On 3/20/08, Claudio Jeker <[EMAIL PROTECTED]> wrote: > On Wed, Mar 19, 2008 at 04:59:40PM -0700, Matthew Dempsky wrote: > > OpenBSD's currently limited to using interfaces with an index < 32 for > > multicast, and on one of my machines I created and destroyed enough > > virtual interfaces during experimentation that some of the interfaces > > currently in use and that I would like to route multicast traffic to > > have indexes >= 32. > > That's a bug in the multicast code that needs fixing.
Yeah, but in the mean-time I need a workaround. :-) > > Fortunately, I tested this idea first, because it actually leads to a > > kernel panic. :-) > > I somewhat expected that. Nobody ever expected the ifindex to wrap. Looks like at least markus@ did. ;-) > See attached diff which should help finding free slots (at least it helped > in my case). It will only reuse the last if_index and not previous free > slots. See if.c:if_attachsetup() use of the static if_index. Cool. > The main issue with this diff is that ifindexes are reused and some > userland apps (mainly SNMP) require that the ifindex is unique and not > reused. I don't care about SNMP but I wanted to warn you about that. I prefer SNMP giving bad results in extreme cases to a kernel panic. :-) Also, I suspect this might be necessary for correctness when IPv6 is used. Since ifindex2ifnet is used in the IPv6 code for choosing which interface a link-scoped address points to, and the ifnet struct is free()'d when the interface detaches, it looks like running (e.g.) 'ping6 fe80:100::1' after interface 256 has been destroyed would lead to undefined memory accesses. (Though I'm still learning how the networking stack works; maybe somewhere higher is able to identify that if_index is invalid even when ifindex2ifnet[if_index] != NULL.)
