Weird behaviour when channelling go-capnproto2 over websocket

[Vaelatern]

My setup is a little convoluted. I use capnproto rpc over websockets (to simplify deployment) between my programs. However, the websocket server keeps failing to read a header somehow and crashes.

rpc: writing return: failed to write msg: websocket closed: failed to read header: EOF

and on the Dialer side:

rpc: writing finish: failed to write msg: context canceled

Sometimes both ends cancel their contexts at the same time. I'm not sure why.

[Vaelatern]

Turns out this happens when no headers are sent at all, I'm trying to send a header with my requests now to see if that changes something.

[Vaelatern]

I do believe this is a bug... this is using the net.Conn wrapper, and I can't think of a reason an expected header sent from the websocket library to the websocket library would be absent.

[nhooyr]

Hi I’m on vacation right now. I don’t have my laptop with me so I cannot help debug. I’ll be back in 2 weeksish and will look into this then. In the mean time def post anything you find.

I’m not sure why the lib sometimes sends no headers. Can you try using the default http client without any proxy or anything, just connect both sides over localhost and lmk if you can reproduce. It sounds like a bug in net/http.

[Vaelatern]

These are actually "net/http" and localhost already. I'm going to try adding debug prints to see exactly what the stack trace actually is.

[nhooyr]

Could you post a reproducible example if possible? Does this happen every time?

[Vaelatern]

Yes, every time and reliably, very shortly after program start. I'm trying to minimize the example as well.

[Vaelatern]

This is what readHeader() returns before the error shows up. I'm going to research more.

fin: true
rsv1: false
rsv2: false
rsv3: false
opcode: 2
payload length: 48
masked: true
maskKey: [184, 234, 196, 219]

Edit: I was wrong, that's the initial message when all is well. Of course readHeader doesn't reply anything when it's got an EOF from the reader.

[Vaelatern]

https://gist.github.com/Vaelatern/dd8aae465e3807af5d189143377f5008 A minimum broken example.

[nhooyr]

Found it. The net conn wrapper needs to stop the timer if the time passed is the zero Value. Right now a zero value deadline cancels the ctx.

If you make a PR I’m happy to merge.

[Vaelatern]

Wonderful! I'll prepare a PR and I expect to write that in the next 12 hours. Thank you!

Edit: Looks like I overestimated my understanding. As best I read it, the wrapper never gets a time as an argument.

[Vaelatern]

It ever calls SetDeadline? I'll find out.