Re: [PATCH net v3 2/4] flow_offload: fix incorrect cb_priv check for flow_block_cb

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

On Tue, Jun 16, 2020 at 05:47:17PM +0200, Simon Horman wrote:
> On Tue, Jun 16, 2020 at 11:18:16PM +0800, wenxu wrote:
> > 
> > 在 2020/6/16 22:34, Simon Horman 写道:
> > > On Tue, Jun 16, 2020 at 10:20:46PM +0800, wenxu wrote:
> > >> 在 2020/6/16 18:51, Simon Horman 写道:
> > >>> On Tue, Jun 16, 2020 at 11:19:38AM +0800, wenxu@ucloud.cn wrote:
> > >>>> From: wenxu <wenxu@ucloud.cn>
> > >>>>
> > >>>> In the function __flow_block_indr_cleanup, The match stataments
> > >>>> this->cb_priv == cb_priv is always false, the flow_block_cb->cb_priv
> > >>>> is totally different data with the flow_indr_dev->cb_priv.
> > >>>>
> > >>>> Store the representor cb_priv to the flow_block_cb->indr.cb_priv in
> > >>>> the driver.
> > >>>>
> > >>>> Fixes: 1fac52da5942 ("net: flow_offload: consolidate indirect flow_block infrastructure")
> > >>>> Signed-off-by: wenxu <wenxu@ucloud.cn>
> > >>> Hi Wenxu,
> > >>>
> > >>> I wonder if this can be resolved by using the cb_ident field of struct
> > >>> flow_block_cb.
> > >>>
> > >>> I observe that mlx5e_rep_indr_setup_block() seems to be the only call-site
> > >>> where the value of the cb_ident parameter of flow_block_cb_alloc() is
> > >>> per-block rather than per-device. So part of my proposal is to change
> > >>> that.
> > >> I check all the xxdriver_indr_setup_block. It seems all the cb_ident parameter of
> > >>
> > >> flow_block_cb_alloc is per-block. Both in the nfp_flower_setup_indr_tc_block
> > >>
> > >> and bnxt_tc_setup_indr_block.
> > >>
> > >>
> > >> nfp_flower_setup_indr_tc_block:
> > >>
> > >> struct nfp_flower_indr_block_cb_priv *cb_priv;
> > >>
> > >> block_cb = flow_block_cb_alloc(nfp_flower_setup_indr_block_cb,
> > >>                                                cb_priv, cb_priv,
> > >>                                                nfp_flower_setup_indr_tc_release);
> > >>
> > >>
> > >> bnxt_tc_setup_indr_block:
> > >>
> > >> struct bnxt_flower_indr_block_cb_priv *cb_priv;
> > >>
> > >> block_cb = flow_block_cb_alloc(bnxt_tc_setup_indr_block_cb,
> > >>                                                cb_priv, cb_priv,
> > >>                                                bnxt_tc_setup_indr_rel);
> > >>
> > >>
> > >> And the function flow_block_cb_is_busy called in most place. Pass the
> > >>
> > >> parameter as cb_priv but not cb_indent .
> > > Thanks, I see that now. But I still think it would be useful to understand
> > > the purpose of cb_ident. It feels like it would lead to a clean solution
> > > to the problem you have highlighted.
> > 
> > I think The cb_ident means identify.  It is used to identify the each flow block cb.
> > 
> > In the both flow_block_cb_is_busy and flow_block_cb_lookup function check
> > 
> > the block_cb->cb_ident == cb_ident.
> 
> Thanks, I think that I now see what you mean about the different scope of
> cb_ident and your proposal to allow cleanup by flow_indr_dev_unregister().
> 
> I do, however, still wonder if there is a nicer way than reaching into
> the structure and manually setting block_cb->indr.cb_priv
> at each call-site.
> 
> Perhaps a variant of flow_block_cb_alloc() for indirect blocks
> would be nicer?

A follow up patch to add this new variant would be good. Probably
__flow_block_indr_binding() can go away with this new variant to set
up the indirect flow block.