Subject: Re: proper use of non-blocking read

Re: proper use of non-blocking read

From: Daniel Stenberg <>
Date: Wed, 10 Apr 2013 23:28:03 +0200 (CEST)

On Wed, 10 Apr 2013, Eric Frias wrote:

> The main problem I'm battling is what happens when I need to do a read on a
> channel, and I don't know if there's going to be any data there to read.
> If there is, I want to read it. If there's no data, I want to be able to do
> other stuff like write to the channel or read or write to another channel.
> Seems pretty normal.

But stuff you'd do on another channel isn't the problem here, right? Just
trying to check that I follow you properly.

> So let's say I call libssh2_channel_read_ex() and it returns
> LIBSSH2_ERROR_EAGAIN. My problem is that I can't tell what this means.
> It could mean that:
> * there was no data on that channel, and it's safe to go write some data
> or service another channel. or,
> * in the process of doing the read, libssh2 wandered into
> _libssh2_channel_receive_window_adjust, sending a message to the ssh
> server. While sending the adjust message, the TCP send buffer filled
> up, and the OS level send(2) call returned EAGAIN

Yes. EAGAIN is a bit restrictive in that it only says libssh2 couldn't deliver
any data to you and you need to call libssh2_session_block_directions() to
figure out which direction of the communication that caused EAGAIN.

> Of course, this second case is the one that is giving me problems. Although
> I haven't seen it stated explicitly anywhere, it looks like if I get
> LIBSSH2_ERROR_EAGAIN due to a failed write, the only thing I can safely do
> with libssh2 is wait and call the same libssh2_channel_read_ex() function
> again with the same arguments until it manages to finish the write. If I
> were to make some other call to libssh2 which could result in network
> traffic, libssh2 will give me a slap on the wrist (in the form of a
> LIBSSH2_ERROR_BAD_USE) and then I'm in trouble.

This is a flaw in the current implementation and I would really like us to fix
it. I believe someone else also mentioned this limitation a short while ago. I
guess you're being hit by a check in the transport layer like in
send_existing() ?

> Am I missing something that would let me use the public API safely? Right
> now I'm playing around with workarounds like only allowing my code to switch
> and call another function if session->packet.olen == 0 after an
> LIBSSH2_ERROR_EAGAIN, but that's ugly and I don't have much confidence that
> I'm on the right track.

I think the proper way to implement this would be sure that the queue for
outgoing messages on the channel is still being drained properly and in the
right order even if you try another function that would send data.

Personally I've been backlogged badly recently so I haven't been able to do
much libssh2 hacking and I don't think there's much improvement for me in the
near future either.

I think creating a good test case that repeats the situation in a fairly good
way could be a first step, then work on fixing code to work with it.

Received on 2013-04-10