func maybeDecorateGRPCError(wrapped cmdFn) cmdFn {
return func(cmd *cobra.Command, args []string) error {
err := wrapped(cmd, args)
{
unwrappedErr := errors.Cause(err)
if unwrappedErr == nil {
return err
}
_, isSendError := unwrappedErr.(*roachpb.SendError)
isGRPCError := grpcutil.IsClosedConnection(unwrappedErr)
_, isNetError := unwrappedErr.(*net.OpError)
if !(isSendError || isGRPCError || isNetError) {
return err // intentionally return original to keep wrapping
}
}
format := `unable to connect or connection lost.
Please check the address and credentials such as certificates (if attempting to
communicate with a secure cluster).
%s`
return errors.Errorf(format, err)
}
}
// GRPCDial calls grpc.Dial with the options appropriate for the context.
func (ctx *Context) GRPCDial(target string, opts ...grpc.DialOption) (*grpc.ClientConn, error) {
ctx.conns.Lock()
meta, ok := ctx.conns.cache[target]
if !ok {
meta = &connMeta{}
ctx.conns.cache[target] = meta
}
ctx.conns.Unlock()
meta.Do(func() {
var dialOpt grpc.DialOption
if ctx.Insecure {
dialOpt = grpc.WithInsecure()
} else {
tlsConfig, err := ctx.GetClientTLSConfig()
if err != nil {
meta.err = err
return
}
dialOpt = grpc.WithTransportCredentials(credentials.NewTLS(tlsConfig))
}
dialOpts := make([]grpc.DialOption, 0, 2+len(opts))
dialOpts = append(dialOpts, dialOpt)
dialOpts = append(dialOpts, grpc.WithBackoffMaxDelay(maxBackoff))
dialOpts = append(dialOpts, opts...)
if log.V(1) {
log.Infof(ctx.masterCtx, "dialing %s", target)
}
meta.conn, meta.err = grpc.DialContext(ctx.masterCtx, target, dialOpts...)
if meta.err == nil {
if err := ctx.Stopper.RunTask(func() {
ctx.Stopper.RunWorker(func() {
err := ctx.runHeartbeat(meta.conn, target)
if err != nil && !grpcutil.IsClosedConnection(err) {
log.Error(ctx.masterCtx, err)
}
ctx.removeConn(target, meta)
})
}); err != nil {
meta.err = err
// removeConn and ctx's cleanup worker both lock ctx.conns. However,
// to avoid racing with meta's initialization, the cleanup worker
// blocks on meta.Do while holding ctx.conns. Invoke removeConn
// asynchronously to avoid deadlock.
go ctx.removeConn(target, meta)
}
}
})
return meta.conn, meta.err
}
func (ctx *Context) removeConnLocked(key string, meta *connMeta) {
if log.V(1) {
log.Infof(ctx.masterCtx, "closing %s", key)
}
if conn := meta.conn; conn != nil {
if err := conn.Close(); err != nil && !grpcutil.IsClosedConnection(err) {
if log.V(1) {
log.Errorf(ctx.masterCtx, "failed to close client connection: %s", err)
}
}
}
delete(ctx.conns.cache, key)
}
// start dials the remote addr and commences gossip once connected. Upon exit,
// the client is sent on the disconnected channel. This method starts client
// processing in a goroutine and returns immediately.
func (c *client) start(
g *Gossip,
disconnected chan *client,
rpcCtx *rpc.Context,
stopper *stop.Stopper,
nodeID roachpb.NodeID,
breaker *circuit.Breaker,
) {
stopper.RunWorker(func() {
ctx, cancel := context.WithCancel(c.AnnotateCtx(context.Background()))
var wg sync.WaitGroup
defer func() {
// This closes the outgoing stream, causing any attempt to send or
// receive to return an error.
//
// Note: it is still possible for incoming gossip to be processed after
// this point.
cancel()
// The stream is closed, but there may still be some incoming gossip
// being processed. Wait until that is complete to avoid racing the
// client's removal against the discovery of its remote's node ID.
wg.Wait()
disconnected <- c
}()
consecFailures := breaker.ConsecFailures()
var stream Gossip_GossipClient
if err := breaker.Call(func() error {
// Note: avoid using `grpc.WithBlock` here. This code is already
// asynchronous from the caller's perspective, so the only effect of
// `WithBlock` here is blocking shutdown - at the time of this writing,
// that ends ups up making `kv` tests take twice as long.
conn, err := rpcCtx.GRPCDial(c.addr.String())
if err != nil {
return err
}
if stream, err = NewGossipClient(conn).Gossip(ctx); err != nil {
return err
}
return c.requestGossip(g, stream)
}, 0); err != nil {
if consecFailures == 0 {
log.Warningf(ctx, "node %d: failed to start gossip client: %s", nodeID, err)
}
return
}
// Start gossiping.
log.Infof(ctx, "node %d: started gossip client to %s", nodeID, c.addr)
if err := c.gossip(ctx, g, stream, stopper, &wg); err != nil {
if !grpcutil.IsClosedConnection(err) {
g.mu.Lock()
if c.peerID != 0 {
log.Infof(ctx, "node %d: closing client to node %d (%s): %s", nodeID, c.peerID, c.addr, err)
} else {
log.Infof(ctx, "node %d: closing client to %s: %s", nodeID, c.addr, err)
}
g.mu.Unlock()
}
}
})
}
