Re: [PATCH 1/5] freepgt: free_pgtables use vma list

[Hugh Dickins <hugh@veritas.com>]

On Tue, 22 Mar 2005, David S. Miller wrote:
> On Tue, 22 Mar 2005 19:36:46 +0000 (GMT)
> Hugh Dickins <hugh@veritas.com> wrote:
> 
> > I notice that although both i386 and sparc64 use pgtable-nopud.h, the
> > i386 pud_clear does nothing at all and the sparc64 pud_clear resets to 0.
> 
> This was a dead end.  I386 doesn't do anything with pud_clear() in
> order to work around a chip erratum.
> 
> IA64 does clear in pud_clear() just like sparc64.

My mind kept flipping back and forth on whether it was pud_clear().
I agree, I can't see that it's the issue now.

> I think it's the floor/ceiling stuff.
> 
> At that pud_clear(), we do it if floor-->ceiling (after masking)
> covers the whole PUD.  Not whether start/end do, which is what
> the code sort of does right now.

Not an explanation I understood ;)

> "start" and "end" say which specific entries we might be purging.

Yes.

> "floor" and "ceiling" say that once that purging is done, the
> extent of the potential address space freed.

Well, they specify the limits beyond which we cannot free,
because there's other stuff still making use of addresses beyond.

> I cooked up a quick patch that changes the logic to:
> 
> 	floor &= PUD_MASK;
> 	ceiling &= PUD_MASK;
> 	if (floor - 1 >= ceiling - 1)
> 		return;

That can't be right.  In exit_mmap, for example, floor will be 0
throughout, and the condition will be true for all values of ceiling.

> 	pmd = pmd_offset(pud, start);
> 	pud_clear(pud);
> 	pmd_free_tlb(tlb, pmd);
> 
> and things start to basically work.

Because none of your higher level tables are being freed at all:
eventually you'll run out of memory.

I still can't see what's wrong with the code that's already
there.  My brain is seizing up, I'm taking a break.

Hugh