Re: [RFC PATCH 1/2] lib/percpu-list: Per-cpu list with associated per-cpu locks

[Peter Zijlstra]

On Wed, Feb 17, 2016 at 10:10:02PM +1100, Dave Chinner wrote:
> On Wed, Feb 17, 2016 at 12:00:40PM +0100, Peter Zijlstra wrote:

> > Yeah, that is pretty terrible. Maybe a visitor interface is advisable?
> > 
> > visit_percpu_list_entries(struct percpu_list *head, void (*visitor)(struct list_head *pos, void *data), void *data)
> > {
> > 	int cpu;
> > 
> > 	for_each_possible_cpu(cpu) {
> > 		spinlock_t *lock = per_cpu_ptr(&head->lock, cpu);
> > 		struct list_head *head = per_cpu_ptr(&head->list, cpu);
> > 		struct list_head *pos, *tmp;
> > 
> > 		spin_lock(lock);
> > 		for (pos = head->next, tmp = pos->next; pos != head; pos = tmp)
> > 			visitor(pos, data);
> 
> I thought about this - it's the same problem as the list_lru walking
> functions. That is, the visitor has to be able to drop the list lock
> to do blocking operations, so the lock has to be passed to the
> visitor/internal loop context somehow, and the way the callers can
> use it need to be documented.

But you cannot drop the lock and guarantee fwd progress. The moment you
drop the lock, you have to restart the iteration from the head, since
any iterator you had might now be pointing into space.