Re: [lwip-users] Race condition with lwIP-1.3.2 in PXELINUX-4.10-pre16+ on VMware/KVM platforms

[address@hidden]

Gene Cumm wrote:
> For the past several months, Syslinux has been trying to integrate 
> lwIP into PXELINUX.  At the moment, there appears to be a bug of some 
> sort when using VMware platforms or some KVM platforms.  Over the last 
> few weeks, I've been working with hpa and trying to add additional 
> debug checks to trace the program flow.  I've noticed that it is able 
> to send a TCP segment but then switches to receive the incoming 
> segment before being added to unacked by the output thread.  RTT from 
> SYN to SYN/ACK is 0.023 ms per Wireshark on the host.  The output 
> thread does eventually return but doesn't do so in time before the 
> socket is reset.
>
> Examining top-of-git, I don't see anything that would change this 
> behavior and I have attempted to integrate this version in my copy for 
> now to hopefully make it easier to work with lwIP.  I see at least 
> three possible solutions:
> 1) Don't allow the interface driver to switch threads.
> 2) tcp_process() should handle the condition of the packet being in 
> unsent.
> 3) Allow the output thread to wrap up a packet, queuing it into 
> unacked, before tcp_process() searches unacked.
>
> Any suggestions?
That all sounds like you allow multiple threads to use lwIP at the same 
time, which is not supported. Could it be that you call the input 
function from one thread while calling the write (or output) function 
from another thread? If so, you have to change your setup to pass RX 
packets into the tcpip_thread (using tcpip_input() as input function 
passed to netif_add()).

In general, lwIP does not include threading protection when modifying 
internal state variables. It relies on the API layer (and/or the lwIP 
port to your "hardware" or OS) to either process everything in one 
thread or take care of a global lock before using lwIP resources.

Simon