On Mon, Nov 09, 2015 at 12:30:08PM -0800, Dave Hansen wrote:
> On 11/09/2015 11:52 AM, Matthew Wilcox wrote:
> > On Sun, Nov 08, 2015 at 02:27:39PM -0500, Dan Williams wrote:
> >> As Dave points out when dealing with the contents of a page we use
> >> PAGE_SIZE and PAGE_SHIFT, similary for huge pages use HPAGE_SIZE and
> >> HPAGE_SHIFT.
> >
> > Hang on, no. We want to use PMD_SIZE here so that when we add PGD_SIZE
> > support, it's all nicely symmetrical.
>
> This is splitting hairs a bit, but here goes.
>
> Do we care about the pagetables, or do we care about the page sizes?
>
> If we are doing pagetable things, we use P{TE,MD,UD,GD}_SIZE. If we are
> doing something related to page (contents) themselves, we use
> H?PAGE_SIZE. This makes it clear whether we're dealing with pagetables,
> or *contents*.
>
> In the end, they're the same thing on our architecture at the moment, so
> it matters very little.
Agreed we're splitting hairs here but I'm not sure why the hair can
be split :-) We come here from the page fault handler, where we've
discovered that there is no page table for a PMD-shaped hole in the
address space, and we'd like a PMD inserted by the fault handler.
In that sense, we care about PMD_SIZE.
Another problem is that the name 'HPAGE' doesn't obviously say "I cover
a PMD-sized hole" -- hugetlbfs can use 1GB pages as well as 2MB pages.
Then there's the problem that there *is* no equivalent to HPAGE for PUDs
or PGDs -- if PMDs are "huge", then what are the names for pages pointed
to by PUDs and PGDs?
Why is it important to distinguish between "the size of something pointed
to by a PMD" and "the size of the contents"? Actually, I'm not sure I
even understand how to summarise the distinction.
--
To unsubscribe from this list: send the line "unsubscribe linux-block" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
