On 22.07.2015 14:56, Daniel Borkmann wrote:

On 07/22/2015 11:23 AM, Konstantin Khlebnikov wrote:

quoted

In dev_queue_xmit() net_cls protected with rcu-bh.

...

quoted

Signed-off-by: Konstantin Khlebnikov <redacted>
---
  net/core/netclassid_cgroup.c |    3 ++-
  1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/net/core/netclassid_cgroup.c b/net/core/netclassid_cgroup.c
index 1f2a126f4ffa..6441f47b1a8f 100644
--- a/net/core/netclassid_cgroup.c
+++ b/net/core/netclassid_cgroup.c

@@ -23,7 +23,8 @@ static inline struct cgroup_cls_state

*css_cls_state(struct cgroup_subsys_state

  struct cgroup_cls_state *task_cls_state(struct task_struct *p)
  {
-    return css_cls_state(task_css(p, net_cls_cgrp_id));
+    return css_cls_state(task_css_check(p, net_cls_cgrp_id,
+                        rcu_read_lock_bh_held()));

Did you also check that after your patch this doesn't trigger on ingress
either? There, this code path could be invoked under rcu_read_lock(). So,
perhaps you need to check for both.

I haven't seen warnings with this patch. rcu_read_lock_held() is
checked inside rcu_dereference_check() inside task_css_check().

So, overall check expression is:

lockdep_is_held(&cgroup_mutex) ||  <- in task_css_set_check
lockdep_is_held(&css_set_rwsem) ||
((task)->flags & PF_EXITING) ||
rcu_read_lock_bh_held() || <- mine
rcu_read_lock_held() <- in rcu_dereference_check

quoted

  }
  EXPORT_SYMBOL_GPL(task_cls_state);

-- 
Konstantin