Re: [lwip-devel] TCP FIN problem.

[address@hidden]

Kieran Mansley wrote:
> > > Slightly confused, debugging further...
> > >       
> > I'd bet that there's a difference between "struct tcp_pcb" and "struct
> > tcp_pcb_listen" on a 64 bit build that means casting between them isn't
> > valid.  This is deliberately OK on a 32 bit build, but perhaps the
> > structure fields get a different layout with padding on a 64 bit build,
> > and so when the accept field in the pcb is referenced (after the cast in
> > tcp_accept() to struct tcp_pcb_listen)  it's writing to the rcv_nxt
> > field.
> >     
>
> Yeah, there's a bug here - the tcp_accept field isn't in the
> TCP_PCB_COMMON stuff, and so accessing it via a cast is wrong.  I'm not
> sure why we need it at all in the case of a connected PCB to be honest.
If I remember correctly, when receiving a new SYN on a listen_pcb, a 
tcp_pcb is allocated and the accept callback function pointer is copied 
into this new pcb. The actual call to the callback is then done using 
this new tcp_pcb, not the original listen_pcb. In C++, this could be 
solve by derving from the TCP_PCB_COMMON, but that's not possible here...

While this seems relatively easy to fix in this special case, it might 
still be a problem in the rest of the stack: the method of defining 
common parts in PCBs is used throughout the protocols, where it could 
lead to different structure layout as well.

Simon