Re: [RFC PATCH v2 4/4] get rid of busy-waiting in shrink_dcache_tree()

[Al Viro <viro@zeniv.linux.org.uk>]

On Fri, Apr 03, 2026 at 05:04:26PM -0700, Linus Torvalds wrote:
> On Fri, 3 Apr 2026 at 16:58, Al Viro <viro@zeniv.linux.org.uk> wrote:
> >
> > FWIW, I wonder if we would be better off with something like
> >
> > struct completion_list {
> >         struct completion completion;
> >         struct completion_list *next;
> > };
> >
> > embedded into select_data, with dentry->waiters being a pointer
> > to that.  It's just that dragging struct select_data into that,
> > even as an opaque pointer, tastes wrong ;-/
> >
> > Preferences?
> 
> I have no strong preferences in this area.
> 
> The opaque pointer is fine by me, because it's type-safe without
> exposing internal implementation or having people being able to much
> around with the fields.
> 
> But if you prefer a completion list, i don't see that being a problem
> either. So entirely up to you.

BTW, how much would you hate something like

/*
 * Check if dentry is already dying and if it is arrange to be notified
 * once it's gone.
 *
 * caller must hold rcu_read_lock and dentry->d_lock
 */
static bool d_deathwatch(struct dentry *dentry, struct completion_list *node)
{
	if (dentry->d_lockref.count >= 0)
		return false;		// it's not dying yet

	init_completion(&node->completion);
	if (unlikely(dentry->d_flags & DCACHE_DENTRY_KILLED)) {
		// it's already gone
		complete(&node->completion);
		return true;
	}
	node->next = dentry->waiters;
	dentry->waiters = node;
	return true;
}

with
		if (data.victim) {
			struct dentry *v = data.victim;
			struct completion_list wait;

			spin_lock(&v->d_lock);
			if (d_deathwatch(v, &wait)) {
				// Another thread got around to killing it.
				// if there's something else to evict, do so
				// then wait for the victim to be gone
				spin_unlock(&v->d_lock);
				rcu_read_unlock();
				if (!list_empty(&data.dispose))
					shrink_dentry_list(&data.dispose);
				wait_for_completion(&wait.completion);
				continue;
			}
			if (!lock_for_kill(v)) {
				spin_unlock(&v->d_lock);
				rcu_read_unlock();
			} else {
				shrink_kill(v);
			}
		}

in shrink_dentry_tree()?