net: don't keep reading from UDP resolver after truncated packet

[ianlancetaylor]

When the host or dig programs see a malformed packet from a resolver when using UDP, they fall back to using TCP. The net package resolver does not do this; it simply ignores the malformed packet (in (*dnsPacketConn).dnsRoundTrip in net/dnsclient_unix.go). This was done for #13281. I suggest that we do the same.

This is showing up right now for me when I run go test -test.run=TestLookupLongTXT net. I see this:

--- FAIL: TestLookupLongTXT (10.00s)
	lookup_test.go:334: lookup golang.rsc.io on 127.0.0.1:53: read udp 127.0.0.1:39779->127.0.0.1:53: i/o timeout
FAIL
FAIL	net	10.023s

If I run dig -t txt golang.rsc.io the output starts with

;; Warning: Message parser reports malformed message packet.
;; Truncated, retrying in TCP mode.

I suggest that we keep the current behavior for the !resp.IsResponseTo(query) case but change the behavior for an Unpack failure to drop right back to TCP.

CC @mdempsky

[ianlancetaylor]

This should be re-checked given the work done for #21160.

[ianlancetaylor]

See also #22857.

[ianlancetaylor]

CC @mikioh @iangudger

[iangudger]

With the new DNS client, the chances of truncated DNS messages causing problems is greatly reduced, not not eliminated. I believe that now in cases where not all answers are contained in the read UDP message we will either try another DNS server or error out. I agree that trying TCP would be better, but at least failing fast is better than waiting for a timeout.