Santiago Garcia Mantinan wrote:
> I'm running current Jessie's Debian except for the kernel which is a 3.10.69
> kernel on a mips based router (http://wiki.openwrt.org/toh/arcadyan/arv7519)
> the machine doesn't have floating point so I'm running with floating point
> emulated.
>
> All the services I'm running on Debian run Ok (including apache, postfix,
> squid3, isc-dhcp-server, ...) but bind9 hangs after some time, the time
> doesn't seem to follow any pattern, it can be from less than 10 minutos to
> some hours (not many, not a full day).
>
> Tying to see what happens with this problem I came around this forum:
> http://community.ubnt.com/t5/EdgeMAX/How-To-Install-BIND-DNS-Server-on-EdgeRouter/td-p/558349/page/3
> a message from napsterbater that points to:
> https://lists.isc.org/pipermail/bind-users/2008-August/072137.html
> on wich people suggests to use --disable-atomic to solve crashes.
>
> I have to explain that my bind doesn't crash, it just hangs without replying
> to queries or to a reload or stop command issued from the init.d script.
>
> However as I didn't have any better solution I tried the --disable-atomic
> way.
>
> To do this I just added to the rules file after the kfreebsd exception this
> code:
>
> ifeq ($(DEB_HOST_ARCH),mips)
> EXTRA_FEATURES=--disable-atomic
> endif
>
> And rebuilt bind9's packages with this code on our rules.
>
> This build has been in production for more than 27 hours without problems,
> so It seem that the atomic fix is working.
>
> There might be some other ways to solve this problem, if so and you need
> some testing, please let me know.
Hi, Santiago:
If adding --disable-atomic to configure works around the problem,
there's a problem in the implementation of the atomic operations for
your architecture.
I asked on the debian-mips mailing list and found that there's a problem
in the isc_atomic_cmpxchg() implementation on mips. (Note that this
function has been patched in the Debian package, see [0] and [1].)
Can you try compiling with the attached patch (and without
--disable-atomic) and see if it fixes the problem?
Thanks!
[0]
http://anonscm.debian.org/cgit/users/lamont/bind9.git/commit/?id=14cb2d0750593e2d51c7e028a678f3fd00775fea
[1]
http://anonscm.debian.org/cgit/users/lamont/bind9.git/commit/?id=98b8d193db59e79d7f98df904852988ed1dd71da
--
Robert Edmonds
edmo...@debian.org
>From 801989597d5e7ebad86bb1d993975dfc83983133 Mon Sep 17 00:00:00 2001
From: Robert Edmonds <edmo...@debian.org>
Date: Mon, 7 Sep 2015 21:27:57 -0400
Subject: [PATCH] Add earlyclobber constraint in isc_atomic_cmpxchg() on mips
According to James Cowgill,
The tmp register constraint should be earlyclobber ("=&r") because
the other two registers (%3 and %4) will be read again after tmp
(%2) is written if the code loops around. At the moment, GCC could
make tmp and cmpval/val refer to the same register.
(https://lists.debian.org/debian-mips/2015/09/msg00041.html)
Addresses-Debian-Bug: 778720
Signed-off-by: Robert Edmonds <edmo...@debian.org>
---
lib/isc/mips/include/isc/atomic.h | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/lib/isc/mips/include/isc/atomic.h b/lib/isc/mips/include/isc/atomic.h
index e3657d7..13b5cff 100644
--- a/lib/isc/mips/include/isc/atomic.h
+++ b/lib/isc/mips/include/isc/atomic.h
@@ -79,7 +79,7 @@ isc_atomic_cmpxchg(isc_int32_t *p, int cmpval, int val) {
" beqz %2, 1b \n"
"2: move %0, $1 \n"
" .set pop \n"
- : "=&r"(orig), "+R" (*p), "=r" (tmp)
+ : "=&r"(orig), "+R" (*p), "=&r" (tmp)
: "r"(cmpval), "r"(val)
: "memory");
--
2.5.1
