Fix race condition in ipv6_add_addr

Here's a patch originally from Keir Fraser, which we included in SLES10, but which we forgot to submit upstream so far.

During stress testing, machines were frequently crashing in __ipv6_ifa_notify on dst_hold(&ifp->rt.u_dst), with ifp->rt being a NULL pointer.

The attached patch fixes the problem.

Thanks,
Olaf

Olaf Kirch   |  --- o --- Nous sommes du soleil we love when we play
[email protected] |    / | \   sol.dhoop.naytheet.ah kin.ir.samse.qurax

From: Keir Fraser <[email protected]>
Subject: ipv6_add_addr should install dstentry earlier

ipv6_add_addr allocates a struct inet6_ifaddr and a dstentry, but it doesn't install the dstentry in ifa->rt until after it releases the addrconf_hash_lock. This means other CPUs will be able to see the new address while it hasn't been initialized completely yet.

One possible fix would be to grab the ifp->lock spinlock when creating the address struct; a simpler fix is to just move the assignment.

@@ -549,6 +549,8 @@
 	ifa->flags = flags | IFA_F_TENTATIVE;
 	ifa->cstamp = ifa->tstamp = jiffies;
 
+	ifa->rt = rt;
+
 	ifa->idev = idev;
 	in6_dev_hold(idev);
 	/* For caller */

@@ -575,8 +577,6 @@
 	}
 #endif
 
-	ifa->rt = rt;
-
 	in6_ifa_hold(ifa);
 	write_unlock(&idev->lock);
 out2:

In article <[email protected]> (at Tue, 29 Aug 2006 10:35:36 +0200), Olaf Kirch <[email protected]> says:

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

From: YOSHIFUJI Hideaki <[email protected]>
Date: Tue, 29 Aug 2006 18:21:28 +0900 (JST)

Applied, thanks everyone.

Although the only reason this is a problem is due to the fact that we use rwlock's here. If addrconf_lock were a normal spinlock, or taken as a writer in this code path, the problematic cases would not be possible.

I guess this is a reminder that I need to revisit my patches which move all IPV6 address and inet6_dev changes out of software interrupt context, plus add use of RCU. :-)