Hi Alex,
I'm interested in the same issue, and posted patches a few hours ago
too (Cc:ed you.) It can be interesting/helpful for you.
On Fri, Aug 30, 2013 at 11:58:17AM -0500, Alex Thorlton wrote:
> This patch changes out the page_table_lock for the pud_page ptl in the
> THP fault path; pretty self-explanatory. I got lazy and commented out
> the spinlock assertion in follow_trans_huge_pmd instead of digging up
> the pud_page ptl in this function. This is just a proof of concept, so
> I didn't feel that it was too important to keep around for now.
>
> ---
> mm/huge_memory.c | 4 ++--
> mm/memory.c | 6 +++---
> 2 files changed, 5 insertions(+), 5 deletions(-)
>
> diff --git a/mm/huge_memory.c b/mm/huge_memory.c
> index a92012a..d3b34e2f 100644
> --- a/mm/huge_memory.c
> +++ b/mm/huge_memory.c
> @@ -1240,10 +1240,10 @@ struct page *follow_trans_huge_pmd(struct
> vm_area_struct *vma,
> pmd_t *pmd,
> unsigned int flags)
> {
> - struct mm_struct *mm = vma->vm_mm;
> +// struct mm_struct *mm = vma->vm_mm;
> struct page *page = NULL;
>
> - assert_spin_locked(&mm->page_table_lock);
> +// assert_spin_locked(&mm->page_table_lock);
>
> if (flags & FOLL_WRITE && !pmd_write(*pmd))
> goto out;
> diff --git a/mm/memory.c b/mm/memory.c
> index af84bc0..5b4e910 100644
> --- a/mm/memory.c
> +++ b/mm/memory.c
> @@ -1527,15 +1527,15 @@ struct page *follow_page_mask(struct vm_area_struct
> *vma,
> split_huge_page_pmd(vma, address, pmd);
> goto split_fallthrough;
> }
> - spin_lock(&mm->page_table_lock);
> + spin_lock(&pud_page(*pud)->ptl);
> if (likely(pmd_trans_huge(*pmd))) {
> if (unlikely(pmd_trans_splitting(*pmd))) {
> - spin_unlock(&mm->page_table_lock);
> + spin_unlock(&pud_page(*pud)->ptl);
> wait_split_huge_page(vma->anon_vma, pmd);
> } else {
> page = follow_trans_huge_pmd(vma, address,
> pmd, flags);
> - spin_unlock(&mm->page_table_lock);
> + spin_unlock(&pud_page(*pud)->ptl);
> *page_mask = HPAGE_PMD_NR - 1;
> goto out;
> }
I think that other ptl holders should use pud_page->ptl rather than
mm->page_table_lock. Otherwise we have a race that other threads can
change *pmd when running on this code.
Thanks,
Naoya Horiguchi
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
