Re: [PATCH 00/21] mm: Preemptibility -v6

[Andrea Arcangeli]

On Wed, Jan 19, 2011 at 06:10:39PM +0100, Peter Zijlstra wrote:
> > 19/21 mm-convert_i_mmap_lock_and_anon_vma-_lock_to_mutexes.patch
> >       I suggest doing the anon_vma lock->mutex conversion separately here.
> >       Acked-by: Hugh Dickins <hughd@xxxxxxxxxx>
> >       except that in the past we have renamed a lock when we've done this
> >       kind of conversion, so I'd expect anon_vma->mutex throughout now.
> >       Or am I just out of date?  I don't feel very strongly about it.
> 
> Done.. however:
> 
> Index: linux-2.6/include/linux/huge_mm.h
> ===================================================================
> --- linux-2.6.orig/include/linux/huge_mm.h
> +++ linux-2.6/include/linux/huge_mm.h
> @@ -91,12 +91,8 @@ extern void __split_huge_page_pmd(struct
>  #define wait_split_huge_page(__anon_vma, __pmd)				\
>  	do {								\
>  		pmd_t *____pmd = (__pmd);				\
> -		spin_unlock_wait(&(__anon_vma)->root->lock);		\
> -		/*							\
> -		 * spin_unlock_wait() is just a loop in C and so the	\
> -		 * CPU can reorder anything around it.			\
> -		 */							\
> -		smp_mb();						\
> +		anon_vma_lock(__anon_vma);				\
> +		anon_vma_unlock(__anon_vma);				\
>  		BUG_ON(pmd_trans_splitting(*____pmd) ||			\
>  		       pmd_trans_huge(*____pmd));			\
>  	} while (0)
> 
> Andrea, is that smp_mb() simply to avoid us doing anything before the
> lock is free? Why isn't there an mb() before to ensure nothing leaks
> past it from the other end?

The idea would be that by the time we read the pmd set to
pmd_trans_splitting with the page_table_lock hold, we're guaranteed
we'll see the anon-vma locked (if it's still locked). So it's ok if
spin_unlock_wait happens before reading the pmd_trans_splitting check
inside the CPU (all it matters is for it not to happen before the
spin_lock(&page_table_lock) which it can't by the acquire semantics of
the spinlock).

So in short we know we start with the anon_vma locked, and we just
wait as long as needed.

So we only need to protect to the stuff after spin_unlock_wait().
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/