SkyDNS2是SkyDNS Version 2.x的统称，其官方文档只有README.md，网上能找到的资料也不多，因此需要我们自行对代码进行一定的分析，才能对其有更好的理解，这就是本文的工作，通过走读SkyDNS的代码，了解其内部架构及其工作原理。

说明

• SkyDNS2的github地址: https://github.com/skynetservices/skydns
• Version: v2.5.3a

SkyDNS架构

关于SkyDNS是什么？.... 这些知识，请前往官网了解。

下面我直接把我阅读代码后理解的SkyDNS架构贴出来：

SkyDNS工作原理

SkyDNS Server的工作，依赖后端Key-Value存储的支持。当前支持etcd或etcd3作为Backend（架构图中蓝色部分），为SkyDNS提供配置和数据的管理。

通过环境变量ETCD_MACHINES进行etcd cluster的配置，如果Backend为etcd3，还需要设置etcd中/v2/keys//skydns/config/etcd3为true。SkyDNS中有etcd client模块，负责与ETCD_MACHINES的通信。

SkyDNS主要对应的etcd key path如下：

/v2/keys/skydns/config
/v2/keys/skydns/local/skydns/east/production/rails
/v2/keys/skydns/local/skydns/dns/stub
/v2/keys/skydns/local/skydns/...

通过如下环境变量的配置，支持prometheus监控（架构图中棕色部分）。如果想disable prometheus监控，则配置环境变量PROMETHEUS_PORT的值为0即可。

Port      = os.Getenv("PROMETHEUS_PORT")
Path      = envOrDefault("PROMETHEUS_PATH", "/metrics")
Namespace = envOrDefault("PROMETHEUS_NAMESPACE", "skydns")
Subsystem = envOrDefault("PROMETHEUS_SUBSYSTEM", "skydns")

如果/v2/keys/skydns/config/nameservers有值，则SkyDNS解析不了的Domain，会forward到对应的这些IP:Port构成的nameservers，由它们进行解析（架构图中绿色部分）。

参考官方文档https://github.com/skynetservices/skydns/blob/master/README.md完成参数配置后，便可启动SkyDNS。

SkyDNS Server的启动过程如下：

• 创建etcd client对象；
• dns_addr 和 nameservers参数合法性检查；
• 加载启动参数到etcd，覆盖/v2/keys/skydns/config中原有数据；
• 配置SkyDNS Server参数的default值，并创建SkyDNS server对象；
• 去etcd中加载.../dns/stub/<domain>/xx数据作为server的stub zones数据，并启动对.../dns/stub/的watcher，一旦有数据更新，就加载到server的stub zones数据中；
• 注册SkyDNS metrics到prometheus；
• 然后在/v2/keys/skydns/config/dns_addr配置的interface和port上开启tcp/udp监听服务并block住，由此开始提供DSN服务。

在github.com/skynetservices/skydns/server/server.go中的ServeDNS方法覆盖了miekg/dns/server中的ServeMux.ServeDNS方法，由自实现的ServeDNS提供来处理DNS client的请求。

github.com/skynetservices/skydns/server/server.go// ServeDNS is the handler for DNS requests, responsible for parsing DNS request, possibly forwarding
// it to a real dns server and returning a response.
func (s *server) ServeDNS(w dns.ResponseWriter, req *dns.Msg) {...// Check cache first.m1 := s.rcache.Hit(q, dnssec, tcp, m.Id)if m1 != nil {...// Still round-robin even with hits from the cache.// Only shuffle A and AAAA records with each other.if q.Qtype == dns.TypeA || q.Qtype == dns.TypeAAAA {s.RoundRobin(m1.Answer)}...return}for zone, ns := range *s.config.stub {if strings.HasSuffix(name, "." + zone) || name == zone {metrics.ReportRequestCount(req, metrics.Stub)resp := s.ServeDNSStubForward(w, req, ns)if resp != nil {s.rcache.InsertMessage(cache.Key(q, dnssec, tcp), resp)}metrics.ReportDuration(resp, start, metrics.Stub)metrics.ReportErrorCount(resp, metrics.Stub)return}}...if name == s.config.Domain {if q.Qtype == dns.TypeSOA {m.Answer = []dns.RR{s.NewSOA()}return}if q.Qtype == dns.TypeDNSKEY {if s.config.PubKey != nil {m.Answer = []dns.RR{s.config.PubKey}return}}}if q.Qclass == dns.ClassCHAOS {if q.Qtype == dns.TypeTXT {switch name {case "authors.bind.":fallthroughcase s.config.Domain:hdr := dns.RR_Header{Name: q.Name, Rrtype: dns.TypeTXT, Class: dns.ClassCHAOS, Ttl: 0}authors := []string{"Erik St. Martin", "Brian Ketelsen", "Miek Gieben", "Michael Crosby"}for _, a := range authors {m.Answer = append(m.Answer, &dns.TXT{Hdr: hdr, Txt: []string{a}})}for j := 0; j < len(authors)*(int(dns.Id())%4+1); j++ {q := int(dns.Id()) % len(authors)p := int(dns.Id()) % len(authors)if q == p {p = (p + 1) % len(authors)}m.Answer[q], m.Answer[p] = m.Answer[p], m.Answer[q]}returncase "version.bind.":fallthroughcase "version.server.":hdr := dns.RR_Header{Name: q.Name, Rrtype: dns.TypeTXT, Class: dns.ClassCHAOS, Ttl: 0}m.Answer = []dns.RR{&dns.TXT{Hdr: hdr, Txt: []string{Version}}}returncase "hostname.bind.":fallthroughcase "id.server.":// TODO(miek): machine name to returnhdr := dns.RR_Header{Name: q.Name, Rrtype: dns.TypeTXT, Class: dns.ClassCHAOS, Ttl: 0}m.Answer = []dns.RR{&dns.TXT{Hdr: hdr, Txt: []string{"localhost"}}}return}}// still here, failm.SetReply(req)m.SetRcode(req, dns.RcodeServerFailure)return}switch q.Qtype {case dns.TypeNS:if name != s.config.Domain {break}// Lookup s.config.DnsDomainrecords, extra, err := s.NSRecords(q, s.config.dnsDomain)if isEtcdNameError(err, s) {m = s.NameError(req)return}m.Answer = append(m.Answer, records...)m.Extra = append(m.Extra, extra...)case dns.TypeA, dns.TypeAAAA:records, err := s.AddressRecords(q, name, nil, bufsize, dnssec, false)if isEtcdNameError(err, s) {m = s.NameError(req)return}m.Answer = append(m.Answer, records...)case dns.TypeTXT:records, err := s.TXTRecords(q, name)if isEtcdNameError(err, s) {m = s.NameError(req)return}m.Answer = append(m.Answer, records...)case dns.TypeCNAME:records, err := s.CNAMERecords(q, name)if isEtcdNameError(err, s) {m = s.NameError(req)return}m.Answer = append(m.Answer, records...)case dns.TypeMX:records, extra, err := s.MXRecords(q, name, bufsize, dnssec)if isEtcdNameError(err, s) {m = s.NameError(req)return}m.Answer = append(m.Answer, records...)m.Extra = append(m.Extra, extra...)default:fallthrough // also catch other types, so that they return NODATAcase dns.TypeSRV:records, extra, err := s.SRVRecords(q, name, bufsize, dnssec)if err != nil {if isEtcdNameError(err, s) {m = s.NameError(req)return}logf("got error from backend: %s", err)if q.Qtype == dns.TypeSRV { // Otherwise NODATAm = s.ServerFailure(req)return}}// if we are here again, check the types, because an answer may only// be given for SRV. All other types should return NODATA, the// NXDOMAIN part is handled in the above code. TODO(miek): yes this// can be done in a more elegant manor.if q.Qtype == dns.TypeSRV {m.Answer = append(m.Answer, records...)m.Extra = append(m.Extra, extra...)}}if len(m.Answer) == 0 { // NODATA responsem.Ns = []dns.RR{s.NewSOA()}m.Ns[0].Header().Ttl = s.config.MinTtl}
}

上面代码逻辑比较复杂，细节上需要你慢慢去理解，简短的可以总结如下：

• 如架构图中标注的线路1：如果在SkyDNS维护的cache中找到对应Msg，则从cache中读取并返回Msg给DNS client;
• 如架构图中标注的线路2：如果在cache中没有对应的记录，并且是需要DNS forward的场景（比如name匹配到stub zones等），则将请求forward到对应的DNS servers进行处理；
• 如架构图中标注的线路3：如果在cache中没有对应的记录，并且Question Type为A/AAAA，SRV等类型时，就通过etcd client去etcd cluster中获取对应的Rule，并构造Msg返回。

总结

通过走读SkyDNS的代码，了解其内部架构及其工作原理。

Mark

SkyDNS2 Changes since version 1：

• Does away with Raft and uses etcd (which uses raft).
• Makes it possible to query arbitrary domain names.
• Is a thin layer above etcd, that translates etcd keys and values to the DNS.
• Does DNSSEC with NSEC3 instead of NSEC.