Maybe a bug in RaisePendingStreamEvents

[fomy]

    if (socket->epoll & MTCP_EPOLLOUT) {
        struct tcp_send_vars *sndvar = stream->sndvar;
        if (!sndvar->sndbuf || 
                (sndvar->sndbuf && sndvar->sndbuf->len < sndvar->snd_wnd)) {
            if (!(socket->events & MTCP_EPOLLOUT)) {
                TRACE_EPOLL("Socket %d: Adding write event\n", socket->id);
                AddEpollEvent(ep, USR_SHADOW_EVENT_QUEUE, socket, MTCP_EPOLLOUT);
            }
        }
    }

Hi,

I notice that only when sndvar->sndbuf->len < sndvar->snd_wnd, mtcp will raise an event. Should this condition be sndvar->snd_wnd > 0?

Thanks

[eunyoung14]

Hi, this block checks whether there is send buffer available.

1. mTCP's sending buffer is created on demand. If sndbuf is null, it means that there was no prior mtcp_write call and we can create new send buffer in the future.
2. If sending buffer is already present, it checks whether there is enough room in the buffer. If the length of payloads in the buffer is less than the send window, it means that the app can send more data.

If one of those two conditions is satisfied, mTCP raises an MTCP_EPOLLOUT event so that the application can send. If sndvar->snd_wnd is 0, I think the application shouldn't send anything.

[fomy]

Sure.

But your code is

sndvar->sndbuf->len < sndvar->snd_wnd

rather than sndvar->snd_wnd > 0

[fomy]

Here is the code where snd_wnd is calculated:

sndvar->snd_wnd = sndvar->sndbuf->size - sndvar->sndbuf->len;

sndvar->snd_wnd indicates how many room available in the buf, and sndbuf->len indicates how many bytes are already in the buf.

sndvar->sndbuf->len < sndvar->snd_wnd will cause half of the buf cannot be used.

[eunyoung14]

Now it makes sense! I was confused a little before you edited the comment. I'll fix the condition to sndvar->snd_wnd > 0. Thanks for reporting.