Re: [RFC][PATCH 5/9] mm: Rename pmd_read_atomic()

[Jason Gunthorpe <jgg@ziepe.ca>]

On Mon, Nov 30, 2020 at 12:27:10PM +0100, Peter Zijlstra wrote:
> There's no point in having the identical routines for PTE/PMD have
> different names.
> 
> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
>  include/linux/pgtable.h    |    7 +------
>  mm/hmm.c                   |    2 +-
>  mm/mapping_dirty_helpers.c |    2 +-
>  mm/mprotect.c              |    2 +-
>  mm/userfaultfd.c           |    2 +-
>  5 files changed, 5 insertions(+), 10 deletions(-)
> 
> +++ b/include/linux/pgtable.h
> @@ -1244,11 +1244,6 @@ static inline int pud_trans_unstable(pud
>  #endif
>  }
>  
> -static inline pmd_t pmd_read_atomic(pmd_t *pmdp)
> -{
> -	return pmdp_get_lockless(pmdp);
> -}
> -
>  #ifndef arch_needs_pgtable_deposit
>  #define arch_needs_pgtable_deposit() (false)
>  #endif
> @@ -1275,7 +1270,7 @@ static inline pmd_t pmd_read_atomic(pmd_
>   */
>  static inline int pmd_none_or_trans_huge_or_clear_bad(pmd_t *pmd)
>  {
> -	pmd_t pmdval = pmd_read_atomic(pmd);
> +	pmd_t pmdval = pmdp_get_lockless(pmd);
>  	/*
>  	 * The barrier will stabilize the pmdval in a register or on
>  	 * the stack so that it will stop changing under the code.
> +++ b/mm/hmm.c
> @@ -356,7 +356,7 @@ static int hmm_vma_walk_pmd(pmd_t *pmdp,
>  		 * huge or device mapping one and compute corresponding pfn
>  		 * values.
>  		 */
> -		pmd = pmd_read_atomic(pmdp);
> +		pmd = pmdp_get_lockless(pmdp);
>  		barrier();
>  		if (!pmd_devmap(pmd) && !pmd_trans_huge(pmd))
>  			goto again;

The pagewalk API doesn't call the functions with interrupts disabled,
doesn't this mean we hit this assertion?

+#if CONFIG_PGTABLE_LEVELS > 2
+static inline pmd_t pmdp_get_lockless(pmd_t *pmdp)
+{
+       pmd_t pmd;
+
+       lockdep_assert_irqs_disabled();
+

It is only holding the read side of the mmap_sem here

Jason