Re: [linux-next:master 6618/6917] kernel/sched/psi.c:1230:13: sparse: error: incompatible types in comparison expression (different address spaces)

[Matthew Wilcox <willy@infradead.org>]

On Tue, Feb 12, 2019 at 07:56:10AM -0800, Paul E. McKenney wrote:
> On Mon, Feb 11, 2019 at 05:36:06PM -0800, Matthew Wilcox wrote:
> > radix_tree_iter_resume is, happily, gone from my xarray-conv tree.
> > __radix_tree_lookup, __radix_tree_replace, radix_tree_iter_replace and
> > radix_tree_iter_init are still present, but hopefully not for too much
> > longer.  For example, __radix_tree_replace() is (now) called only from
> > idr_replace(), and there are only 12 remaining callers of idr_replace().
> 
> Will this reduce the number of uses of rcu_dereference_raw()?  Or do they
> simply migrate into Xarray?

Unlike the radix tree (which let you do whatever awful locking scheme you
wanted), the XArray requires that you use the spinlock embedded in the
root of the data structure to protect against simultaneous modification.
So all dereferences within the XArray code look like this:

(if either under lock, or rcu lock held):
        return rcu_dereference_check(node->slots[offset],
                                                lockdep_is_held(&xa->xa_lock));

(if we know the lock is held):
        return rcu_dereference_protected(node->slots[offset],
                                                lockdep_is_held(&xa->xa_lock));

The XArray API doesn't expose slot pointers to its clients.  It hides them
inside the xa_state's pointer to the current xa_node.