Re: [RFC][PATCH] make /proc/pid/pagemap work with huge pages and return page size

[Matt Mackall <mpm@selenic.com>]

On Fri, 2008-02-29 at 14:30 -0800, Dave Hansen wrote:
> On Thu, 2008-02-28 at 17:15 -0800, Matt Mackall wrote:
> > > The only issue is that this is *after* the code has decided that a
> > > particular virtual area is for huge pages.  The best arch-generic
> > > interface I know for that is: is_vm_hugetlb_page(), but that is
> > > VMA-based.  Perhaps we should change the pagemap walk to pass the VMA
> > > around. 
> > 
> > I'd rather avoid that. Requiring a VMA to poke at these things shouldn't
> > -really- be necessary.
> 
> Yeah, this is strictly true.  But, it also assumes that we have all of
> the data about where large pages are based on the contents of the
> pagetables alone.  It's possible that we can derive this, or code up a
> bunch of arch-specific functions to do this, but it certainly doesn't
> exist today.  I'm just not keen on going and learning about how each and
> every architecture encodes hugetlb information in their pagetables. :(
> 
> The fact is that we treat pagetables in hugetlb areas completely
> differently than normal pages.  All of the generic functions that deal
> with pagetables magically punt over to the hugetlb code when they see a
> hugetlb VMA.
> 
> I think although it isn't strictly necessary, it is going to save a ton
> of work to just pass the VMAs around.

Well in any case, step 1 should be to finalize the interface. Let's
start with a single patch that does nothing but move the exported bits
around and makes room for the page shift bits. Are we agreed on what
that should look like now?

ps: my laptop died on Tuesday, so I'll be struggling to even read email
until its replacement arrives.

-- 
Mathematics is the supreme nostalgia of our time.