func waitFor(cl *doozer.Conn, path string) {
var rev int64
for {
ev, err := cl.Wait(path, rev)
if err != nil {
panic(err)
}
if ev.IsSet() && len(ev.Body) > 0 {
break
}
rev = ev.Rev + 1
}
}
func activate(st *store.Store, self string, c *doozer.Conn) int64 {
rev, _ := st.Snap()
for _, base := range store.Getdir(st, calDir) {
p := calDir + "/" + base
v, rev := st.Get(p)
if rev != store.Dir && v[0] == "" {
seqn, err := c.Set(p, rev, []byte(self))
if err != nil {
log.Println(err)
continue
}
return seqn
}
}
for {
ch, err := st.Wait(calGlob, rev+1)
if err != nil {
panic(err)
}
ev, ok := <-ch
if !ok {
panic(io.EOF)
}
rev = ev.Rev
// TODO ev.IsEmpty()
if ev.IsSet() && ev.Body == "" {
seqn, err := c.Set(ev.Path, ev.Rev, []byte(self))
if err != nil {
log.Println(err)
continue
}
return seqn
} else if ev.IsSet() && ev.Body == self {
return ev.Seqn
}
}
return 0
}
func follow(st *store.Store, cl *doozer.Conn, rev int64, stop chan bool) {
for {
ev, err := cl.Wait("/**", rev)
if err != nil {
panic(err)
}
// store.Clobber is okay here because the event
// has already passed through another store
mut := store.MustEncodeSet(ev.Path, string(ev.Body), store.Clobber)
st.Ops <- store.Op{ev.Rev, mut}
rev = ev.Rev + 1
select {
case <-stop:
return
default:
}
}
}
// Elect chooses a seed node, and returns a connection to a cal.
// If this process is the seed, returns nil.
func elect(name, id, laddr string, b *doozer.Conn) *doozer.Conn {
// advertise our presence, since we might become a cal
nspath := "/ctl/ns/" + name + "/" + id
r, err := b.Set(nspath, 0, []byte(laddr))
if err != nil {
panic(err)
}
// fight to be the seed
_, err = b.Set("/ctl/boot/"+name, 0, []byte(id))
if err, ok := err.(*doozer.Error); ok && err.Err == doozer.ErrOldRev {
// we lost, lookup addresses again
cl := lookupAndAttach(b, name)
if cl == nil {
panic("failed to attach after losing election")
}
// also delete our entry, since we're not officially a cal yet.
// it gets set again in peer.Main when we become a cal.
err := b.Del(nspath, r)
if err != nil {
panic(err)
}
return cl
} else if err != nil {
panic(err)
}
return nil // we are the seed node -- don't attach
}
// Find possible addresses for cluster named name.
func lookup(b *doozer.Conn, name string) (as []string) {
rev, err := b.Rev()
if err != nil {
panic(err)
}
path := "/ctl/ns/" + name
names, err := b.Getdir(path, rev, 0, -1)
if err == doozer.ErrNoEnt {
return nil
} else if err, ok := err.(*doozer.Error); ok && err.Err == doozer.ErrNoEnt {
return nil
} else if err != nil {
panic(err)
}
path += "/"
for _, name := range names {
body, _, err := b.Get(path+name, &rev)
if err != nil {
panic(err)
}
as = append(as, string(body))
}
return as
}
func Main(clusterName, self, buri, rwsk, rosk string, cl *doozer.Conn, udpConn *net.UDPConn, listener, webListener net.Listener, pulseInterval, fillDelay, kickTimeout int64, hi int64) {
listenAddr := listener.Addr().String()
canWrite := make(chan bool, 1)
in := make(chan consensus.Packet, 50)
out := make(chan consensus.Packet, 50)
st := store.New()
pr := &proposer{
seqns: make(chan int64, alpha),
props: make(chan *consensus.Prop),
st: st,
}
calSrv := func(start int64) {
go gc.Pulse(self, st.Seqns, pr, pulseInterval)
go gc.Clean(st, hi, time.Tick(1e9))
var m consensus.Manager
m.Self = self
m.DefRev = start
m.Alpha = alpha
m.In = in
m.Out = out
m.Ops = st.Ops
m.PSeqn = pr.seqns
m.Props = pr.props
m.TFill = fillDelay
m.Store = st
m.Ticker = time.Tick(10e6)
go m.Run()
}
hostname, err := os.Hostname()
if err != nil {
hostname = "unknown"
}
if cl == nil { // we are the only node in a new cluster
set(st, "/ctl/name", clusterName, store.Missing)
set(st, "/ctl/node/"+self+"/addr", listenAddr, store.Missing)
set(st, "/ctl/node/"+self+"/hostname", hostname, store.Missing)
set(st, "/ctl/node/"+self+"/version", Version, store.Missing)
set(st, "/ctl/cal/0", self, store.Missing)
if buri == "" {
set(st, "/ctl/ns/"+clusterName+"/"+self, listenAddr, store.Missing)
}
calSrv(<-st.Seqns)
// Skip ahead alpha steps so that the registrar can provide a
// meaningful cluster.
for i := 0; i < alpha; i++ {
st.Ops <- store.Op{1 + <-st.Seqns, store.Nop}
}
canWrite <- true
go setReady(pr, self)
} else {
setC(cl, "/ctl/node/"+self+"/addr", listenAddr, store.Clobber)
setC(cl, "/ctl/node/"+self+"/hostname", hostname, store.Clobber)
setC(cl, "/ctl/node/"+self+"/version", Version, store.Clobber)
rev, err := cl.Rev()
if err != nil {
panic(err)
}
stop := make(chan bool, 1)
go follow(st, cl, rev+1, stop)
errs := make(chan error)
go func() {
e, ok := <-errs
if ok {
panic(e)
}
}()
doozer.Walk(cl, rev, "/", cloner{st.Ops, cl, rev}, errs)
close(errs)
st.Flush()
ch, err := st.Wait(store.Any, rev+1)
if err == nil {
<-ch
}
go func() {
n := activate(st, self, cl)
calSrv(n)
advanceUntil(cl, st.Seqns, n+alpha)
stop <- true
canWrite <- true
go setReady(pr, self)
if buri != "" {
b, err := doozer.DialUri(buri, "")
if err != nil {
panic(err)
}
setC(
b,
"/ctl/ns/"+clusterName+"/"+self,
listenAddr,
store.Missing,
)
}
}()
}
shun := make(chan string, 3) // sufficient for a cluster of 7
go member.Clean(shun, st, pr)
go server.ListenAndServe(listener, canWrite, st, pr, rwsk, rosk)
if rwsk == "" && rosk == "" && webListener != nil {
web.Store = st
web.ClusterName = clusterName
go web.Serve(webListener)
}
go func() {
for p := range out {
n, err := udpConn.WriteTo(p.Data, p.Addr)
if err != nil {
log.Println(err)
continue
}
if n != len(p.Data) {
log.Println("packet len too long:", len(p.Data))
continue
}
}
}()
selfAddr, ok := udpConn.LocalAddr().(*net.UDPAddr)
if !ok {
panic("no UDP addr")
}
lv := liveness{
timeout: kickTimeout,
ival: kickTimeout / 2,
self: selfAddr,
shun: shun,
}
for {
t := time.Now().UnixNano()
buf := make([]byte, maxUDPLen)
n, addr, err := udpConn.ReadFromUDP(buf)
if err == syscall.EINVAL {
return
}
if err != nil {
log.Println(err)
continue
}
buf = buf[:n]
lv.mark(addr, t)
lv.check(t)
in <- consensus.Packet{addr, buf}
}
}
func setC(cl *doozer.Conn, path, body string, rev int64) {
_, err := cl.Set(path, rev, []byte(body))
if err != nil {
panic(err)
}
}
func advanceUntil(cl *doozer.Conn, ver <-chan int64, done int64) {
for <-ver < done {
cl.Nop()
}
}