IPv6 badness continues

This problem seems to be appearing much sooner, in 2.6.10 and 2.6.10-bk kernels. Herbert, your patch did not apply to recent kernels, so I did not boot with it.

The first 'Badness in dst_release' message (of many) from 2.6.10-bk1, on x86 SMP, is found below.

Standard FC2 OS with IPv6 6to4 setup as described on http://linux.yyz.us/ipv6-fc2-howto.html

Sorry, this is my main fileserver and router, so I need to reboot it...

Jeff

Badness in dst_release at include/net/dst.h:149

 [<f8c14a57>] ip6_dst_check+0x67/0x80 [ipv6]
 [<f8c0b2d4>] ip6_dst_lookup+0x1b4/0x1d0 [ipv6]
 [<f8c1e4b4>] udpv6_sendmsg+0x2b4/0x950 [ipv6]
 [<c0337954>] udp_recvmsg+0x64/0x300
 [<c033f12d>] inet_sendmsg+0x4d/0x60
 [<c02ed4bd>] sock_sendmsg+0xdd/0x100
 [<c011d694>] find_busiest_group+0xd4/0x300
 [<c01fa662>] copy_from_user+0x42/0x70
 [<c01371a0>] autoremove_wake_function+0x0/0x60
 [<c02eef5f>] sys_sendmsg+0x18f/0x1f0
 [<c011e431>] __wake_up_common+0x41/0x70
 [<c011e49e>] __wake_up+0x3e/0x60
 [<c0137877>] wake_futex+0x37/0x70
 [<c013792b>] futex_wake+0x7b/0xd0
 [<c01fa662>] copy_from_user+0x42/0x70
 [<c02ef402>] sys_socketcall+0x242/0x260

I upgraded my problematic machine to FC3, which somehow broke 6to4, and I didn't bother to reenable it again because the local DNS server (dnscache) just drops AAAA queries on the floor anyway, leading to lots of timeouts when trying to resolve host names, etc. Ever since 6to4 was turned off on that box I stopped seeing these messages.

--L

Surely it doesn't just drop AAAA queries on the floor, but sends an error message?

If so, maybe then the problem is that glibc resolver cannot properly handle such errors without undue timeouts.

Actually I've just sent a message on libc-alpha@ describing a similar timeout issue:

http://sources.redhat.com/ml/libc-alpha/2005-01/msg00007.html

Maybe you could run a little bit of debug on whether the server sends errors (which kind) and how the resolver reacts.

Pekka Savola                 "You each name yourselves king, yet the
Netcore Oy                    kingdom bleeds."
Systems. Networks. Security. -- George R.R. Martin: A Clash of Kings

hmmm, that's annoying. My router is still FC2, but I'll check that out when I upgrade it to FC3.

Jeff

Sorry, I just tried again and can't reproduce this behavior -- when doing an AAAA query on a random domain name I get a clean 0/0/0 answer indicating no RRs of the desired type having been found, and doing an AAAA query on a domain name that does have RRs of that type, I do get the desired records.

I'm trying to recall what specific problem it was that I was seeing..

--L

Hello.

In article <[email protected]> (at Tue, 04 Jan 2005 01:40:32 -0500), Jeff Garzik <[email protected]> says:

:

Ok, now I doubt sk_dst_check() code path.

When ip6_dst_check() (or its variants) returns NULL, it will release refcnt while it leaves sk->dst_cache non-NULL.
Later, caller sets it to NULL by sk_dst_reset() (or its variants), and it decrements refcnt again.

I think we should NOT release refcnt in sk_dst_check() (or its variants) even we return NULL because that reference was held for sk->dst_cache and it is available yet. Instead, we should release refcnt when we really reset the dst in sk_dst_reset() (or its variants). (Alternatively, we may always set sk->dst_cache to NULL and release refcnt when we're about to return NULL in sk_dst_check()
(or its variants).)

Am I miss something?

===== net/decnet/dn_route.c 1.29 vs edited =====

@@ -253,7 +253,6 @@
  */
 static struct dst_entry *dn_dst_check(struct dst_entry *dst, __u32 cookie)
 {
-	dst_release(dst);
 	return NULL;
 }
 

===== net/ipv4/route.c 1.99 vs edited =====

@@ -1321,7 +1321,6 @@
 
 static struct dst_entry *ipv4_dst_check(struct dst_entry *dst, u32 cookie)
 {
-	dst_release(dst);
 	return NULL;
 }
 

===== net/ipv6/route.c 1.102 vs edited =====

@@ -578,8 +578,6 @@
 
 	if (rt && rt->rt6i_node && (rt->rt6i_node->fn_sernum == cookie))
 		return dst;
-
-	dst_release(dst);
 	return NULL;
 }
 

===== net/xfrm/xfrm_policy.c 1.61 vs edited =====

@@ -1000,8 +1000,6 @@
 {
 	if (!stale_bundle(dst))
 		return dst;
-
-	dst_release(dst);
 	return NULL;
 }
 

Hideaki YOSHIFUJI @ USAGI Project <[email protected]>
GPG FP: 9022 65EB 1ECF 3AD1 0BDF  80D8 4807 F894 E062 0EEA

Does every path which destroys an ipv6 socket do the sk_dst_reset()? If so, your patch should be valid.

In article <[email protected]> (at Thu, 6 Jan 2005 19:53:33 -0800), "David S. Miller" <[email protected]> says:

Currently, inet6_destroy_sock() in net/ipv6/af_inet6.c, which is called from udpv6_destory_sock() etc., takes care of this. In that means, yes.

However, one for ipv4 doesn't do this.

I'll come up with new changesets.

Thanks.

--yoshfuji

I think the current approach is safe. The reason is that sk_dst_cache has its own reference which gets dropped in __sk_dst_reset. So even though ip6_dst_check has dropped a refcnt, it should not hit zero unless there is a bug somewhere else.

Visit Openswan at http://www.openswan.org/
Email: Herbert Xu ~{PmV>HI~} <[email protected]>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt

Here is that patch rediffed against the current bk. BTW, I'd like to access the machine with the problem. However, in order for it to be useful I also need access to the build directory of your kernel so that I can match things up. Please mail me privately to set it up.

Thanks,

Visit Openswan at http://www.openswan.org/
Email: Herbert Xu ~{PmV>HI~} <[email protected]>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt

===== net/ipv6/route.c 1.102 vs edited =====

@@ -379,6 +379,7 @@
 	int err;
 
 	write_lock_bh(&rt6_lock);
+	WARN_ON(rt->u.dst->obsolete);
 	err = fib6_add(&ip6_routing_table, rt, nlh, _rtattr);
 	write_unlock_bh(&rt6_lock);
 

Sorry, I thought you meant that it didn't apply :)

Here is the corrected version.

Visit Openswan at http://www.openswan.org/
Email: Herbert Xu ~{PmV>HI~} <[email protected]>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt

===== net/ipv6/route.c 1.102 vs edited =====

@@ -379,6 +379,7 @@
 	int err;
 
 	write_lock_bh(&rt6_lock);
+	WARN_ON(rt->u.dst.obsolete);
 	err = fib6_add(&ip6_routing_table, rt, nlh, _rtattr);
 	write_unlock_bh(&rt6_lock);
 

I applied the patch and tried it out. The kernel was practically unusable, as the warning hit as soon as the kernel booted:

Sorry. Should've tested it first. Please try this one instead.

Visit Openswan at http://www.openswan.org/
Email: Herbert Xu ~{PmV>HI~} <[email protected]>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt

===== net/ipv6/route.c 1.102 vs edited =====

@@ -379,6 +379,7 @@
 	int err;
 
 	write_lock_bh(&rt6_lock);
+	WARN_ON(rt->u.dst.obsolete > 0);
 	err = fib6_add(&ip6_routing_table, rt, nlh, _rtattr);
 	write_unlock_bh(&rt6_lock);