On Sun, 2006-06-18 at 18:05, James Carlson wrote:
> Lars Tunkrans writes:
> > Im not sure but I have now started up bind 9.2.4 ( native )
> > and bind 9.3.2 ( gcc compiled )
> > on two different Athlon 64 based pc's using SXCR build 41.
> > After I start Named I can talk to it with rndc but as soon as I send a
> > query to it with nslookup
> > , one of Named's LWP's ends up in a tight loop and the query times out.
> > Have to "kill -9 " it, to get rid of it and free up the cpu.
>
> Sounds like CR 6440104.
Workaround: downgrade /usr/lib/dns/libisc.so.7.1.5 to an older version
(the one from snv_38, which is what I had easily accessible, works for
me. I suspect everything up to build 40 should work).
what's going on here?
the SO_TIMESTAMP/SCM_TIMESTAMP implementation introduced in snv_37 is,
unfortunately, really buggy in a bunch of ways. For one, it produces
broken ancillary data (found in struct
msghdr.msg_control/msg_controllen) which causes any reasonable ancillary
data walker to go into an infinite loop.
BIND's "libisc" library conditionally uses SO_TIMESTAMP if it was
visible in the header files available at compile time.
Versions of BIND built on systems running bits older than snv_37 would
not see this and thus would not use it and would avoid the bug. I
suspect build 41 contained the first BIND packages which were build on
snv_37 or later..
- Bill
_______________________________________________
networking-discuss mailing list
[email protected]
