[PATCH] mm: swap: avoid to writepage when a page is !PageSwapCache

[PATCH] mm: swap: avoid to writepage when a page is !PageSwapCache

[Liu Ping Fan]

There is race between do_swap_page() and swap_writepage(), if
do_swap_page() had deleted a page from swap cache, there is no need
to write it. So changing the ret of try_to_free_swap() to make
swap_writepage() aware of this scene.

Signed-off-by: Liu Ping Fan <pingfank@linux.vnet.ibm.com>
---
 mm/swapfile.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/mm/swapfile.c b/mm/swapfile.c
index 4c524f7..9d80671 100644
--- a/mm/swapfile.c
+++ b/mm/swapfile.c
@@ -910,7 +910,7 @@ int try_to_free_swap(struct page *page)
 	VM_BUG_ON_PAGE(!PageLocked(page), page);
 
 	if (!PageSwapCache(page))
-		return 0;
+		return -1;
 	if (PageWriteback(page))
 		return 0;
 	if (page_swapcount(page))
-- 
1.8.1.4

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: [PATCH] mm: swap: avoid to writepage when a page is !PageSwapCache

[Hugh Dickins]

On Wed, 2 Jul 2014, Liu Ping Fan wrote:

> There is race between do_swap_page() and swap_writepage(), if
> do_swap_page() had deleted a page from swap cache, there is no need
> to write it. So changing the ret of try_to_free_swap() to make
> swap_writepage() aware of this scene.

Is this an inefficiency that you have noticed in practice,
or something that you think you spotted by code inspection?

I don't see how it can happen: all the places I know of that call
swap_writepage() (including vmscan.c's mapping->a_ops->writepage)
have not dropped page lock since setting or checking PageSwapCache,
and page lock is supposed to protect against deletion from swap cache.

Has that changed?  Please point out where.

> 
> Signed-off-by: Liu Ping Fan <pingfank@linux.vnet.ibm.com>
> ---
>  mm/swapfile.c | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
> 
> diff --git a/mm/swapfile.c b/mm/swapfile.c
> index 4c524f7..9d80671 100644
> --- a/mm/swapfile.c
> +++ b/mm/swapfile.c
> @@ -910,7 +910,7 @@ int try_to_free_swap(struct page *page)
>  	VM_BUG_ON_PAGE(!PageLocked(page), page);
>  
>  	if (!PageSwapCache(page))
> -		return 0;
> +		return -1;

Previously it returned either 0 or 1, which is what __try_to_reclaim_swap()
says it returns; so better to stick to 0 or 1, unless you have good reason
to add a distinct value.

It's true that by the time __try_to_reclaim_swap() has got the page lock,
the page might have been removed from swap cache, and we could then treat
that as swap_was_freed (even though it was not freed by the caller).

But it's a very narrow window, and no great advantage to do so:
I don't think it's worth changing try_to_free_swap() semantics for,
but you could persuade us.

Hugh

>  	if (PageWriteback(page))
>  		return 0;
>  	if (page_swapcount(page))
> -- 
> 1.8.1.4

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: [PATCH] mm: swap: avoid to writepage when a page is !PageSwapCache

[Liu ping fan]

On Wed, Jul 2, 2014 at 12:29 PM, Hugh Dickins <hughd@google.com> wrote:
> On Wed, 2 Jul 2014, Liu Ping Fan wrote:
>
>> There is race between do_swap_page() and swap_writepage(), if
>> do_swap_page() had deleted a page from swap cache, there is no need
>> to write it. So changing the ret of try_to_free_swap() to make
>> swap_writepage() aware of this scene.
>
> Is this an inefficiency that you have noticed in practice,
> or something that you think you spotted by code inspection?
>

just  spotted by code inspection.
> I don't see how it can happen: all the places I know of that call
> swap_writepage() (including vmscan.c's mapping->a_ops->writepage)
> have not dropped page lock since setting or checking PageSwapCache,
> and page lock is supposed to protect against deletion from swap cache.
>
> Has that changed?  Please point out where.
>

No, my fault.  Thanks for making me aware of this.
>>
>> Signed-off-by: Liu Ping Fan <pingfank@linux.vnet.ibm.com>
>> ---
>>  mm/swapfile.c | 2 +-
>>  1 file changed, 1 insertion(+), 1 deletion(-)
>>
>> diff --git a/mm/swapfile.c b/mm/swapfile.c
>> index 4c524f7..9d80671 100644
>> --- a/mm/swapfile.c
>> +++ b/mm/swapfile.c
>> @@ -910,7 +910,7 @@ int try_to_free_swap(struct page *page)
>>       VM_BUG_ON_PAGE(!PageLocked(page), page);
>>
>>       if (!PageSwapCache(page))
>> -             return 0;
>> +             return -1;
>
> Previously it returned either 0 or 1, which is what __try_to_reclaim_swap()
> says it returns; so better to stick to 0 or 1, unless you have good reason
> to add a distinct value.
>
> It's true that by the time __try_to_reclaim_swap() has got the page lock,
> the page might have been removed from swap cache, and we could then treat
> that as swap_was_freed (even though it was not freed by the caller).
>
> But it's a very narrow window, and no great advantage to do so:
> I don't think it's worth changing try_to_free_swap() semantics for,
> but you could persuade us.
>

Got it, and it is meaningless to do that.

Thanks,
Fan

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>