client:fix ttrpc send hang

[ningmingxiao]

fix:https://github.com/containerd/ttrpc/issues/174 @dmcgowan @fuweid @thaJeztah can you review my pr? thank you.

[dmcgowan]

The client context scope should not be considered during send here. Since the context is at the client scope, it would be expected that unblocking sends would be done by closing the underlying connection. The sender could be legitimately waiting on writing to the connection.

What is the caller error that would cause a hang here? The caller has timed out by something is waiting for ttrpc to finish? Seems that should be handled by the caller not ttrpc.

[ningmingxiao]

I use containerd 1.7.3 https://github.com/containerd/containerd/blob/v1.7.3/services/tasks/local.go#L340 containerd try to connect shim to get container status, it hanged 1818 minutes.

ctx timeout is 2s.

func getProcessState(ctx context.Context, p runtime.Process) (*task.Process, error) {
    ctx, cancel := timeout.WithContext(ctx, stateTimeout)
    defer cancel()

    state, err := p.State(ctx)

createStream ctx never timeout. It should check timeout in dispatch func.

func (c *Client) createStream(flags uint8, b []byte) (*stream, error) {
...
    select {
    case <-c.ctx.Done():

// NewClient creates a new ttrpc client using the given connection
func NewClient(conn net.Conn, opts ...ClientOpts) *Client {
    ctx, cancel := context.WithCancel(context.Background())

[kevpar]

As I mentioned here, this may be a duplicate of an issue we already fixed.