Re: [PATCH 1/2] mm: Allow architectures to request 'old' entries when prefaulting

[Matthew Wilcox]

On Mon, Dec 14, 2020 at 09:54:06AM -0800, Linus Torvalds wrote:
> > I expected to hate it more, but it looks reasonable. Opencoded
> > xas_for_each() smells bad, but...
> 
> I think the open-coded xas_for_each() per se isn't a problem, but I
> agree that the startup condition is a bit ugly. And I'm actually
> personally more confused by why xas_retry() is needed here, bit not in
> many other places. That is perhaps more obvious now that it shows up
> twice.
> 
> Adding Willy to the cc in case he has comments on that, and can
> explain it to me in small words.
> 
> [ https://lore.kernel.org/lkml/20201214160724.ewhjqoi32chheone@box/
> for context ]

The xas_retry() is something I now regret, but haven't got annoyed enough
by it yet to fix (also, other projects).  It originated in the radix
tree where we would get a radix_tree_node and then iterate over it in
header macros.  If we're holding the rcu_read_lock() and somebody else
deletes an entry leaving the entry at index 0 as the only index in the
tree, we tell the RCU readers to rewalk the tree from the top by putting
a retry entry in place of the real entry.

It's not entirely clear to me now why we did that.  Just leave the entry
alone and the RCU-walkers will see it, then the rest of the node is empty.

As to why we need to do this in some places and not others; you can
only see a retry entry if you're only protected by the RCU lock.  If
you're protected by the spinlock, you can't see any nodes which
contain retry entries.

But I now think we should just get rid of retry entries.  Maybe I'm
missing a good reason to keep them.