|
|
@@ -45,53 +45,33 @@ const (
|
|
|
ConnWriteTimeout = 5 * time.Second
|
|
|
)
|
|
|
|
|
|
-type Sender interface {
|
|
|
- // StartStreaming enables streaming in the sender using the given writer,
|
|
|
- // which provides a fast and efficient way to send appendEntry messages.
|
|
|
- StartStreaming(w WriteFlusher, to types.ID, term uint64) (done <-chan struct{}, err error)
|
|
|
- Update(u string)
|
|
|
- // Send sends the data to the remote node. It is always non-blocking.
|
|
|
- // It may be fail to send data if it returns nil error.
|
|
|
- Send(m raftpb.Message) error
|
|
|
- // Stop performs any necessary finalization and terminates the Sender
|
|
|
- // elegantly.
|
|
|
- Stop()
|
|
|
-
|
|
|
- // Pause pauses the sender. The sender will simply drops all incoming
|
|
|
- // messages without retruning an error.
|
|
|
- Pause()
|
|
|
-
|
|
|
- // Resume resumes a paused sender.
|
|
|
- Resume()
|
|
|
-}
|
|
|
-
|
|
|
-func NewSender(tr http.RoundTripper, u string, id types.ID, cid types.ID, p Processor, fs *stats.FollowerStats, shouldstop chan struct{}) *sender {
|
|
|
- s := &sender{
|
|
|
+func NewPeer(tr http.RoundTripper, u string, id types.ID, cid types.ID, r Raft, fs *stats.FollowerStats, shouldstop chan struct{}) *peer {
|
|
|
+ p := &peer{
|
|
|
id: id,
|
|
|
active: true,
|
|
|
tr: tr,
|
|
|
u: u,
|
|
|
cid: cid,
|
|
|
- p: p,
|
|
|
+ r: r,
|
|
|
fs: fs,
|
|
|
shouldstop: shouldstop,
|
|
|
batcher: NewBatcher(100, appRespBatchMs*time.Millisecond),
|
|
|
propBatcher: NewProposalBatcher(100, propBatchMs*time.Millisecond),
|
|
|
q: make(chan *raftpb.Message, senderBufSize),
|
|
|
}
|
|
|
- s.wg.Add(connPerSender)
|
|
|
+ p.wg.Add(connPerSender)
|
|
|
for i := 0; i < connPerSender; i++ {
|
|
|
- go s.handle()
|
|
|
+ go p.handle()
|
|
|
}
|
|
|
- return s
|
|
|
+ return p
|
|
|
}
|
|
|
|
|
|
-type sender struct {
|
|
|
+type peer struct {
|
|
|
id types.ID
|
|
|
cid types.ID
|
|
|
|
|
|
tr http.RoundTripper
|
|
|
- p Processor
|
|
|
+ r Raft
|
|
|
fs *stats.FollowerStats
|
|
|
shouldstop chan struct{}
|
|
|
|
|
|
@@ -115,201 +95,210 @@ type sender struct {
|
|
|
paused bool
|
|
|
}
|
|
|
|
|
|
-func (s *sender) StartStreaming(w WriteFlusher, to types.ID, term uint64) (<-chan struct{}, error) {
|
|
|
- s.strmSrvMu.Lock()
|
|
|
- defer s.strmSrvMu.Unlock()
|
|
|
- if s.strmSrv != nil {
|
|
|
+// StartStreaming enables streaming in the peer using the given writer,
|
|
|
+// which provides a fast and efficient way to send appendEntry messages.
|
|
|
+func (p *peer) StartStreaming(w WriteFlusher, to types.ID, term uint64) (<-chan struct{}, error) {
|
|
|
+ p.strmSrvMu.Lock()
|
|
|
+ defer p.strmSrvMu.Unlock()
|
|
|
+ if p.strmSrv != nil {
|
|
|
// ignore lower-term streaming request
|
|
|
- if term < s.strmSrv.term {
|
|
|
- return nil, fmt.Errorf("out of data streaming request: term %d, request term %d", term, s.strmSrv.term)
|
|
|
+ if term < p.strmSrv.term {
|
|
|
+ return nil, fmt.Errorf("out of data streaming request: term %d, request term %d", term, p.strmSrv.term)
|
|
|
}
|
|
|
// stop the existing one
|
|
|
- s.strmSrv.stop()
|
|
|
- s.strmSrv = nil
|
|
|
+ p.strmSrv.stop()
|
|
|
+ p.strmSrv = nil
|
|
|
}
|
|
|
- s.strmSrv = startStreamServer(w, to, term, s.fs)
|
|
|
- return s.strmSrv.stopNotify(), nil
|
|
|
+ p.strmSrv = startStreamServer(w, to, term, p.fs)
|
|
|
+ return p.strmSrv.stopNotify(), nil
|
|
|
}
|
|
|
|
|
|
-func (s *sender) Update(u string) {
|
|
|
- s.mu.Lock()
|
|
|
- defer s.mu.Unlock()
|
|
|
- s.u = u
|
|
|
+func (p *peer) Update(u string) {
|
|
|
+ p.mu.Lock()
|
|
|
+ defer p.mu.Unlock()
|
|
|
+ p.u = u
|
|
|
}
|
|
|
|
|
|
+// Send sends the data to the remote node. It is always non-blocking.
|
|
|
+// It may be fail to send data if it returns nil error.
|
|
|
// TODO (xiangli): reasonable retry logic
|
|
|
-func (s *sender) Send(m raftpb.Message) error {
|
|
|
- s.mu.RLock()
|
|
|
- pause := s.paused
|
|
|
- s.mu.RUnlock()
|
|
|
+func (p *peer) Send(m raftpb.Message) error {
|
|
|
+ p.mu.RLock()
|
|
|
+ pause := p.paused
|
|
|
+ p.mu.RUnlock()
|
|
|
if pause {
|
|
|
return nil
|
|
|
}
|
|
|
|
|
|
- s.maybeStopStream(m.Term)
|
|
|
- if shouldInitStream(m) && !s.hasStreamClient() {
|
|
|
- s.initStream(types.ID(m.From), types.ID(m.To), m.Term)
|
|
|
- s.batcher.Reset(time.Now())
|
|
|
+ p.maybeStopStream(m.Term)
|
|
|
+ if shouldInitStream(m) && !p.hasStreamClient() {
|
|
|
+ p.initStream(types.ID(m.From), types.ID(m.To), m.Term)
|
|
|
+ p.batcher.Reset(time.Now())
|
|
|
}
|
|
|
|
|
|
var err error
|
|
|
switch {
|
|
|
case isProposal(m):
|
|
|
- s.propBatcher.Batch(m)
|
|
|
- case canBatch(m) && s.hasStreamClient():
|
|
|
- if !s.batcher.ShouldBatch(time.Now()) {
|
|
|
- err = s.send(m)
|
|
|
+ p.propBatcher.Batch(m)
|
|
|
+ case canBatch(m) && p.hasStreamClient():
|
|
|
+ if !p.batcher.ShouldBatch(time.Now()) {
|
|
|
+ err = p.send(m)
|
|
|
}
|
|
|
case canUseStream(m):
|
|
|
- if ok := s.tryStream(m); !ok {
|
|
|
- err = s.send(m)
|
|
|
+ if ok := p.tryStream(m); !ok {
|
|
|
+ err = p.send(m)
|
|
|
}
|
|
|
default:
|
|
|
- err = s.send(m)
|
|
|
+ err = p.send(m)
|
|
|
}
|
|
|
// send out batched MsgProp if needed
|
|
|
// TODO: it is triggered by all outcoming send now, and it needs
|
|
|
// more clear solution. Either use separate goroutine to trigger it
|
|
|
// or use streaming.
|
|
|
- if !s.propBatcher.IsEmpty() {
|
|
|
+ if !p.propBatcher.IsEmpty() {
|
|
|
t := time.Now()
|
|
|
- if !s.propBatcher.ShouldBatch(t) {
|
|
|
- s.send(s.propBatcher.Message)
|
|
|
- s.propBatcher.Reset(t)
|
|
|
+ if !p.propBatcher.ShouldBatch(t) {
|
|
|
+ p.send(p.propBatcher.Message)
|
|
|
+ p.propBatcher.Reset(t)
|
|
|
}
|
|
|
}
|
|
|
return err
|
|
|
}
|
|
|
|
|
|
-func (s *sender) send(m raftpb.Message) error {
|
|
|
+func (p *peer) send(m raftpb.Message) error {
|
|
|
// TODO: don't block. we should be able to have 1000s
|
|
|
// of messages out at a time.
|
|
|
select {
|
|
|
- case s.q <- &m:
|
|
|
+ case p.q <- &m:
|
|
|
return nil
|
|
|
default:
|
|
|
log.Printf("sender: dropping %s because maximal number %d of sender buffer entries to %s has been reached",
|
|
|
- m.Type, senderBufSize, s.u)
|
|
|
+ m.Type, senderBufSize, p.u)
|
|
|
return fmt.Errorf("reach maximal serving")
|
|
|
}
|
|
|
}
|
|
|
|
|
|
-func (s *sender) Stop() {
|
|
|
- close(s.q)
|
|
|
- s.wg.Wait()
|
|
|
- s.strmSrvMu.Lock()
|
|
|
- if s.strmSrv != nil {
|
|
|
- s.strmSrv.stop()
|
|
|
- s.strmSrv = nil
|
|
|
+// Stop performs any necessary finalization and terminates the peer
|
|
|
+// elegantly.
|
|
|
+func (p *peer) Stop() {
|
|
|
+ close(p.q)
|
|
|
+ p.wg.Wait()
|
|
|
+ p.strmSrvMu.Lock()
|
|
|
+ if p.strmSrv != nil {
|
|
|
+ p.strmSrv.stop()
|
|
|
+ p.strmSrv = nil
|
|
|
}
|
|
|
- s.strmSrvMu.Unlock()
|
|
|
- if s.strmCln != nil {
|
|
|
- s.strmCln.stop()
|
|
|
+ p.strmSrvMu.Unlock()
|
|
|
+ if p.strmCln != nil {
|
|
|
+ p.strmCln.stop()
|
|
|
}
|
|
|
}
|
|
|
|
|
|
-func (s *sender) Pause() {
|
|
|
- s.mu.Lock()
|
|
|
- defer s.mu.Unlock()
|
|
|
- s.paused = true
|
|
|
+// Pause pauses the peer. The peer will simply drops all incoming
|
|
|
+// messages without retruning an error.
|
|
|
+func (p *peer) Pause() {
|
|
|
+ p.mu.Lock()
|
|
|
+ defer p.mu.Unlock()
|
|
|
+ p.paused = true
|
|
|
}
|
|
|
|
|
|
-func (s *sender) Resume() {
|
|
|
- s.mu.Lock()
|
|
|
- defer s.mu.Unlock()
|
|
|
- s.paused = false
|
|
|
+// Resume resumes a paused peer.
|
|
|
+func (p *peer) Resume() {
|
|
|
+ p.mu.Lock()
|
|
|
+ defer p.mu.Unlock()
|
|
|
+ p.paused = false
|
|
|
}
|
|
|
|
|
|
-func (s *sender) maybeStopStream(term uint64) {
|
|
|
- if s.strmCln != nil && term > s.strmCln.term {
|
|
|
- s.strmCln.stop()
|
|
|
- s.strmCln = nil
|
|
|
+func (p *peer) maybeStopStream(term uint64) {
|
|
|
+ if p.strmCln != nil && term > p.strmCln.term {
|
|
|
+ p.strmCln.stop()
|
|
|
+ p.strmCln = nil
|
|
|
}
|
|
|
- s.strmSrvMu.Lock()
|
|
|
- defer s.strmSrvMu.Unlock()
|
|
|
- if s.strmSrv != nil && term > s.strmSrv.term {
|
|
|
- s.strmSrv.stop()
|
|
|
- s.strmSrv = nil
|
|
|
+ p.strmSrvMu.Lock()
|
|
|
+ defer p.strmSrvMu.Unlock()
|
|
|
+ if p.strmSrv != nil && term > p.strmSrv.term {
|
|
|
+ p.strmSrv.stop()
|
|
|
+ p.strmSrv = nil
|
|
|
}
|
|
|
}
|
|
|
|
|
|
-func (s *sender) hasStreamClient() bool {
|
|
|
- return s.strmCln != nil && !s.strmCln.isStopped()
|
|
|
+func (p *peer) hasStreamClient() bool {
|
|
|
+ return p.strmCln != nil && !p.strmCln.isStopped()
|
|
|
}
|
|
|
|
|
|
-func (s *sender) initStream(from, to types.ID, term uint64) {
|
|
|
- strmCln := newStreamClient(from, to, term, s.p)
|
|
|
- s.mu.Lock()
|
|
|
- u := s.u
|
|
|
- s.mu.Unlock()
|
|
|
- if err := strmCln.start(s.tr, u, s.cid); err != nil {
|
|
|
+func (p *peer) initStream(from, to types.ID, term uint64) {
|
|
|
+ strmCln := newStreamClient(from, to, term, p.r)
|
|
|
+ p.mu.Lock()
|
|
|
+ u := p.u
|
|
|
+ p.mu.Unlock()
|
|
|
+ if err := strmCln.start(p.tr, u, p.cid); err != nil {
|
|
|
log.Printf("rafthttp: start stream client error: %v", err)
|
|
|
return
|
|
|
}
|
|
|
- s.strmCln = strmCln
|
|
|
+ p.strmCln = strmCln
|
|
|
}
|
|
|
|
|
|
-func (s *sender) tryStream(m raftpb.Message) bool {
|
|
|
- s.strmSrvMu.Lock()
|
|
|
- defer s.strmSrvMu.Unlock()
|
|
|
- if s.strmSrv == nil || m.Term != s.strmSrv.term {
|
|
|
+func (p *peer) tryStream(m raftpb.Message) bool {
|
|
|
+ p.strmSrvMu.Lock()
|
|
|
+ defer p.strmSrvMu.Unlock()
|
|
|
+ if p.strmSrv == nil || m.Term != p.strmSrv.term {
|
|
|
return false
|
|
|
}
|
|
|
- if err := s.strmSrv.send(m.Entries); err != nil {
|
|
|
+ if err := p.strmSrv.send(m.Entries); err != nil {
|
|
|
log.Printf("rafthttp: send stream message error: %v", err)
|
|
|
- s.strmSrv.stop()
|
|
|
- s.strmSrv = nil
|
|
|
+ p.strmSrv.stop()
|
|
|
+ p.strmSrv = nil
|
|
|
return false
|
|
|
}
|
|
|
return true
|
|
|
}
|
|
|
|
|
|
-func (s *sender) handle() {
|
|
|
- defer s.wg.Done()
|
|
|
- for m := range s.q {
|
|
|
+func (p *peer) handle() {
|
|
|
+ defer p.wg.Done()
|
|
|
+ for m := range p.q {
|
|
|
start := time.Now()
|
|
|
- err := s.post(pbutil.MustMarshal(m))
|
|
|
+ err := p.post(pbutil.MustMarshal(m))
|
|
|
end := time.Now()
|
|
|
|
|
|
- s.mu.Lock()
|
|
|
+ p.mu.Lock()
|
|
|
if err != nil {
|
|
|
- if s.errored == nil || s.errored.Error() != err.Error() {
|
|
|
- log.Printf("sender: error posting to %s: %v", s.id, err)
|
|
|
- s.errored = err
|
|
|
+ if p.errored == nil || p.errored.Error() != err.Error() {
|
|
|
+ log.Printf("sender: error posting to %s: %v", p.id, err)
|
|
|
+ p.errored = err
|
|
|
}
|
|
|
- if s.active {
|
|
|
- log.Printf("sender: the connection with %s becomes inactive", s.id)
|
|
|
- s.active = false
|
|
|
+ if p.active {
|
|
|
+ log.Printf("sender: the connection with %s becomes inactive", p.id)
|
|
|
+ p.active = false
|
|
|
}
|
|
|
if m.Type == raftpb.MsgApp {
|
|
|
- s.fs.Fail()
|
|
|
+ p.fs.Fail()
|
|
|
}
|
|
|
} else {
|
|
|
- if !s.active {
|
|
|
- log.Printf("sender: the connection with %s becomes active", s.id)
|
|
|
- s.active = true
|
|
|
- s.errored = nil
|
|
|
+ if !p.active {
|
|
|
+ log.Printf("sender: the connection with %s becomes active", p.id)
|
|
|
+ p.active = true
|
|
|
+ p.errored = nil
|
|
|
}
|
|
|
if m.Type == raftpb.MsgApp {
|
|
|
- s.fs.Succ(end.Sub(start))
|
|
|
+ p.fs.Succ(end.Sub(start))
|
|
|
}
|
|
|
}
|
|
|
- s.mu.Unlock()
|
|
|
+ p.mu.Unlock()
|
|
|
}
|
|
|
}
|
|
|
|
|
|
// post POSTs a data payload to a url. Returns nil if the POST succeeds,
|
|
|
// error on any failure.
|
|
|
-func (s *sender) post(data []byte) error {
|
|
|
- s.mu.RLock()
|
|
|
- req, err := http.NewRequest("POST", s.u, bytes.NewBuffer(data))
|
|
|
- s.mu.RUnlock()
|
|
|
+func (p *peer) post(data []byte) error {
|
|
|
+ p.mu.RLock()
|
|
|
+ req, err := http.NewRequest("POST", p.u, bytes.NewBuffer(data))
|
|
|
+ p.mu.RUnlock()
|
|
|
if err != nil {
|
|
|
return err
|
|
|
}
|
|
|
req.Header.Set("Content-Type", "application/protobuf")
|
|
|
- req.Header.Set("X-Etcd-Cluster-ID", s.cid.String())
|
|
|
- resp, err := s.tr.RoundTrip(req)
|
|
|
+ req.Header.Set("X-Etcd-Cluster-ID", p.cid.String())
|
|
|
+ resp, err := p.tr.RoundTrip(req)
|
|
|
if err != nil {
|
|
|
return err
|
|
|
}
|
|
|
@@ -318,14 +307,14 @@ func (s *sender) post(data []byte) error {
|
|
|
switch resp.StatusCode {
|
|
|
case http.StatusPreconditionFailed:
|
|
|
select {
|
|
|
- case s.shouldstop <- struct{}{}:
|
|
|
+ case p.shouldstop <- struct{}{}:
|
|
|
default:
|
|
|
}
|
|
|
- log.Printf("rafthttp: conflicting cluster ID with the target cluster (%s != %s)", resp.Header.Get("X-Etcd-Cluster-ID"), s.cid)
|
|
|
+ log.Printf("rafthttp: conflicting cluster ID with the target cluster (%s != %s)", resp.Header.Get("X-Etcd-Cluster-ID"), p.cid)
|
|
|
return nil
|
|
|
case http.StatusForbidden:
|
|
|
select {
|
|
|
- case s.shouldstop <- struct{}{}:
|
|
|
+ case p.shouldstop <- struct{}{}:
|
|
|
default:
|
|
|
}
|
|
|
log.Println("rafthttp: this member has been permanently removed from the cluster")
|