Re: [QUESTION] llist: Comment releasing 'must delete' restriction before traversing

[Byungchul Park]

On Tue, Jul 31, 2018 at 07:30:52AM -0700, Paul E. McKenney wrote:
> On Tue, Jul 31, 2018 at 06:29:50PM +0900, Byungchul Park wrote:
> > On Mon, Jul 30, 2018 at 09:30:42PM -0700, Paul E. McKenney wrote:
> > > On Tue, Jul 31, 2018 at 09:58:36AM +0900, Byungchul Park wrote:
> > > > Hello folks,
> > > > 
> > > > I'm careful in saying.. and curious about..
> > > > 
> > > > In restrictive cases like only addtions happen but never deletion, can't
> > > > we safely traverse a llist? I believe llist can be more useful if we can
> > > > release the restriction. Can't we?
> > > 
> > > Yes, but please give a thought to the people looking at your code some
> > > time down the line.  If you are doing this, lots of comments, please.
> > 
> > Yes, I will. Thank you for the comment.
> > 
> > > Here are the approaches that I am aware of:
> > > 
> > > 1.	Normal RCU.  Use list_add_rcu(), list_del_rcu(), and friends.
> > > 
> > > 2.	Things are added but never deleted.  Use list_add_rcu() and
> > > 	friends, but since you don't ever delete anything, you never
> > > 	use list_del_rcu(), synchronize_rcu(), call_rcu(), and friends.
> > 
> > I think rcu list also works well. But I decided to use llist because
> > llist is simpler and has one less pointer.
> 
> No.
> 
> To see this, look at llist_for_each() below, which is absolutely -not-
> able to reliably traverse lists while nodes are being inserted.

Ah.. Not only 'node' but also 'pos->next' can cause a problem w/o
rcu_dereference or similar.. I need consider another way.

Or I might need memory barrier between READ_ONCE(head->first) and
llist_for_each() to make sure safely to read 'pos->next' when I see
a value of 'head->first'.

> #define llist_for_each(pos, node)			\
> 	for ((pos) = (node); pos; (pos) = (pos)->next)
> 
> Now, you could introduce an llist_for_each_rcu() that used rcu_dereference
> or similar (thus handling insertion, but that is not what your patches
> currently do.

Right.

> > Just to be sure, let me explain my use case more:
> > 
> >    1. Introduced a global list where single linked list is sufficient.
> >    2. All operations I need is to add items and traverse the list.
> >    3. Ensure the operations always happen within irq-disabled section.
> >    4. I'm considering CONFIG_ARCH_HAVE_NMI_SAFE_CMPXCHG properly.
> >    5. The list can be accessed by every CPU concurrently.
> > 
> > Of course, I can use normal double list with a lock or rcu list. But I
> > think it doesn't have to be protected by even rcu in that case. I wanted
> > to use the simplest one all requiremnets are satisfied with and I
> > thought it's llist. Thoughts?
> 
> If you want lockless reader traversal, you need rcu_dereference().

Yes, I need that or similar anyway.

Thanks a lot, Paul!