net/rpc: race condition in Server.ServeCodec

[willfaught]

Please answer these questions before submitting your issue. Thanks!

What version of Go are you using (go version)?

It was 1.6 (on a work computer I don't have access to anymore).

What operating system and processor architecture are you using (go env)?

It was Mac OS 10.11 (on a work computer I don't have access to anymore).

What did you do?

There's a race condition in Server.ServeCodec, where the connection is closed without waiting for all calls to finish:

477         go service.call(server, sending, mtype, req, argv, replyv, codec)
478     }
479     codec.Close()

In this code, the go statement is executed in a loop, and then the codec is closed immediately after the loop ends, which in the case of the jsonrpc codec closes the connection used by the calls before they can finish, resulting in errors.

[quentinmit]

/cc @robpike @rsc @adg

I think the intent here is for the loop to block again when it loops on readRequest.

How are you managing to exit the loop?

We could of course fix this by adding a sync.WaitGroup to track the outstanding RPCs, but then you would not be able to tell from the RPC handler when the connection is closed and you should abort.

[robpike]

If this is related to #5792, as I think it is, it's really hard to fix. The read half is relatively easy but the write half is not. See also #16844, which was introduced partly in response to this problem.

[willfaught]

How are you managing to exit the loop?

By closing the connection to shut down the server gracefully.

[rsc]

CL sent.

[gopherbot]

Change https://golang.org/cl/79515 mentions this issue: net/rpc: wait for responses to be written before closing Codec

[willfaught]

Awesome! Thanks @rsc!