[lwip-users] problems with LWIP on Xilinx ML403

[Mike Caplinger]

I've been trying to do large bulk data transfers using the sockets API 
on an ML403 board using the version of LWIP bundled with EDK 8.2i.  I 
won't bore you with all the total stalls of LWIP I've experienced, but I 
have noticed that there seems to be a huge bug in the Xilinx 
implementation of sys_arch.  The function sys_arch_sem_wait is supposed 
to return the amount of time it had to wait to get the semaphore, or 
SYS_ARCH_TIMEOUT if it failed to get it.  But the v2pro port version in 
EDK 8.2i simply returns the timeout immediately if it got the semaphore, 
regardless how long it had to wait, if at all.  This causes the TCP 
thread, which is using a 250 msec timeout handler to call tcp_tmr,  to 
go completely nuts during a transfer, since tcp_tmr ends up getting 
called at extremely high rates.  Retransmits and timeouts happen too 
fast, RTO is completely misestimated (I've seen it stuck up at 32767) etc.

Maybe this isn't an issue when using the raw interface, and maybe this 
is why nobody interested in performance uses the socket interface.

Has anyone else encountered this, or am I crazy?  Maybe later versions 
of EDK addressed this?