by the libc resolver ?

Status changed to WaitingForReply.

[gopherbot]

Comment 3 by [email protected]:

I am simply building Go from source via src/all.bash. How do I enable using the cgo
versions of those functions? I don't think they're using cgo right now, since local
changes I made to the Go version of lookup() took effect after rebuilding.
Right now my go version is the latest tip:
go version devel +ae79f385177d Wed Jun 12 14:05:13 2013 +1000 linux/amd64

[davecheney]

Comment 4:

cgo is enabled by default on all platforms. You can check the status
of cgo with `go env`

[gopherbot]

Comment 5 by [email protected]:

go env reports:
$ go env
GOARCH="amd64"
GOBIN=""
GOCHAR="6"
GOEXE=""
GOHOSTARCH="amd64"
GOHOSTOS="linux"
GOOS="linux"
GOPATH="/home/julius/go:/home/julius/gosrc"
GORACE=""
GOROOT="/home/julius/go"
GOTOOLDIR="/home/julius/go/pkg/tool/linux_amd64"
CC="gcc"
GOGCCFLAGS="-g -O2 -fPIC -m64 -pthread"
CGO_ENABLED="1"
$ which go
/home/julius/go/bin/go
$ ls -l /home/julius/go/bin/go
-rwxrwxr-x 1 julius julius 8096584 Jun 12 13:49 /home/julius/go/bin/go
So this is the version I just built. I added a panic() just before the Dial("udp", ...)
in net.tryOneName() and sure enough, it panics when I do LookupSRV().
Maybe there is no CGO version of this functionality yet?

[minux]

Comment 6:

i confirm that there isn't cgo version for this yet.
does getaddrinfo(3) support SRV lookup?
(AI_SRVLOOKUP??)

[rsc]

Comment 7:

Even if we get cgo doing this, the pure Go version should be fixed.

Labels changed: added priority-later, go1.2, removed priority-triage.

Status changed to Accepted.

[gopherbot]

Comment 8 by [email protected]:

Seems like it's not supported: http://sourceware.org/bugzilla/show_bug.cgi?id=2099

[rsc]

Comment 9:

Let's get the pure DNS version working, but no cgo fallback to it.

Labels changed: added go1.2maybe, removed go1.2.

[gopherbot]

Comment 10 by [email protected]:

Considered including https://github.com/miekg/dns in the core? Might be too heavy, but
it's a great, idiomatic library in a similar style to net/http.

[rsc]

Comment 11:

Labels changed: added feature.

[bradfitz]

Comment 12:

This issue was closed by revision 0a3cb7e.

Status changed to Fixed.