failed to close writer: context canceled

[lukaspj]

I'm using this library to create a zero-trust tunnel into a Kubernetes cluster, when I encountered this error message:

failed to write msg: failed to close writer: failed to acquire lock: context canceled

So what happened was that when client closed the connection, I would send a "connection closed" message over the WebSocket connection, but I (mistakenly) used the request context for the write command.
This context would get cancelled somewhere around this Write call, sometimes after message was written but before it released the lock.

As a result, the lock would never get released and the websocket was no longer usable.

The solution was to create a new context for the write:

	writeCtx, _ := context.WithTimeout(context.Background(), 2*time.Second)

	err = c.Websocket.Write(writeCtx, websocket.MessageText, marshal)

And I don't think I should have used the request context to begin with, but I do think it's not optimal behaviour that the websocket can be left in a broken state by mistimed context cancel.

In any case, just thought I'd share it. Thanks for maintaining this awesome library!

[mafredri]

Hey @lukaspj, thanks for the user experience report.

You are definitely not the first! This is a very common mistake to make. There is one valid use-case for r.Context() after the connection being hijacked¹, but in general it's almost never what you want.

I see this issue is further exacerbated by the readme and examples using it incorrectly (https://github.com/coder/websocket?tab=readme-ov-file#server).

We should at least update the readmes and examples, but I'll try to think about some alternative solutions as well.

Perhaps the most obvious way would be to put a context on websocket.Conn, but defining it's behavior would be a bit tricky.

Anyway, if you have anything else to add, how you wish this had worked instead, what lead you down the wrong path, etc, please feel free to share.

¹ A valid use-case is when you're running the websocket logic in a separate goroutine, and want it to stop when the HTTP handler exits. Usually the logic runs in the HTTP handler, though, where using r.Context() is useless. (https://pkg.go.dev/net/http#Hijacker)

[lukaspj]

I think you pin-pointed it, if the readme and examples explained it, it would probably help.

Additionally, it would be nice if the function documentation would state this danger of closing the context. Something like "if the context is cancelled, it can cause a deadlock" or just "the context should be the context of the websocket not the context of the message being written". Idk if there's a nice way to write it tbh, but would be nice if it showed up on the documentation somewhere on "how to use this library".

[spikecurtis]

@mafredri I don't think a docs update alone will resolve this issue.

@lukaspj seems to be saying that if the context you pass to Write() gets canceled, it can leave the Websocket in a broken state (a lock that never gets released). While it's true that you probably don't want to use the r.Context() for calls to Write() in most cases, I would still consider it a bug if an ill-timed context timeout leaves the websocket in a broken state.

[mafredri]

@spikecurtis I was thinking about the original report, but that's a good callout. We should split the second issue into a new ticket.

There may be multiple issues in how connection closure is propagated and context cancellation is handled. There seems to be multiple code-paths depending on scenario.

I also noticed while looking at #476 that a read/write context that is cancelled actually ends up closing the connection abruptly, whereas I would expect we just give up on the write.

[maggie44]

Hey @lukaspj, thanks for the user experience report.

You are definitely not the first! This is a very common mistake to make. There is one valid use-case for r.Context() after the connection being hijacked1, but in general it's almost never what you want.

I see this issue is further exacerbated by the readme and examples using it incorrectly (https://github.com/coder/websocket?tab=readme-ov-file#server).

We should at least update the readmes and examples, but I'll try to think about some alternative solutions as well.

Perhaps the most obvious way would be to put a context on websocket.Conn, but defining it's behavior would be a bit tricky.

Anyway, if you have anything else to add, how you wish this had worked instead, what lead you down the wrong path, etc, please feel free to share.

1 A valid use-case is when you're running the websocket logic in a separate goroutine, and want it to stop when the HTTP handler exits. Usually the logic runs in the HTTP handler, though, where using r.Context() is useless. (https://pkg.go.dev/net/http#Hijacker)

The HiJacker docs referenced note:

 The original Request's Context remains valid and
	// is not canceled until the Request's ServeHTTP method
	// returns.

This sounds to me like the context is still useful after a hijack, I'm not following why "in general it's almost never what you want"? Wanting to stop when the HTTP handler exits is probably a very common use case? And not too different to having a context created in the handler function.

In the PR that has been merged (https://github.com/coder/websocket/pull/477/files), it says:

// Set the context as needed. Use of r.Context() is not recommended
// to avoid surprising behavior (see http.Hijacker).

It doesn't go on to explain why r.Context() is not recommended, it doesn't explain what "surprising behaviour" is, and the http.Hijacker seems to suggest that r.Context() is valid after a Hijack. Could you clarify what is meant here?

If a request.Context or any other context is used in Write, it shouldn't lead to a broken websocket connection. If Write only accepts context.Background, is there any value in passing a context at all?

[mafredri]

@maggie44 certainly. In my experience, a common use-case is that the websocket is managed within the http.Handler, thus when r.Context() is used, a common misconception is that the context will be cancelled when the other side disconnects, which is not the case.

In the aforementioned scenario, the r.Context() can't be utilized in any way and is mainly a dangerous gotcha.

That's not to say there aren't use-cases, one could imagine the following method on a service:

func (s *Server) handleWS(w http.ResponseWriter, r *http.Request) {
	conn, _ := websocket.Accept(s.ctx, r, nil) // <- Note, r.Context() would be wrong here because it will **never** cancel.
    s.connHandler <- conn
    <-s.ctx.Done()
}

I'd be interested to know if there's a relatively common scenario in which you would find the behavior useful?

[maggie44]

I think the behaviour described in the first post is what I would have expected:

failed to write msg: failed to close writer: failed to acquire lock: context canceled

c.Websocket.Write has used r.Context(). The websocket has been closed, all the functions have returned up the chain to the handler where the websocket was accepted which returns nil, which triggers a cancel on r.Context() ("The original Request's Context remains valid and is not canceled until the Request's ServeHTTP method returns"):

func handler(w http.ResponseWriter, r *http.Request) {
	websocket.Accept(... )
	// Doing things. 
	things(r.Context())
	// Returns nil
}

A c.Websocket.Write in the initial post is trying to write something to a connection that has closed, which we know because the handler has returned, so a context canceled error is returned through .Write. This context cancelled accurately reflects what happened, and can be handled gracefully. I.e. there isn't anyone to write the websocket message to anymore, so aborts.

Where it is problematic is in this scenario it doesn't release the lock, and it seems the Write hangs, but up until that stage it seems that r.Context() is doing what I would expect.

It sounds like r.Context() in the Hijack scenario is equivalent to:

func handler(w http.ResponseWriter, r *http.Request) {
	ctx, cancel := context.WithCancel(context.Background())
	defer cancel()

	websocket.Accept(... )
	// Doing things. 
	things(ctx)
	// Returns nil
}

As you mentioned, this isn't what is expected in terms of the user disconnecting doesn't call a cancel, but is a useful way to attach multiple functions, process, particularly goroutines to the broader user connection. One of those things is communicating to writers and readers that nobody is here anymore to hear the requests? An example might be using the Ping() functionality in this package. I might start a goroutine that loops every 10 seconds, to call a ping for keep alive. I would pass r.Context() to this so that it will keep pinging until the connection is closed.

I think long story short, the Write should probably be able to accept r.Context() and the cancel it calls without blocking the websocket. And that r.Context() would be a better context to pass down than context.Background(), even if it isn't performing as it would in a non-hijack scenario.