vxlan.go 6.1 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283
  1. package vxlan
  2. import (
  3. "encoding/json"
  4. "fmt"
  5. "net"
  6. "sync"
  7. "time"
  8. log "github.com/coreos/flannel/Godeps/_workspace/src/github.com/golang/glog"
  9. "github.com/coreos/flannel/Godeps/_workspace/src/github.com/vishvananda/netlink"
  10. "github.com/coreos/flannel/backend"
  11. "github.com/coreos/flannel/subnet"
  12. "github.com/coreos/flannel/pkg/ip"
  13. "github.com/coreos/flannel/pkg/task"
  14. )
  15. const (
  16. defaultVNI = 1
  17. )
  18. type VXLANBackend struct {
  19. sm *subnet.SubnetManager
  20. rawCfg json.RawMessage
  21. cfg struct {
  22. VNI int
  23. Port int
  24. }
  25. dev *vxlanDevice
  26. stop chan bool
  27. wg sync.WaitGroup
  28. rts routes
  29. }
  30. func New(sm *subnet.SubnetManager, config json.RawMessage) backend.Backend {
  31. vb := &VXLANBackend{
  32. sm: sm,
  33. rawCfg: config,
  34. stop: make(chan bool),
  35. }
  36. vb.cfg.VNI = defaultVNI
  37. return vb
  38. }
  39. func newSubnetAttrs(pubIP net.IP, mac net.HardwareAddr) (*subnet.LeaseAttrs, error) {
  40. data, err := json.Marshal(&vxlanLeaseAttrs{hardwareAddr(mac)})
  41. if err != nil {
  42. return nil, err
  43. }
  44. return &subnet.LeaseAttrs{
  45. PublicIP: ip.FromIP(pubIP),
  46. BackendType: "vxlan",
  47. BackendData: json.RawMessage(data),
  48. }, nil
  49. }
  50. func (vb *VXLANBackend) Init(extIface *net.Interface, extIP net.IP) (*backend.SubnetDef, error) {
  51. // Parse our configuration
  52. if len(vb.rawCfg) > 0 {
  53. if err := json.Unmarshal(vb.rawCfg, &vb.cfg); err != nil {
  54. return nil, fmt.Errorf("error decoding UDP backend config: %v", err)
  55. }
  56. }
  57. devAttrs := vxlanDeviceAttrs{
  58. vni: uint32(vb.cfg.VNI),
  59. name: fmt.Sprintf("flannel.%v", vb.cfg.VNI),
  60. vtepIndex: extIface.Index,
  61. vtepAddr: extIP,
  62. vtepPort: vb.cfg.Port,
  63. }
  64. var err error
  65. for {
  66. vb.dev, err = newVXLANDevice(&devAttrs)
  67. if err == nil {
  68. break
  69. } else {
  70. log.Error("VXLAN init: ", err)
  71. log.Info("Retrying in 1 second...")
  72. // wait 1 sec before retrying
  73. time.Sleep(1*time.Second)
  74. }
  75. }
  76. sa, err := newSubnetAttrs(extIP, vb.dev.MACAddr())
  77. if err != nil {
  78. return nil, err
  79. }
  80. sn, err := vb.sm.AcquireLease(sa, vb.stop)
  81. if err != nil {
  82. if err == task.ErrCanceled {
  83. return nil, err
  84. } else {
  85. return nil, fmt.Errorf("failed to acquire lease: %v", err)
  86. }
  87. }
  88. // vxlan's subnet is that of the whole overlay network (e.g. /16)
  89. // and not that of the individual host (e.g. /24)
  90. vxlanNet := ip.IP4Net{
  91. IP: sn.IP,
  92. PrefixLen: vb.sm.GetConfig().Network.PrefixLen,
  93. }
  94. if err = vb.dev.Configure(vxlanNet); err != nil {
  95. return nil, err
  96. }
  97. return &backend.SubnetDef{
  98. Net: sn,
  99. MTU: vb.dev.MTU(),
  100. }, nil
  101. }
  102. func (vb *VXLANBackend) Run() {
  103. vb.wg.Add(1)
  104. go func() {
  105. vb.sm.LeaseRenewer(vb.stop)
  106. vb.wg.Done()
  107. }()
  108. log.Info("Watching for L2/L3 misses")
  109. misses := make(chan *netlink.Neigh, 100)
  110. // Unfortunately MonitorMisses does not take a cancel channel
  111. // as there's no wait to interrupt netlink socket recv
  112. go vb.dev.MonitorMisses(misses)
  113. log.Info("Watching for new subnet leases")
  114. evts := make(chan subnet.EventBatch)
  115. vb.wg.Add(1)
  116. go func() {
  117. vb.sm.WatchLeases(evts, vb.stop)
  118. vb.wg.Done()
  119. }()
  120. defer vb.wg.Wait()
  121. for {
  122. select {
  123. case miss := <-misses:
  124. vb.handleMiss(miss)
  125. case evtBatch := <-evts:
  126. vb.handleSubnetEvents(evtBatch)
  127. case <-vb.stop:
  128. return
  129. }
  130. }
  131. }
  132. func (vb *VXLANBackend) Stop() {
  133. close(vb.stop)
  134. }
  135. func (vb *VXLANBackend) Name() string {
  136. return "VXLAN"
  137. }
  138. // So we can make it JSON (un)marshalable
  139. type hardwareAddr net.HardwareAddr
  140. func (hw hardwareAddr) MarshalJSON() ([]byte, error) {
  141. return []byte(fmt.Sprintf("%q", net.HardwareAddr(hw))), nil
  142. }
  143. func (hw *hardwareAddr) UnmarshalJSON(b []byte) error {
  144. if len(b) < 2 || b[0] != '"' || b[len(b)-1] != '"' {
  145. return fmt.Errorf("error parsing hardware addr")
  146. }
  147. b = b[1:len(b)-1]
  148. mac, err := net.ParseMAC(string(b))
  149. if err != nil {
  150. return err
  151. }
  152. *hw = hardwareAddr(mac)
  153. return nil
  154. }
  155. type vxlanLeaseAttrs struct {
  156. VtepMAC hardwareAddr
  157. }
  158. func (vb *VXLANBackend) handleSubnetEvents(batch subnet.EventBatch) {
  159. for _, evt := range batch {
  160. switch evt.Type {
  161. case subnet.SubnetAdded:
  162. log.Info("Subnet added: ", evt.Lease.Network)
  163. if evt.Lease.Attrs.BackendType != "vxlan" {
  164. log.Warningf("Ignoring non-vxlan subnet: type=%v", evt.Lease.Attrs.BackendType)
  165. continue
  166. }
  167. var attrs vxlanLeaseAttrs
  168. if err := json.Unmarshal(evt.Lease.Attrs.BackendData, &attrs); err != nil {
  169. log.Error("Error decoding subnet lease JSON: ", err)
  170. continue
  171. }
  172. vb.rts.set(evt.Lease.Network, evt.Lease.Attrs.PublicIP.ToIP(), net.HardwareAddr(attrs.VtepMAC))
  173. case subnet.SubnetRemoved:
  174. log.Info("Subnet removed: ", evt.Lease.Network)
  175. vb.rts.remove(evt.Lease.Network)
  176. if evt.Lease.Attrs.BackendType != "vxlan" {
  177. log.Warningf("Ignoring non-vxlan subnet: type=%v", evt.Lease.Attrs.BackendType)
  178. continue
  179. }
  180. var attrs vxlanLeaseAttrs
  181. if err := json.Unmarshal(evt.Lease.Attrs.BackendData, &attrs); err != nil {
  182. log.Error("Error decoding subnet lease JSON: ", err)
  183. continue
  184. }
  185. if len(attrs.VtepMAC) > 0 {
  186. vb.dev.DelL2(net.HardwareAddr(attrs.VtepMAC), evt.Lease.Attrs.PublicIP.ToIP())
  187. }
  188. default:
  189. log.Error("Internal error: unknown event type: ", int(evt.Type))
  190. }
  191. }
  192. }
  193. func (vb *VXLANBackend) handleMiss(miss *netlink.Neigh) {
  194. switch {
  195. case len(miss.IP) == 0 && len(miss.HardwareAddr) == 0:
  196. log.Info("Ignoring nil miss")
  197. case len(miss.IP) == 0:
  198. vb.handleL2Miss(miss)
  199. case len(miss.HardwareAddr) == 0:
  200. vb.handleL3Miss(miss)
  201. default:
  202. log.Infof("Ignoring not a miss: %v, %v", miss.HardwareAddr, miss.IP)
  203. }
  204. }
  205. func (vb *VXLANBackend) handleL2Miss(miss *netlink.Neigh) {
  206. log.Infof("L2 miss: %v", miss.HardwareAddr)
  207. rt := vb.rts.findByVtepMAC(miss.HardwareAddr)
  208. if rt == nil {
  209. log.Infof("Route for %v not found", miss.HardwareAddr)
  210. return
  211. }
  212. if err := vb.dev.AddL2(miss.HardwareAddr, rt.vtepIP); err != nil {
  213. log.Errorf("AddL2 failed: %v", err)
  214. } else {
  215. log.Info("AddL2 succeeded")
  216. }
  217. }
  218. func (vb *VXLANBackend) handleL3Miss(miss *netlink.Neigh) {
  219. log.Infof("L3 miss: %v", miss.IP)
  220. rt := vb.rts.findByNetwork(ip.FromIP(miss.IP))
  221. if rt == nil {
  222. log.Infof("Route for %v not found", miss.IP)
  223. return
  224. }
  225. if err := vb.dev.AddL3(miss.IP, rt.vtepMAC); err != nil {
  226. log.Errorf("AddL3 failed: %v", err)
  227. } else {
  228. log.Info("AddL3 succeeded")
  229. }
  230. }