Read error after EOF on Windows, but not on Unix

[klutzy]

If user tries to read tcp socket after EOF, uv_read_start() returns UV_ENOTCONN on Windows, but not on Unix. This may make inconsistency on error handling.

On Windows, src/win/stream.c checks UV_HANDLE_READABLE flag, which is turned off at EOF.

On Unix, however, it does not return error, instead read_cb is called normally (with zero-length buffer). There is UV_STREAM_READABLE flag but it is not used for the purpose I think.

[saghul]

This comment in the code mentions that case exactly: https://github.com/joyent/libuv/blob/master/src/unix/stream.c#L1328

[klutzy]

Oh sorry! I was only reading codes not comments. My bad...
I found this behavior while fixing Rust; it has "read EOF twice" test which gets error on windows therefore fails, but it seems that it should fail if tcp state tracking is implemented in future.

[saghul]

No worries! I just added the comment "for the record" :-)

[piscisaureus]

A patch to make this more consistent would be welcome.
I think the unix behaviour is preferable here.

[txdv]

Why callback with 0 length? Instead of returning immediately since we already know that we are disconnected?

The state of the socket is known immediately, we should return the error code immediately as well.

[saghul]

We need a consistent model for error reporting, there is already an issue open about it. The consensus seems to be to report invalid values right away and everything else as a callback. This would make things more consistent with this model.

[txdv]

That is exactly what I am proposing, or is your argument directed at @piscisaureus ?

[saghul]

No, you are proposing to return the error immeditely, I'm agreeing with @piscisaureus.

[txdv]

So the consensus is to report invalid values right away, but you are agreeing with @piscisaureus to mimic the unix behavior which calls the read_cb with an empty buffer instead of returning an error code immediately?

I'm seriously confused right now.

[saghul]

By invalid values I mean parameter sanity checks, for which we usually return EINVAL.

[saghul]

#829

[txdv]

Only return errors for bad arguments (EINVAL) with a compile-time option to turn those into asserts for easier debugging.

We are already returning ENOMEM immediately... And we have been going the direction of returning error codes immediately if there is no need for them do defer (consistency between platforms) and now you dig up a 1 year old issue with a short description which advises us to move against what he have been already doing.

[saghul]

This needs to be well thought out, and that's not going to happen before 1.0.0. We need to go though all cases. Some things are obvious: invalid parameter == EINVAL immediately, but some others not necessarily (EMFILE, scoket creation error, and so on).

[txdv]

Ok, can we start discussing this before 1.0.0?

Also, where do we want to discuss this? In bnoordhuis issue you referenced?

[saghul]

I don't have the time to discuss this right now. I'd like to discuss it as part of a multi-chapter RFC for 2.0.0., which I'll publish when we are done with 1.0.0.

[txdv]

Discussing this shouldn't be the time intensive task... But whatever.

What happened to that other RFC? Can I start implementing it? Is there anything else but my pull request for the write_after_connect standing in front of 1.0.0?

[saghul]

Discussing this shouldn't be the time intensive task... But whatever.

It is for me.

What happened to that other RFC? Can I start implementing it? Is there anything else but my pull request for the write_after_connect standing in front of 1.0.0?

I told you before, I will implement that one. I already started, but well, things got in between. Also I needed to get a touch of reality before finishing up the rest of the RFC. Right now it's a bunch of notes which I need to tidy up.

At this point I'm not sure write_after_connect will land before 1.0.0.

[txdv]

Can I write an RFC for the return codes then?

[saghul]

Sure.
On Oct 16, 2014 7:52 PM, "Andrius Bentkus" notifications@github.com wrote:

Can I write an RFC for the return codes then?

—
Reply to this email directly or view it on GitHub
#948 (comment).