websocket transport disconnected occasionally

[tomnotcat]

I use videoroom plugin with websocket transport, most time janus works well, but occasionally websocket will disconnect, the janus log file output:
[libwebsockets][ERR] ****** 0x7fe6cc0026e0: Sending new 187 (<81>~), pending truncated ...
It's illegal to do an lws_write outside of
the writable callback: fix your code

[lminiero]

That's exactly what we do, if you even bothered to check the code. But it you think that's not the case, I will consider "fixing my code" when you say "please".

[tomnotcat]

I'm sorry for your misunderstand, “fix your code” is part of libwebsockets log, Not added by me

[lminiero]

Sorry about that, I shouldn't read mails early in the morning... 😊
It's weird we get that error, though, because we definitely only do writes from the lws callbacks. I'll reopen and check on Monday.

[tomnotcat]

I'm appreciated for all your guys amazing work!

I googled and found this post, warmcat/libwebsockets#852

It seems we should test lws_send_pipe_choked before call lws_write when receive LWS_CALLBACK_SERVER_WRITEABLE

But I'm not sure whether this will fix the bug, I'm going to test it for a while.

[lminiero]

Ah, thanks for the pointer! This comment makes sense indeed:

However two or more lws_write() inside the writable callback can also cause it, if the first one was unable to send everything.

So you should marshall what you want to write into a buffer and write it once (on lws_write() call) per writable callback.

as we do have a mechanism to send pending data. I'll have to check if we try to send more stuff in sequence, which is likely. I'll keep you posted.

[lminiero]

Mh, no, that's not our case: I verified and in the LWS_CALLBACK_CLIENT_WRITEABLE callback we only do a singe lws_write and then call lws_callback_on_writable again before returning, waiting for the callback to be invoked again. This means we never call lws_write more than once in the same callback.

[tomnotcat]

I tested for some days in production environment,
After add lws_send_pipe_choked check, this error never happen again.
I created a pull request that fix this bug:
#2107

[lminiero]

Closing as the fix was merged.