On Sat, Mar 03, 2018 at 09:55:22AM -0500, Jeff Layton wrote:
> On Mon, 2018-02-19 at 11:45 -0800, Matthew Wilcox wrote:
> > From: Matthew Wilcox <mawilcox@xxxxxxxxxxxxx>
> >
> > This results in no change in structure size on 64-bit x86 as it fits in
> > the padding between the gfp_t and the void *.
> >
>
> While the patch itself looks fine, we should take note that this will
> likely increase the size of radix_tree_root on 32-bit arches.
>
> I don't think that's necessarily a deal breaker, but there are a lot of
> users of radix_tree_root. Many of those users have their own spinlock
> for radix tree accesses, and could be trivially changed to use the
> xa_lock. That would need to be done piecemeal though.
>
> A less disruptive idea might be to just create some new struct that's a
> spinlock + radix_tree_root, and then use that going forward in the
> xarray conversion. That might be better anyway if you're considering a
> more phased approach for getting this merged.
Well, it's a choice. If we do:
struct xarray {
spinlock_t xa_lock;
struct radix_tree_root root;
};
then the padding on 64-bit turns that into a 24-byte struct. So do we
spend the extra 4 bytes on 32-bit and have the struct the way we want it
to look from the beginning, or do we spend the extra 8 bytes on 64-bit
and have to redo the struct accessors after the conversions are complete?
I chose option (a), but reasonable people can disagree on that choice.
