Re: [PATCH nf] netfilter: nf_tables: unconditionally flush pending work before notifier

[Hillf Danton <hdanton@sina.com>]

On Mon, 8 Jul 2024 13:58:31 +0200 Florian Westphal <fw@strlen.de>
> Hillf Danton <hdanton@sina.com> wrote:
> > On Sun, 7 Jul 2024 10:08:24 +0200 Florian Westphal <fw@strlen.de>
> > > Hillf Danton <hdanton@sina.com> wrote:
> > > > > I think this change might be useful as it also documents
> > > > > this requirement.
> > > > 
> > > > Yes it is boy and the current reproducer triggered another warning [1,2].
> > > > 
> > > > [1] https://lore.kernel.org/lkml/20240706231332.3261-1-hdanton@sina.com/
> > > 
> > > The WARN is incorrect.  The destroy list can be non-empty; i already
> > > tried to explain why.
> > >
> > That warning as-is could be false positive but it could be triggered with a
> > single netns.
> 
> How?
> 
You saw the below cpu diagram, no?

> > 	cpu1		cpu2		cpu3
> > 	---		---		---
> > 					nf_tables_trans_destroy_work()
> > 					spin_lock(&nf_tables_destroy_list_lock);
> > 
> > 					// 1) clear the destroy list
> > 					list_splice_init(&nf_tables_destroy_list, &head);
> > 					spin_unlock(&nf_tables_destroy_list_lock);
> > 
> > 			nf_tables_commit_release()
> > 			spin_lock(&nf_tables_destroy_list_lock);
> > 
> > 			// 2) refill the destroy list
> > 			list_splice_tail_init(&nft_net->commit_list, &nf_tables_destroy_list);
> > 			spin_unlock(&nf_tables_destroy_list_lock);
> > 			schedule_work(&trans_destroy_work);
> > 			mutex_unlock(&nft_net->commit_mutex);
> 
> So you're saying work can be IDLE after schedule_work()?
> 
I got your point but difficult to explain you. In simple words,
like runqueue, workqueue has latency.

> I'm not following at all.

This does not matter but is why I added tj to the cc list.