On Thu, Sep 05, 2013 at 02:48:18PM +0530, Aneesh Kumar K.V wrote:
> Naoya Horiguchi <n-horigu...@ah.jp.nec.com> writes:
>
> > Hi Aneesh,
> >
> > On Wed, Sep 04, 2013 at 12:43:19PM +0530, Aneesh Kumar K.V wrote:
> >> Naoya Horiguchi <n-horigu...@ah.jp.nec.com> writes:
> >>
> >> > Currently all of page table handling by hugetlbfs code are done under
> >> > mm->page_table_lock. So when a process have many threads and they heavily
> >> > access to the memory, lock contention happens and impacts the
> >> > performance.
> >> >
> >> > This patch makes hugepage support split page table lock so that we use
> >> > page->ptl of the leaf node of page table tree which is pte for normal
> >> > pages
> >> > but can be pmd and/or pud for hugepages of some architectures.
> >> >
> >> > ChangeLog v2:
> >> > - add split ptl on other archs missed in v1
> >> >
> >> > Signed-off-by: Naoya Horiguchi <n-horigu...@ah.jp.nec.com>
> >> > ---
> >> > arch/powerpc/mm/hugetlbpage.c | 6 ++-
> >> > arch/tile/mm/hugetlbpage.c | 6 ++-
> >> > include/linux/hugetlb.h | 20 ++++++++++
> >> > mm/hugetlb.c | 92
> >> > ++++++++++++++++++++++++++-----------------
> >> > mm/mempolicy.c | 5 ++-
> >> > mm/migrate.c | 4 +-
> >> > mm/rmap.c | 2 +-
> >> > 7 files changed, 90 insertions(+), 45 deletions(-)
> >> >
> >> > diff --git v3.11-rc3.orig/arch/powerpc/mm/hugetlbpage.c
> >> > v3.11-rc3/arch/powerpc/mm/hugetlbpage.c
> >> > index d67db4b..7e56cb7 100644
> >> > --- v3.11-rc3.orig/arch/powerpc/mm/hugetlbpage.c
> >> > +++ v3.11-rc3/arch/powerpc/mm/hugetlbpage.c
> >> > @@ -124,6 +124,7 @@ static int __hugepte_alloc(struct mm_struct *mm,
> >> > hugepd_t *hpdp,
> >> > {
> >> > struct kmem_cache *cachep;
> >> > pte_t *new;
> >> > + spinlock_t *ptl;
> >> >
> >> > #ifdef CONFIG_PPC_FSL_BOOK3E
> >> > int i;
> >> > @@ -141,7 +142,8 @@ static int __hugepte_alloc(struct mm_struct *mm,
> >> > hugepd_t *hpdp,
> >> > if (! new)
> >> > return -ENOMEM;
> >> >
> >> > - spin_lock(&mm->page_table_lock);
> >> > + ptl = huge_pte_lockptr(mm, new);
> >> > + spin_lock(ptl);
> >>
> >>
> >> Are you sure we can do that for ppc ?
> >> new = kmem_cache_zalloc(cachep, GFP_KERNEL|__GFP_REPEAT);
> >
> > Ah, thanks. new is not a pointer to one full page occupied by page
> > table entries, so trying to use struct page of it is totally wrong.
> >
> >> The page for new(pte_t) could be shared right ? which mean a deadlock ?
> >
> > Yes, that's disastrous.
> >
> >> May be you should do it at the pmd level itself for ppc
>
> The pgd page also cannot be used because pgd also comes from kmem
> cache.
>
> >
> > Yes, that's possible, but I simply drop the changes in __hugepte_alloc()
> > for now because this lock seems to protect us from the race between
> > concurrent
> > calls of __hugepte_alloc(), not between allocation and read/write access.
> > Split ptl is used to avoid race between read/write accesses, so I think
> > that using different types of locks here is not dangerous.
> > # I guess that that's why we now use mm->page_table_lock for __pte_alloc()
> > # and its family even if USE_SPLIT_PTLOCKS is true.
>
> A simpler approach could be to make huge_pte_lockptr arch
> specific and leave it as mm->page_table_lock for ppc
OK, I'll do this.
Thanks,
Naoya
--
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/
