Re: [ECOS] uipc_socket.c (and cyg_tcp_maxidle)

[Hans Hübner <hans at huebner dot org>]

On Wed, 15 Jun 2005, Will Lentz wrote:

> I may have found a potential bug in
> packages/net/bsd_tcpip/current/src/sys/kern/uipc_socket.c (or I may be
> completely wrong :-).
>
> At the end of sodealloc(), the following code exists:
>       zfreei(so->so_zone, so);
>       wakeup(so->so_zone);
> The problem is that zfreei() changes so->so_zone.  Shouldn't wakeup() be
> done on the original so->so_zone?  I only noticed this problem by:
> 1- while(1) {
>   sock = socket( AF_INET, SOCK_STREAM, IPPROTO_TCP );
>   connect( sock, ... );
>   close( sock );
>   }
>   Eventually this pauses in socket() (in cyg_tsleep()) when you run out
> of eCos sockets.
>
> 2- After 2*MSL or so, cyg_wakeup() gets called with chan == 0x0.  Why?
> The zfreei() call in sodealloc() changes so->so_zone to 0 before the
> wakeup() call.

If I read the source correctly, zfreei() is a function, not a macro, and it 
is passed the zone argument by value (as is default with C).  Thus, zfreei() 
cannot change the so structure itself and the value of so->so_zone will be 
unmodified after the call.  sleep() and wakeup() take void* as argument and do 
not interpret the value, so to them it is not of concern whether the passed 
pointer is actually valid.

I looked into this because I had some problems with eCos and an embedded web 
server.  Seemingly, the server worked fine but stopped serving after some 10 
requests had been processed.  Looking into this, I found that the number of 
sockets in our default configuration was very low (16).  As TCP sockets stay 
allocated after the connection has closed until the tcp_maxidle period has 
expired, no new connections could be accepted until that period (which is in 
the minutes range in the default configuration) passed.

Not having looked at the TCP specification, I decided that reducing the 
tcp_maxidle parameter to a very low value combined with an increase in the 
number of sockets would help me out:

  extern int cyg_tcp_maxidle;
  cyg_tcp_maxidle = 200;

I am still not completely satisfied with the solution, because I don't really 
know whether this change will result in bad interactions with other TCP 
network partners.  Also, I have found the numbers reported for TCP 
accept/established/close do not correspond and I suspect that there are in 
fact other bugs which get triggered by hitting the socket limit very often.

I hope this is useful to some.

Regards,
Hans

--
Before posting, please read the FAQ: http://ecos.sourceware.org/fom/ecos
and search the list archive: http://ecos.sourceware.org/ml/ecos-discuss