Hi Jürgen,
> The patch below we already use since 2006. > We had problems with the TCP/IP stack directly after booting because it > already received packets before being initialized completely. > Here our CVS log: > > move > splx(s); > from line 682 to line 739 > so that tcp/ip stack has started up before the ethernet driver is released > (else the not-ready stack will already receive packets) > > Do you agree this is a valid patch? that is hard to tell - probably it is not necessary to lock anything here, especially because the in_control() does many things with the in_ifaddrhead list, without any lock. It just locks when it adds/removes entries from that list, which is really the bare minimum. I had similar problems however: an ICMP packet arriving while there was zero interface addresses registered. => NULL pointer access in sys/netinet/ip_icmp.c see the hunk at ip_icmp.c, lines 651-656. This means that the packet may arrive well before in_ifinit is called. Even before line 293 of in.c enters the address to the list: this place is locked with splnet however. That null pointer access became only apparent after I changed the MMU to exclude page zero, because usually the arm processor would silently allow read accesses to the ISR table. See my patch at http://bugs.ecos.sourceware.org/show_bug.cgi?id=1001649 for details on "MMU: added Access Protection against NULL-Pointer accesses". Before I had invented that patch it was really hard to find such hideous bugs. Regards Bernd Edlinger
