Re: Shared futexes (was [PATCH] FUTEX : new PRIVATE futexes)

[Peter Zijlstra <a.p.zijlstra@chello.nl>]

On Fri, 2007-04-06 at 14:02 +0100, Hugh Dickins wrote:
> On Fri, 6 Apr 2007, Peter Zijlstra wrote:
> > 
> > some thoughts on shared futexes;
> > 
> > Could we get rid of the mmap_sem on the shared futexes in the following
> > manner:
> > 
> >  - do a page table walk to find the pte;
> 
> ("walk" meaning descent down the levels, I presume, rather than across)

indeed.

> I've not had time to digest your proposal, and I'm about to go out:
> let me sound a warning that springs to mind, maybe it's entirely
> inapproriate, but better said than kept silent.
> 
> It looks as if you're supposing that mmap_sem is needed to find_vma,
> but not for going down the pagetables.  It's not a simple as that:
> you need to be careful that a concurrent munmap from another thread
> isn't freeing pagetables from under you.
> 
> Holding (down_read) of mmap_sem is one way to protect against that.
> try_to_unmap doesn't have that luxury: in its case, it's made safe
> by the way free_pgtables does anon_vma_unlink and unlink_file_vma
> before freeing any pagetables, so try_to_unmap etc. won't get there;
> but you can't do that.

Ah, drad.

I had hoped that the pte_lock would solve that race, but that doesn't
cover the upper levels of the tree.

Back to the drawing board..