Re: [lwip-devel] Question regarding calling tcp_output() from tcp_input() each time

[address@hidden]

Roy Shterman wrote:
> I would like to understand if there is some hidden reason beyond 
> calling tcp_output()
> at the end of tcp_input(). we are already calling tcp_output() from 
> tcp_recved() in case it is needed
> by window size and amount of data we need to ack.

You mix up left and right edge of the window:
- from tcp_input(), we ACK the data
- from tcp_recved(), we increase the window size

Since there is no way for the stack to know exactly when the application 
will call tcp_recved(), ACKs must be sent right away to prevent the peer 
thinking the segment got lost.
And applications are free to not update the window for a very long time! 
Plus, keep in mind the minimum timer frequency is 250ms. Delaying an ACK 
for that long might reduce throughput.

> this tcp_output() from tcp_input() is causing enormous amount of empty 
> ack towards the other end point.

Of course that's not too good, but noone has provided a decent, working 
solution for this, yet. Feel free to provide one :-)
One accepted improvement are "delayed ACKs" where we only send an empty 
ACK for every 2nd segment.

Cheers,
Simon