Lingering Deadlock when Write and CloseRead called concurrently in 1.8.6

[RTann]

In the fix for the deadlock introduced in 1.8.5, we have the following:

err = c.writeFrameMu.lock(ctx)
if err != nil {
	return 0, err
}
defer c.writeFrameMu.unlock()

// If the state says a close has already been written, we wait until
// the connection is closed and return that error.
//
// However, if the frame being written is a close, that means its the close from
// the state being set so we let it go through.
c.closeMu.Lock()
wroteClose := c.wroteClose
c.closeMu.Unlock()
if wroteClose && opcode != opClose {
	select {
	case <-ctx.Done():
		return 0, ctx.Err()
	case <-c.closed:
		return 0, c.closeErr
	}
}

...

If there is a (*Conn).Write that passes the if condition, then it waits for at least one of two channels to close, while still holding the writeFrameMu lock. Since c.wroteClose is set before (*Conn).CloseRead gets to call (*Conn).writeFrame (see (*Conn).writeClose), it is possible that (*Conn).CloseRead calls (*Conn).writeFrame after (*Conn).Write gets stuck, so (*Conn).CloseRead will get stuck waiting to get writeFrameMu, and no progress can be made.

One option would be to unlock writeFrameMu immediately after passing the if condition. defer c.writeFrameMu.unlock() would need to be removed since the mutex implementation doesn't allow for unlocking twice.

[nhooyr]

I believe the if condition prevents it. It won't block on the two channels if a close is being written.

I know the code flow is confusing, it'll be cleaned up soon with #239

To be clear, I'm referring to this line: https://github.com/nhooyr/websocket/blob/02861b474d9c29660eff53a3c424d589aaf46d1e/write.go#L259

[nhooyr]

Oh never mind, I see what you're saying, the lock would be held up by the goroutine writing so we would never even get to the if statement. 🤦

[nhooyr]

Luckily, the close write has a timeout of 5s so this bug has likely had minor impact. Great catch @RTann and more the reason for me to do #239

[RTann]

Right. The goroutine who calls (*Conn).Write gets stuck inside the if while still holding the lock, and the goroutine who calls (*Conn).CloseRead gets stuck trying to obtain the lock, so it never gets to reach the if. Apologies for the confusing wording. I wanted to try to detail the chain of events I observed, which I could have worded better for sure.