* Moving TransferParse from file to its own package

* Adding tests for parse

Switched health and autopath plugin to allow any plugins to be used instead
of a hardcoded list. I did not switch federation over since it wasn't
obvious that anything other than kubernetes could be used with it.

Fixes #1291

When z.Tree.Prev returns zero we should break out of this loop, not
use elem as if nothing has happened.

Can be triggered by sending edns0 to unsigned zone.

DS is peculiar and needs some special cases to do the right things.
For a direct query return the DS, but if we're hitting below the
delegation return the NS+DS in a delegation response.

Don't put NS records in the answer section; if we see it is delegated
we should put the records in the AUTHORITY section.

This removes a special case, which is good.

add an example and some typos fixed.

If protocol is https_google we do the health checks wrong, i.e. we're healthchecking the wrong
endpoint, hence the health check code below should not be executed. See issue #1202.
This is an ugly hack and the thing requires a rethink. Possibly in conjunction with moving
to the *forward* plugin.

Qunar (https://qunar.com) is the leading travel booking service in China. They
use CoreDNS for service discovery of the GPU machine learning cloud they build
with TensorFlow and Kubernetes.

There is a very nice talk in KubeCon 2017, presented by Ye Lu:
https://kccncna17.sched.com/event/CU61/all-you-need-to-know-to-build-your-gpu-machine-learning-cloud-b-ye-lu-qunarSigned-off-by: Yong Tang <yong.tang.github@outlook.com>

Hide it

- added dnstapEncoder object which incapsulates marshalling of dnstap
   messages to protobuf and writing data to connection

 - dnstapEncoder writes data directly to connection object. It doesn't
   use the framestream's "write" method, because it writes data to
   intermediate buffer (bufio.Writer) which leads to unnecessary
   data copying and drops the performance

 - dnstapEncoder reuses a preallocated buffer for marshalling dnstap
   messages. Many messages are added to the same buffer. They are
   separated with a "frame length" 4-byte values, so the buffer content
   is writen to connection object in the format compatible with
   framestream library

 - added test which guarantees that dnstapEncoder output is the same
   as framestream Encoder output

 - the performance increase is about 50% in (dio *dnstapIO) serve() method
   of dnstap plugin. The overall coredns performance increase is about 10%
   in the following configuration:

   .:1053 {
       erratic {
           drop 0
           truncate 0
           delay 0
       }
       dnstap tcp://127.0.0.1:6000 full
       errors stdout
   }

   tested with dnsperf tool

This fix tries to address the issue raised in 1275 to clarify
the syntax for multiple endpoints specification.

This fix also adds additional test cases to demo the usage.

This fix fixes 1275.
Signed-off-by: Yong Tang <yong.tang.github@outlook.com>

Looks like the last commit missed a `~` in the docs.
Signed-off-by: Yong Tang <yong.tang.github@outlook.com>

* Add NSID plugin support for CoreDNS

This fix adds NSID plugin support for CoreDNS, as was proposed
in 1256.
Signed-off-by: Yong Tang <yong.tang.github@outlook.com>

* Add test cases for NSID plugin
Signed-off-by: Yong Tang <yong.tang.github@outlook.com>

* Generate code for NSID plugin
Signed-off-by: Yong Tang <yong.tang.github@outlook.com>

* Use hostname as the default (as with bind), and remove unneeded copy
Signed-off-by: Yong Tang <yong.tang.github@outlook.com>

* Add README.md
Signed-off-by: Yong Tang <yong.tang.github@outlook.com>

When we failing the healthcheck we should increate the fails for this
host; which is the *bleeping* point of doing the HC in the first place.

Add the missing atomic.Adds.

Fixes #1247

HC for google_https does not work because it does not HC the correct
set of IPs. Hard to solve in the current code - rather use forward as
the new impl. and leave this as legacy.

Remove superfluous println that we are refreshing the dns.google.com IP
address.

* metrics: use exponentional buckets

See  https://github.com/kubernetes/dns/pull/168

See #1211 for disuccsion; current code is probably the wrong thing to
do; rethink if we need/want this. Comment out the code for now.

* plugin/rewrite: don't set or use ecs.DraftOption

Don't know why we are accessing this and explicitally setting it to
False (the default).
Any kill with fire - makes the build, build again.

* remove this

- Dnstap require protocol, address and port info about peer. So, I
   removed conversion of TCPAddr to IPAddr

* Use seconds for bucketing.
* Add template for high accuracy timing buckets to avoid copy-pasta.

Use caching, just ask etcd for every query. This also improves
throughput because the single lock in inflght is bypassed.

* ipv6 UT for endpoint

* tweak

Remove most 303 TTLs (those get skipped by the test) and use 5, which is
the default for all tests.

The TTL was still zero, instead it should adhere to the k.ttl setting.
Change this and update tests to match.

* clean up and add some ipv6 tests

* aaaa endpoint

* gofmt