Re: [PATCH net 0/2] bond: fix xfrm offload feature during init

[Hangbin Liu]

On Thu, Jan 09, 2025 at 09:26:38AM +0800, Jianbo Liu wrote:
> 
> 
> On 1/8/2025 3:14 PM, Hangbin Liu wrote:
> > On Wed, Jan 08, 2025 at 11:40:05AM +0800, Jianbo Liu wrote:
> > > 
> > > 
> > > On 1/8/2025 10:46 AM, Hangbin Liu wrote:
> > > > On Mon, Jan 06, 2025 at 10:47:16AM +0000, Hangbin Liu wrote:
> > > > > On Thu, Jan 02, 2025 at 11:33:34AM +0800, Jianbo Liu wrote:
> > > > > > > > Re-locking doesn't look great, glancing at the code I don't see any
> > > > > > > > obvious better workarounds. Easiest fix would be to don't let the
> > > > > > > > drivers sleep in the callbacks and then we can go back to a spin lock.
> > > > > > > > Maybe nvidia people have better ideas, I'm not familiar with this
> > > > > > > > offload.
> > > > > > > 
> > > > > > > I don't know how to disable bonding sleeping since we use mutex_lock now.
> > > > > > > Hi Jianbo, do you have any idea?
> > > > > > > 
> > > > > > 
> > > > > > I think we should allow drivers to sleep in the callbacks. So, maybe it's
> > > > > > better to move driver's xdo_dev_state_delete out of state's spin lock.
> > > > > 
> > > > > I just check the code, xfrm_dev_state_delete() and later
> > > > > dev->xfrmdev_ops->xdo_dev_state_delete(x) have too many xfrm_state x
> > > > > checks. Can we really move it out of spin lock from xfrm_state_delete()
> > > > 
> > > > I tried to move the mutex lock code to a work queue, but found we need to
> > > > check (ipsec->xs == xs) in bonding. So we still need xfrm_state x during bond
> > > 
> > > Maybe I miss something, but why need to hold spin lock. You can keep xfrm
> > > state by its refcnt.
> > 
> > Do you mean move the xfrm_dev_state_delete() out of spin lock directly like:
> > 
> 
> Yes. Not feasible?
> 
> > diff --git a/net/xfrm/xfrm_state.c b/net/xfrm/xfrm_state.c
> > index 67ca7ac955a3..6881ddeb4360 100644
> > --- a/net/xfrm/xfrm_state.c
> > +++ b/net/xfrm/xfrm_state.c
> > @@ -766,13 +766,6 @@ int __xfrm_state_delete(struct xfrm_state *x)
> >   		if (x->encap_sk)
> >   			sock_put(rcu_dereference_raw(x->encap_sk));
> > -		xfrm_dev_state_delete(x);
> > -
> > -		/* All xfrm_state objects are created by xfrm_state_alloc.
> > -		 * The xfrm_state_alloc call gives a reference, and that
> > -		 * is what we are dropping here.
> > -		 */
> > -		xfrm_state_put(x);
> >   		err = 0;
> >   	}
> > @@ -787,8 +780,20 @@ int xfrm_state_delete(struct xfrm_state *x)
> >   	spin_lock_bh(&x->lock);
> >   	err = __xfrm_state_delete(x);
> >   	spin_unlock_bh(&x->lock);
> > +	if (err)
> > +		return err;
> > -	return err;
> > +	if (x->km.state == XFRM_STATE_DEAD) {
> > +		xfrm_dev_state_delete(x);
> > +
> > +		/* All xfrm_state objects are created by xfrm_state_alloc.
> > +		 * The xfrm_state_alloc call gives a reference, and that
> > +		 * is what we are dropping here.
> > +		 */
> > +		xfrm_state_put(x);
> > +	}
> > +
> > +	return 0;
> >   }
> >   EXPORT_SYMBOL(xfrm_state_delete);
> > 
> > Then why we need the spin lock in xfrm_state_delete?
> > 
> 
> No, we don't need. But I am trying to understand what you said in your last
> email about adding a new lock, or unlocking spin lock in

I *thought* we need the spin lock in xfrm_state_delete(). So to protect xfrm_state,
we need a new lock. Although it looks redundant. e.g. 

int xfrm_state_delete(struct xfrm_state *x)
{
        int err;

        spin_lock_bh(&x->lock);
        err = __xfrm_state_delete(x);
        spin_unlock_bh(&x->lock);
        if (err)
                return err;

	another_lock(&x->other_lock)
        if (x->km.state == XFRM_STATE_DEAD) {
                xfrm_dev_state_delete(x);
                xfrm_state_put(x);
        }
	another_unlock(&x->other_lock)

        return 0;
}
> bond_ipsec_del_sa(). Anything I missed?

The unlock spin lock in bond_ipsec_del_sa looks like
https://lore.kernel.org/netdev/Z1vfsAyuxcohT7th@fedora/

Thanks
Hangbin