bind-9.5.0b1 problem on ppc64 : rbtdb.c:1532: REQUIRE(prev > 0) failed

[Adam Tkac]

On Tue, Feb 05, 2008 at 09:10:41AM +0700, Fajar A. Nugraha wrote:
> Hi,
> I'm using bind on RHEL5/ppc, on a somewhat busy nameserver. I'm building 
> my own RPM based on Fedora 9 (dev) SRPM, bind-9.5.0-24.b1.
> 
> First I build 32 bit rpm (ppc), which works fine on older ppc machine 
> (power4, CPU, 1GHz). This machine doesn't have the processing power 
> required for our needs. To give an idea of the workload, "rndc status" 
> says "recursive clients" usually over 2000. So we move to newer hardware 
> (power5+, 8 logical CPUs, 1.6 GHz). This is when the problem begins.
> 
> Using the same 32bit binary named process would sometime freeze with no 
> apparent reason. "ps" says its running, but named doesn't response to 
> any queries. Using 64bit binary helps a little, since instead of 
> freezing, now it stops with this error:
> 
> rbtdb.c:1532: REQUIRE(prev > 0) failed
> exiting (due to assertion failure)

Interesting, not only you hit this problem (I didn't tested ppc64
version but ppc version simply stops responding). You will add
yourself to https://bugzilla.redhat.com/show_bug.cgi?id=426613

> 
> digging into the source code, the corresponding code seems to be :
> ./lib/dns/rtdb.c
>     dns_rbtnode_refdecrement(node, &nrefs);
> 
> which leads to ./lib/dns/include/dns/rbt.h
>                 isc_refcount_decrement(&(node)->references, (refs)); \
> 
> which leads to ./lib/isc/include/isc/refcount.h
>                 prev = isc_atomic_xadd(&(rp)->refs, -1);        \
>                 REQUIRE(prev > 0);                              \
> 
> which leads to ./lib/isc/powerpc/include/isc/atomic.h, that contains 
> ppc-specific asm. Way out of my league :-P
> 
> So here's the question :
> - Why did it work correctly on my older machine, but not on this newer 
> machine?
> - Is the new machine perhaps too fast, or have too many logical CPUs?
> - Why did the ppc and ppc64 binary behaves differently? (I like the 
> ppc64 behavior better though)
> 

It is possible that gcc doing some bad stuff, not sure. But generally
ppc binaries are faster than ppc64 binaries

Adam

-- 
Adam Tkac, Red Hat, Inc.