Re: [PATCH 8/9] mm: implement split page table lock for PMD level

[Naoya Horiguchi]

On Fri, Sep 13, 2013 at 04:06:15PM +0300, Kirill A. Shutemov wrote:
> The basic idea is the same as with PTE level: the lock is embedded into
> struct page of table's page.
> 
> Split pmd page table lock only makes sense on big machines.
> Let's say >= 32 CPUs for now.
> 
> We can't use mm->pmd_huge_pte to store pgtables for THP, since we don't
> take mm->page_table_lock anymore. Let's reuse page->lru of table's page
> for that.

Looks nice.

> hugetlbfs hasn't converted to split locking: disable split locking if
> hugetlbfs enabled.

I don't think that we have to disable when hugetlbfs is enabled,
because hugetlbfs code doesn't use huge_pmd_lockptr() or huge_pmd_lock().

> Signed-off-by: Naoya Horiguchi <n-horiguchi@xxxxxxxxxxxxx>
> Signed-off-by: Kirill A. Shutemov <kirill.shutemov@xxxxxxxxxxxxxxx>
> ---
>  include/linux/mm.h       | 31 +++++++++++++++++++++++++++++++
>  include/linux/mm_types.h |  5 +++++
>  kernel/fork.c            |  4 ++--
>  mm/Kconfig               | 10 ++++++++++
>  4 files changed, 48 insertions(+), 2 deletions(-)
> 
> diff --git a/include/linux/mm.h b/include/linux/mm.h
> index d2f8a50..5b3922d 100644
> --- a/include/linux/mm.h
> +++ b/include/linux/mm.h
> @@ -1294,13 +1294,44 @@ static inline void pgtable_page_dtor(struct page *page)
>  	((unlikely(pmd_none(*(pmd))) && __pte_alloc_kernel(pmd, address))? \
>  		NULL: pte_offset_kernel(pmd, address))
>  
> +#if USE_SPLIT_PMD_PTLOCKS
> +
> +static inline spinlock_t *huge_pmd_lockptr(struct mm_struct *mm, pmd_t *pmd)
> +{
> +	return &virt_to_page(pmd)->ptl;
> +}
> +
> +static inline void pgtable_pmd_page_ctor(struct page *page)
> +{
> +	spin_lock_init(&page->ptl);
> +#ifdef CONFIG_TRANSPARENT_HUGEPAGE
> +	page->pmd_huge_pte = NULL;
> +#endif
> +}
> +
> +static inline void pgtable_pmd_page_dtor(struct page *page)
> +{
> +#ifdef CONFIG_TRANSPARENT_HUGEPAGE
> +	VM_BUG_ON(page->pmd_huge_pte);
> +#endif
> +}
> +
> +#define pmd_huge_pte(mm, pmd) (virt_to_page(pmd)->pmd_huge_pte)
> 
> +
> +#else
> +
>  static inline spinlock_t *huge_pmd_lockptr(struct mm_struct *mm, pmd_t *pmd)
>  {
>  	return &mm->page_table_lock;
>  }
>  
> +static inline void pgtable_pmd_page_ctor(struct page *page) {}
> +static inline void pgtable_pmd_page_dtor(struct page *page) {}
> +
>  #define pmd_huge_pte(mm, pmd) ((mm)->pmd_huge_pte)
>  
> +#endif
> +
>  static inline spinlock_t *huge_pmd_lock(struct mm_struct *mm, pmd_t *pmd)
>  {
>  	spinlock_t *ptl = huge_pmd_lockptr(mm, pmd);
> diff --git a/include/linux/mm_types.h b/include/linux/mm_types.h
> index 1c64730..5706ddf 100644
> --- a/include/linux/mm_types.h
> +++ b/include/linux/mm_types.h
> @@ -24,6 +24,8 @@
>  struct address_space;
>  
>  #define USE_SPLIT_PTE_PTLOCKS	(NR_CPUS >= CONFIG_SPLIT_PTE_PTLOCK_CPUS)
> +#define USE_SPLIT_PMD_PTLOCKS	(USE_SPLIT_PTE_PTLOCKS && \
> +		NR_CPUS >= CONFIG_SPLIT_PMD_PTLOCK_CPUS)
>  
>  /*
>   * Each physical page in the system has a struct page associated with
> @@ -130,6 +132,9 @@ struct page {
>  
>  		struct list_head list;	/* slobs list of pages */
>  		struct slab *slab_page; /* slab fields */
> +#if defined(CONFIG_TRANSPARENT_HUGEPAGE) && USE_SPLIT_PMD_PTLOCKS
> +		pgtable_t pmd_huge_pte; /* protected by page->ptl */
> +#endif
>  	};
>  
>  	/* Remainder is not double word aligned */

Can we remove pmd_huge_pte from mm_struct when USE_SPLIT_PMD_PTLOCKS is true?

Thanks,
Naoya Horiguchi
--
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/