Re: [lwip-users] MEMP_NUM_TCP_SEG setting

[Siva Velusamy]

> This loop happens to be circular! Should I assume that I'm calling
> some lwip tcp function from a context that has not been properly
> guarded with mutexes?

Yes.  Almost certainly you have a race between two threads, or perhaps
between your application thread and the timers, that you haven't
protected sufficiently.  As a result, they will both occasionally try
and modify a queue at the same time, and it will become corrupt,
containing a circular reference for example as you point out.

> 2. The second issue that I see (if I don't run into issue 1), is that
> the TX performance keeps dropping slowly. Over 10 minutes it drops
> about 10 Mbps or so..starting from lets say 55 Mbps to 45 Mbps, and
> continuing to drop. Any suggestions on what to look for here? Could it
> be that my timer interval is not quite right?

Firstly I would make sure you've solved issue 1.  I wouldn't have
thought the timer interval was to blame for dropping performance.  I
think someone else has reported something similar recently but couldn't
get to the bottom of it and I can't remember the details now.
Personally I think it is most likely to be a memory leak or something
like that leading to a shortage of packet buffers or something like
that.  A corrupt queue could be to blame though (if it just kept getting
longer and longer, and we have to iterate it each time a packet is
received for example).

Thanks a ton, Kieran. You were absolutely correct on both counts.

My application was structured as follows:

while (1) {
        ethernetif_input(netif);
        transfer_data()
}


What was happening was that ethernetif_input (based on the template in netif/ethernetif.c), takes input data from a queue and sends it to ip_input. When this is happening, a timer interrupt occurred, that called tcp_output within tcp_slowtmr.

I put in a big global lock around "ethernetif_input" in the above while loop, and now I see that everything works fine - consistent speeds and no degradation over time.

What I have not yet understood is that the template ethernetif_input doesn't talk about protecting calls to ip/tcp functions from within that function. Should calls to ip_input be protected via a mutex? It seemed to me that most of the critical functions inside lwip were protected, so I assumed that a timer interrupt could happen when lwip is processing tcp input. However this does not seem to be the case. Do you have any thoughts on that?

Thanks,
-Siva