Re: [PATCH net] net: drop secpath extension before skb deferral free

[Jianbo Liu <jianbol@nvidia.com>]

On Thu, 2024-05-23 at 12:00 +0200, Steffen Klassert wrote:
> On Thu, May 23, 2024 at 06:57:25AM +0000, Jianbo Liu wrote:
> > On Thu, 2024-05-23 at 08:44 +0200, Steffen Klassert wrote:
> > > On Thu, May 23, 2024 at 02:22:38AM +0000, Jianbo Liu wrote:
> > > > On Wed, 2024-05-22 at 11:34 +0200, Steffen Klassert wrote:
> > > > > 
> > > > > Maybe we should directly remove the device from the
> > > > > xfrm_state
> > > > > when the decice goes down, this should catch all the cases.
> > > > > 
> > > > > I think about something like this (untested) patch:
> > > > > 
> > > > > diff --git a/net/xfrm/xfrm_state.c b/net/xfrm/xfrm_state.c
> > > > > index 0c306473a79d..ba402275ab57 100644
> > > > > --- a/net/xfrm/xfrm_state.c
> > > > > +++ b/net/xfrm/xfrm_state.c
> > > > > @@ -867,7 +867,11 @@ int xfrm_dev_state_flush(struct net
> > > > > *net,
> > > > > struct
> > > > > net_device *dev, bool task_vali
> > > > >                                 xfrm_state_hold(x);
> > > > >                                 spin_unlock_bh(&net-
> > > > > > xfrm.xfrm_state_lock);
> > > > >  
> > > > > -                               err = xfrm_state_delete(x);
> > > > > +                               spin_lock_bh(&x->lock);
> > > > > +                               err = __xfrm_state_delete(x);
> > > > > +                               xfrm_dev_state_free(x);
> > > > > +                               spin_unlock_bh(&x->lock);
> > > > > +
> > > > >                                 xfrm_audit_state_delete(x,
> > > > > err ? 0
> > > > > :
> > > > > 1,
> > > > >                                                         task_
> > > > > valid)
> > > > > ;
> > > > >                                 xfrm_state_put(x);
> > > > > 
> > > > > The secpath is still attached to all skbs, but the hang on
> > > > > device
> > > > > unregister should go away.
> > > > 
> > > > It didn't fix the issue.
> > > 
> > > Do you have a backtrace of the ref_tracker?
> > > 
> > > Is that with packet offload?
> > > 
> > 
> > Yes. And it's the same trace I posted before.
> > 
> >  ref_tracker: eth%d@000000007421424b has 1/1 users at
> >       xfrm_dev_state_add+0xe5/0x4d0
> 
> Hm, interesting.
> 
> Can you check if xfrm_dev_state_free() is triggered in that codepath
> and if it actually removes the device from the states?
> 

xfrm_dev_state_free is not triggered. I think it's because I did "ip x
s delall" before unregister netdev.

Besides, as it's possible to sleep in dev's xdo_dev_state_free, there
is a "scheduling while atomic" issue to call xfrm_dev_state_free in
spin_lock.