GNOME Bugzilla – Bug 667692
Kernel stalls when using IPv6
Last modified: 2012-07-24 16:55:40 UTC
Lately, I've been getting the following in dmesg: INFO: rcu_preempt_state detected stalls on CPUs/tasks: {} (detected by 3, t=1314770 jiffies) This happens exactly once every 3 minutes. After a while, any program that uses the network in some way will hang. If I kill networkmanager, the stalls stop and the programs start responding again (and the network keeps functioning). My router uses radvd to provide ipv6 addresses to the network. If I blacklist the ipv6 module, the stalls never happen. I'm using Archlinux, which also has a bug open on this issue: https://bugs.archlinux.org/task/26847. There are also people on other distributions in there with the same problems.
The comment https://bugs.archlinux.org/task/26847#comment86869 says that there is no problem on Fedora kernels. It could be due to incorporating of IPv6 into the Fedora kernel instead of use of a module: * http://lists.fedoraproject.org/pipermail/kernel/2011-June/003105.html * changelog: http://koji.fedoraproject.org/koji/buildinfo?buildID=246867
'ipv6' is now in-kernel on Arch, and yet the stalls continue. Currently testing NM 0.9.3.955, still the same.
Same here, on archlinux. Had to deactive ipv6 support. Kinda annoying.
It would be nice to know what behavior in NM might be triggering this, but in the end, the kernel should never be affected by userland behavior like this. So while we might need to mitigate this in NM (if we can figure out what's actually causing it) there's clearly a kernel bug as well.
The kernel bug appears to have been fixed in 3.4 & 3.3.3: commit 94f826b8076e2cb92242061e92f21b5baa3eccc2 Author: Eric Dumazet <eric.dumazet@gmail.com> Date: Tue Mar 27 09:53:52 2012 +0000 net: fix a potential rcu_read_lock() imbalance in rt6_fill_node()