Re: [lwip-users] TCP windows Full

[Kieran Mansley]

On Tue, 2011-07-12 at 08:55 +0200, brak brak2 wrote:
> > Lack of pbufs to send a packet?
> 
> How can I check pbuf availability?

The LWIP_STATS code is the easiest way to discover if you're running out
of some resource.

> I release pbuf of received package until I process whole received data.
> Then I call pbuf_free:
>               SYS_ARCH_PROTECT(lev);
>               pbuf_free(gpBufsToRelease[i]); //pointer to received data pbuf
>               SYS_ARCH_UNPROTECT(lev);
> 
> 
> >  An error called tcp_recved() (although
> > you say this is not the problem)?
> > The fact that even after a long idle
> > period the advertised window is not getting back to 4096 (its starting
> > value) suggests to me that there could well be a problem with your use
> > of tcp_recved().
> 
> I call tcp_recved for every processed pbuf in the received data chain.
> It looks like this in log:
> Package received:  1460 //total among of received data in tcp_recv
> Package received:  1460
> tcp_recved called: 202 //among of data in tcp_recved call
> tcp_recved called: 256
> tcp_recved called: 256
> tcp_recved called: 256
> Package received:  951
> tcp_recved called: 256
> tcp_recved called: 234
> tcp_recved called: 202
> tcp_recved called: 256
> tcp_recved called: 256
> tcp_recved called: 256
> tcp_recved called: 256
> tcp_recved called: 234
> tcp_recved called: 202
> tcp_recved called: 256
> tcp_recved called: 256
> tcp_recved called: 237
> 
> 
> The call of recved is surrounded with protecting code:
> 
>                       SYS_ARCH_PROTECT(lev);
>                       tcp_recved(gTcpRecAck[i].pcb, gTcpRecAck[i].length);
>                       SYS_ARCH_UNPROTECT(lev);
> 
> So no lwip core code could interrupt execution of tcp_recved.

As long as you call tcp_recved() with the correct total number of bytes,
it should behave properly.  You don't need to call it separately for
each pbuf.  You could try putting some debug print statements in
tcp_recved() to see how much data it thinks is still being used by the
application and receive queue, and compare that to what you think should
be the case.

> > For full details of when lwIP will send an explicit window update, take
> > a look at the source (start in tcp_recved()), but to summarise: if we
> > can increase the advertised window by a significant amount (e.g. 1 MSS)
> > then we do.
> 
> But it is done only in recved or is it check in one of the timer also?

Only in tcp_recved() as this is the only place where the advertised
window is increased.  If we called it from a timer later on it would
just give the same answer, as without another call to tcp_recved() there
would be no more window available.

Kieran