[PATCH v2 1/2] mm: free compound page with correct order

[PATCH v2 1/2] mm: free compound page with correct order

[Yu Zhao]

Compound page should be freed by put_page() or free_pages() with
correct order. Not doing so will cause tail pages leaked.

The compound order can be obtained by compound_order() or use
HPAGE_PMD_ORDER in our case. Some people would argue the latter
is faster but I prefer the former which is more general.

Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Fixes: 97ae17497e99 ("thp: implement refcounting for huge zero page")
Cc: stable@vger.kernel.org (v3.8+)
Signed-off-by: Yu Zhao <yuzhao@google.com>
---
 mm/huge_memory.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/mm/huge_memory.c b/mm/huge_memory.c
index 74c78aa..780d12c 100644
--- a/mm/huge_memory.c
+++ b/mm/huge_memory.c
@@ -200,7 +200,7 @@ retry:
 	preempt_disable();
 	if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
 		preempt_enable();
-		__free_page(zero_page);
+		__free_pages(zero_page, compound_order(zero_page));
 		goto retry;
 	}
 
@@ -232,7 +232,7 @@ static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
 	if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
 		struct page *zero_page = xchg(&huge_zero_page, NULL);
 		BUG_ON(zero_page == NULL);
-		__free_page(zero_page);
+		__free_pages(zero_page, compound_order(zero_page));
 		return HPAGE_PMD_NR;
 	}
 
-- 
2.1.0.rc2.206.gedb03e5

[PATCH v2 2/2] mm: verify compound order when freeing a page

[Yu Zhao]

This allows us to catch the bug fixed in the previous patch
(mm: free compound page with correct order).

Here we also verify whether a page is tail page or not -- tail
pages are supposed to be freed along with their head, not by
themselves.

Reviewed-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Signed-off-by: Yu Zhao <yuzhao@google.com>
---
 mm/page_alloc.c | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/mm/page_alloc.c b/mm/page_alloc.c
index 736d8e1..5bf44e4 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c
@@ -750,6 +750,9 @@ static bool free_pages_prepare(struct page *page, unsigned int order)
 	int i;
 	int bad = 0;
 
+	VM_BUG_ON_PAGE(PageTail(page), page);
+	VM_BUG_ON_PAGE(PageHead(page) && compound_order(page) != order, page);
+
 	trace_mm_page_free(page, order);
 	kmemcheck_free_shadow(page, order);
 
-- 
2.1.0.rc2.206.gedb03e5

Re: [PATCH v2 1/2] mm: free compound page with correct order

[Andrew Morton]

On Wed, 15 Oct 2014 12:20:04 -0700 Yu Zhao <yuzhao@google.com> wrote:

> Compound page should be freed by put_page() or free_pages() with
> correct order. Not doing so will cause tail pages leaked.
> 
> The compound order can be obtained by compound_order() or use
> HPAGE_PMD_ORDER in our case. Some people would argue the latter
> is faster but I prefer the former which is more general.
> 
> Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
> Fixes: 97ae17497e99 ("thp: implement refcounting for huge zero page")
> Cc: stable@vger.kernel.org (v3.8+)

It's two years old and nobody noticed the memory leak, so presumably it
happens rarely.

> Signed-off-by: Yu Zhao <yuzhao@google.com>
> ---
>  mm/huge_memory.c | 4 ++--
>  1 file changed, 2 insertions(+), 2 deletions(-)
> 
> diff --git a/mm/huge_memory.c b/mm/huge_memory.c
> index 74c78aa..780d12c 100644
> --- a/mm/huge_memory.c
> +++ b/mm/huge_memory.c
> @@ -200,7 +200,7 @@ retry:
>  	preempt_disable();
>  	if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
>  		preempt_enable();
> -		__free_page(zero_page);
> +		__free_pages(zero_page, compound_order(zero_page));

This is rare.

>  		goto retry;
>  	}
>  
> @@ -232,7 +232,7 @@ static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
>  	if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
>  		struct page *zero_page = xchg(&huge_zero_page, NULL);
>  		BUG_ON(zero_page == NULL);
> -		__free_page(zero_page);
> +		__free_pages(zero_page, compound_order(zero_page));

But I'm surprised that this is also rare.  It makes me wonder if this
code is working correctly.

>  		return HPAGE_PMD_NR;
>  	}

Were you able to observe the leakage in practice?  If so, under what
circumstances?

Re: [PATCH v2 1/2] mm: free compound page with correct order

[Kirill A. Shutemov]

On Wed, Oct 15, 2014 at 12:30:44PM -0700, Andrew Morton wrote:
> > @@ -232,7 +232,7 @@ static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
> >  	if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
> >  		struct page *zero_page = xchg(&huge_zero_page, NULL);
> >  		BUG_ON(zero_page == NULL);
> > -		__free_page(zero_page);
> > +		__free_pages(zero_page, compound_order(zero_page));
> 
> But I'm surprised that this is also rare.  It makes me wonder if this
> code is working correctly.

This should be rare too. To get here we need a situation when huge zero
page is allocated, but not mapped and we get memory presure to trigger
shrinker.

-- 
 Kirill A. Shutemov

Re: [PATCH v2 2/2] mm: verify compound order when freeing a page

[Kirill A. Shutemov]

On Wed, Oct 15, 2014 at 12:20:05PM -0700, Yu Zhao wrote:
> This allows us to catch the bug fixed in the previous patch
> (mm: free compound page with correct order).
> 
> Here we also verify whether a page is tail page or not -- tail
> pages are supposed to be freed along with their head, not by
> themselves.
> 
> Reviewed-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

I didn't give this tag. It's okay in this case, but please do not assume
that you've got Reviewed-by, unless the person said it explicitly.

> Signed-off-by: Yu Zhao <yuzhao@google.com>

Your Singed-off-by should come first.

-- 
 Kirill A. Shutemov

Re: [PATCH v2 1/2] mm: free compound page with correct order

[Yu Zhao]

On Wed, Oct 15, 2014 at 12:30:44PM -0700, Andrew Morton wrote:
> On Wed, 15 Oct 2014 12:20:04 -0700 Yu Zhao <yuzhao@google.com> wrote:
> 
> > Compound page should be freed by put_page() or free_pages() with
> > correct order. Not doing so will cause tail pages leaked.
> > 
> > The compound order can be obtained by compound_order() or use
> > HPAGE_PMD_ORDER in our case. Some people would argue the latter
> > is faster but I prefer the former which is more general.
> > 
> > Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
> > Fixes: 97ae17497e99 ("thp: implement refcounting for huge zero page")
> > Cc: stable@vger.kernel.org (v3.8+)
> 
> It's two years old and nobody noticed the memory leak, so presumably it
> happens rarely.
> 
> > Signed-off-by: Yu Zhao <yuzhao@google.com>
> > ---
> >  mm/huge_memory.c | 4 ++--
> >  1 file changed, 2 insertions(+), 2 deletions(-)
> > 
> > diff --git a/mm/huge_memory.c b/mm/huge_memory.c
> > index 74c78aa..780d12c 100644
> > --- a/mm/huge_memory.c
> > +++ b/mm/huge_memory.c
> > @@ -200,7 +200,7 @@ retry:
> >  	preempt_disable();
> >  	if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
> >  		preempt_enable();
> > -		__free_page(zero_page);
> > +		__free_pages(zero_page, compound_order(zero_page));
> 
> This is rare.
> 
> >  		goto retry;
> >  	}
> >  
> > @@ -232,7 +232,7 @@ static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
> >  	if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
> >  		struct page *zero_page = xchg(&huge_zero_page, NULL);
> >  		BUG_ON(zero_page == NULL);
> > -		__free_page(zero_page);
> > +		__free_pages(zero_page, compound_order(zero_page));
> 
> But I'm surprised that this is also rare.  It makes me wonder if this
> code is working correctly.
> 
> >  		return HPAGE_PMD_NR;
> >  	}
> 
> Were you able to observe the leakage in practice?  If so, under what
> circumstances?

Yes, not just on our servers (the worst case we saw is 11G leaked on
a 48G machine) but also on our workstations running Ubuntu based distro.

$ cat /proc/vmstat  | grep thp_zero_page_alloc
thp_zero_page_alloc 55
thp_zero_page_alloc_failed 0

This means there is (thp_zero_page_alloc - 1) * (2M - 4K) memory leaked.