Re: [PATCH 03/31] mm/pgtable: kmap_local_page() instead of kmap_atomic()

[Peter Xu]

On Fri, May 26, 2023 at 06:22:58PM -0400, Peter Xu wrote:
> On Sun, May 21, 2023 at 09:52:31PM -0700, Hugh Dickins wrote:
> > pte_offset_map() was still using kmap_atomic(): update it to the
> > preferred kmap_local_page() before making further changes there, in case
> > we need this as a bisection point; but I doubt it can cause any trouble.
> > 
> > Signed-off-by: Hugh Dickins <hughd@xxxxxxxxxx>
> > ---
> >  include/linux/pgtable.h | 4 ++--
> >  1 file changed, 2 insertions(+), 2 deletions(-)
> > 
> > diff --git a/include/linux/pgtable.h b/include/linux/pgtable.h
> > index 8ec27fe69dc8..94235ff2706e 100644
> > --- a/include/linux/pgtable.h
> > +++ b/include/linux/pgtable.h
> > @@ -96,9 +96,9 @@ static inline pte_t *pte_offset_kernel(pmd_t *pmd, unsigned long address)
> >  
> >  #if defined(CONFIG_HIGHPTE)
> >  #define pte_offset_map(dir, address)				\
> > -	((pte_t *)kmap_atomic(pmd_page(*(dir))) +		\
> > +	((pte_t *)kmap_local_page(pmd_page(*(dir))) +		\
> >  	 pte_index((address)))
> > -#define pte_unmap(pte) kunmap_atomic((pte))
> > +#define pte_unmap(pte) kunmap_local((pte))
> >  #else
> >  #define pte_offset_map(dir, address)	pte_offset_kernel((dir), (address))
> >  #define pte_unmap(pte) ((void)(pte))	/* NOP */
> 
> (I think this could be a dumb question if this patch has been running there
>  for years downstream, but still..)
> 
> I assume one major difference of using kmap_local() is page fault will not
> be disabled, while kmap_atomic() will.
> 
> Meanwhile, pte_offset_map() is also used by pte_offset_map_lock(), which
> means before this patch CONFIG_HIGHPTE systems will disable pgfault before
> taking pgtable lock for it, while it will stop doing so after the change.
> 
> Then what happens if a page fault happens on the same pgtable lock range
> that is already taken by the thread context?  What stops the deadlock from
> happening?

Ah, stupid me.  I think such a page fault just cannot happen when holding
the pgtable lock..  I believe the same applies to !HIGHPTE..

Sorry about the noise.

-- 
Peter Xu