Re: [PATCH] ipv4: fix the rcu race between free_fib_info and ip_route_output_slow

[Yanmin Zhang <yanmin_zhang@linux.intel.com>]

On Wed, 2012-05-23 at 06:37 +0200, Eric Dumazet wrote:
> On Tue, 2012-05-22 at 23:23 -0400, David Miller wrote:
> > From: Yanmin Zhang <yanmin_zhang@linux.intel.com>
> > Date: Wed, 23 May 2012 11:02:03 +0800
> > 
> > > 1) Why does free_fib_info call call_rcu instead of releasing fi directly?
> > > I assume other cpu might be accessing it. nexthop_nh->nh_dev is in fi.
> > > If other cpu are accessing it, here resetting to NULL would cause other
> > > cpu panic.
> > 
> > Because fib trie lookups are done with RCU locking, therefore we must
> > use RCU freeing to release the object.
> > 
> > What I was trying to impart to you is that removing the NULL
> > assignment is wrong and that an alternative fix is warranted (hint:
> > consider moving something into the RCU release).
> > --
> 
> Its more than that I'm afraid.
> 
> fi->fib_dev (aka fib_nh[0].nh_dev) need full RCU protection.
The new patch posted at https://lkml.org/lkml/2012/5/22/558 does move the
resetting to RCU protection.

> 
> Also the "fib_info_cnt--;" must stay in free_fib_info() (so that it is
> protected by RTNL)
We would move "fib_info_cnt--;" back to free_fib_info.

Thanks,
Yanmin