On Tue, Aug 15, 2023 at 1:07 PM Matthew Wilcox <willy@xxxxxxxxxxxxx> wrote:
>
> On Tue, Aug 15, 2023 at 03:24:16PM -0400, Peter Xu wrote:
> > On Tue, Aug 15, 2023 at 04:26:44AM +0100, Matthew Wilcox (Oracle) wrote:
> > > +++ b/include/linux/page-flags.h
> > > @@ -99,13 +99,15 @@
> > > */
> > > enum pageflags {
> > > PG_locked, /* Page is locked. Don't touch. */
> > > + PG_writeback, /* Page is under writeback */
> > > PG_referenced,
> > > PG_uptodate,
> > > PG_dirty,
> > > PG_lru,
> > > + PG_head, /* Must be in bit 6 */
> >
> > Could there be some explanation on "must be in bit 6" here?
>
> Not on this line, no. You get 40-50 characters to say something useful.
> Happy to elaborate further in some other comment or in the commit log,
> but not on this line.
>
> The idea behind all of this is that _folio_order moves into the bottom
> byte of _flags_1. Because the order can never be greater than 63 (and
> in practice I think the largest we're going to see is about 30 -- a 16GB
> hugetlb page on some architectures), we know that PG_head and PG_waiters
> will be clear, so we can write (folio->_flags_1 & 0xff) and the compiler
> will just load a byte; it won't actually load the word and mask it.
>
> We can't move PG_head any lower, or we'll risk having a tail page with
> PG_head set (which can happen with the vmemmmap optimisation, but eugh).
> And we don't want to move it any higher because then we'll have a flag
> that cannot be reused in a tail page. Bit 6 is the perfect spot for it.
Is there some compile time assertion that the order cannot overflow into bit 6?
