On Wed, Jul 31, 2013 at 6:43 PM, Hannes Frederic Sowa
[off-list ref] wrote:

On Wed, Jul 31, 2013 at 05:55:13AM -0700, Paul E. McKenney wrote:

quoted

On Wed, Jul 31, 2013 at 04:40:47PM +0530, vinayak menon wrote:

quoted

Hi,

A crash was seen on 3.4.5 kernel during some random wlan operations.

CPU: Single core ARM Cortex A9.

fib_route_seq_next was called with second argument (void *v) as 0xd6e3e360
which is a "freed" object of the "ip_fib_trie" cache. I confirmed that the
object was freed with crash utility.

Sequence: fib_route_seq_next->trie_nextleaf->leaf_walk_rcu

As "v" was a freed object, inside trie_nextleaf(), node_parent_rcu()
returned an invalid tnode. But as I had enabled slab poisoning and the
object was already freed, the tnode was 0x6b6b6b6b. And this was passed to
leaf_walk_rcu and resulted in the crash.

fib_route_seq_start, takes rcu_read_lock(), but free_leaf
calls call_rcu_bh. Can this be the problem ?
Should rcu_read_lock() in fib_route_seq_start be changed to rcu_read_lock_bh()
?

One way or the other, the RCU read-side primitives need to match the RCU
update-side primitives.  Adding netdev...

Already fixed by:

commit 0c03eca3d995e73d691edea8c787e25929ec156d
Author: Eric Dumazet [off-list ref]
Date:   Tue Aug 7 00:47:11 2012 +0000

    net: fib: fix incorrect call_rcu_bh()

    After IP route cache removal, I believe rcu_bh() has very little use and
    we should remove this RCU variant, since it adds some cycles in fast
    path.

    Anyway, the call_rcu_bh() use in fib_true is obviously wrong, since
    some users only assert rcu_read_lock().

Thanks. I missed this somehow.