Re: [PATCH net-next v2 1/7] net: flow_offload: provision conntrack info in ct_metadata

[Vlad Buslov <vladbu@nvidia.com>]

On Tue 17 Jan 2023 at 12:09, Marcelo Ricardo Leitner <marcelo.leitner@gmail.com> wrote:
> On Tue, Jan 17, 2023 at 12:04:32PM -0300, Marcelo Ricardo Leitner wrote:
>> On Fri, Jan 13, 2023 at 05:55:42PM +0100, Vlad Buslov wrote:
>> ...
>> >  struct flow_match {
>> > @@ -288,6 +289,7 @@ struct flow_action_entry {
>> >  		} ct;
>> >  		struct {
>> >  			unsigned long cookie;
>> > +			enum ip_conntrack_info ctinfo;
>> >  			u32 mark;
>> >  			u32 labels[4];
>> >  			bool orig_dir;
>> > diff --git a/net/sched/act_ct.c b/net/sched/act_ct.c
>> > index 0ca2bb8ed026..515577f913a3 100644
>> > --- a/net/sched/act_ct.c
>> > +++ b/net/sched/act_ct.c
>> > @@ -187,6 +187,7 @@ static void tcf_ct_flow_table_add_action_meta(struct nf_conn *ct,
>> >  	/* aligns with the CT reference on the SKB nf_ct_set */
>> >  	entry->ct_metadata.cookie = (unsigned long)ct | ctinfo;
>>                                                    ^^^^^^^^^^^
>
> Hmm. Thought that just came up and still need to dig into, but wanted
> to share/ask already. Would it be a problem to update the cookie later
> on then, to reflect the new ctinfo?

Not sure I'm following, but every time the flow changes state it is
updated in the driver since new metadata is generated by calling
tcf_ct_flow_table_fill_actions() from nf_flow_offload_rule_alloc().

>
>> 
>> >  	entry->ct_metadata.orig_dir = dir == IP_CT_DIR_ORIGINAL;
>> > +	entry->ct_metadata.ctinfo = ctinfo;
>> 
>> tcf_ct_flow_table_restore_skb() is doing:
>>         enum ip_conntrack_info ctinfo = cookie & NFCT_INFOMASK;
>> 
>> Not sure if it really needs this duplication then.
>> 
>> >  
>> >  	act_ct_labels = entry->ct_metadata.labels;
>> >  	ct_labels = nf_ct_labels_find(ct);
>> > -- 
>> > 2.38.1
>> >