Re: [PATCH net-next,v4 07/12] net: sched: use flow block API

[Pablo Neira Ayuso <pablo@netfilter.org>]

On Wed, Aug 14, 2019 at 05:32:34PM +0100, Edward Cree wrote:
> On 09/07/2019 21:55, Pablo Neira Ayuso wrote:
> > This patch adds tcf_block_setup() which uses the flow block API.
> >
> > This infrastructure takes the flow block callbacks coming from the
> > driver and register/unregister to/from the cls_api core.
> >
> > Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
> > ---
> > <snip>
> > @@ -796,13 +804,20 @@ static int tcf_block_offload_cmd(struct tcf_block *block,
> >  				 struct netlink_ext_ack *extack)
> >  {
> >  	struct tc_block_offload bo = {};
> > +	int err;
> >  
> >  	bo.net = dev_net(dev);
> >  	bo.command = command;
> >  	bo.binder_type = ei->binder_type;
> >  	bo.block = block;
> >  	bo.extack = extack;
> > -	return dev->netdev_ops->ndo_setup_tc(dev, TC_SETUP_BLOCK, &bo);
> > +	INIT_LIST_HEAD(&bo.cb_list);
> > +
> > +	err = dev->netdev_ops->ndo_setup_tc(dev, TC_SETUP_BLOCK, &bo);
> > +	if (err < 0)
> > +		return err;
> > +
> > +	return tcf_block_setup(block, &bo);
> >  }
> >  
> >  static int tcf_block_offload_bind(struct tcf_block *block, struct Qdisc *q,
> > @@ -1636,6 +1651,77 @@ void tcf_block_cb_unregister(struct tcf_block *block,
> >  }
> >  EXPORT_SYMBOL(tcf_block_cb_unregister);
> >  
> > +static int tcf_block_bind(struct tcf_block *block,
> > +			  struct flow_block_offload *bo)
> > +{
> > +	struct flow_block_cb *block_cb, *next;
> > +	int err, i = 0;
> > +
> > +	list_for_each_entry(block_cb, &bo->cb_list, list) {
> > +		err = tcf_block_playback_offloads(block, block_cb->cb,
> > +						  block_cb->cb_priv, true,
> > +						  tcf_block_offload_in_use(block),
> > +						  bo->extack);
> > +		if (err)
> > +			goto err_unroll;
> > +
> > +		i++;
> > +	}
> > +	list_splice(&bo->cb_list, &block->cb_list);
> > +
> > +	return 0;
> > +
> > +err_unroll:
> > +	list_for_each_entry_safe(block_cb, next, &bo->cb_list, list) {
> > +		if (i-- > 0) {
> > +			list_del(&block_cb->list);
> > +			tcf_block_playback_offloads(block, block_cb->cb,
> > +						    block_cb->cb_priv, false,
> > +						    tcf_block_offload_in_use(block),
> > +						    NULL);
> > +		}
> > +		flow_block_cb_free(block_cb);
> > +	}
> > +
> > +	return err;
> > +}
> >
> Why has the replay been moved from the function called by the driver
>  (__tcf_block_cb_register()) to work done by the driver's caller based on
>  what the driver has left on this flow_block_offload.cb_list?  This makes
>  it impossible for the driver to (say) unregister a block outside of an
>  explicit request from ndo_setup_tc().
> In my under-development driver, I have a teardown path called on PCI
>  remove, which calls tcf_block_cb_unregister() on all my block bindings
>  (of which the driver keeps track), to ensure that no flow rules are still
>  in place when unregister_netdev() is called;

It's the subsystem that has to release resources when
unregister_netdev() event happens. At least in netfilter, when the
device is going away, the filtering policy is removed, hence the
FLOW_BLOCK_UNBIND is called to release the blocks and, hence, the
offload resources. I remember tc ingress qdisc works like this too.

>  this is needed because some  of the driver's state for certain
>  rules involves taking a reference on  the netdevice (dev_hold()). 
>  Your structural changes here make that  impossible; is there any
>  reason why they're necessary?

May I have access to your driver code?

This would make it easier for me to understand your requirements, and
to discuss changes with you.