mm: migrate.c: migrate PG_readahead flag
diff mbox series

Message ID 1581640185-95731-1-git-send-email-yang.shi@linux.alibaba.com
State In Next
Commit 3f8931fe4d984a613f2abe3befe85bc7213624cb
Headers show
Series
  • mm: migrate.c: migrate PG_readahead flag
Related show

Commit Message

Yang Shi Feb. 14, 2020, 12:29 a.m. UTC
Currently migration code doesn't migrate PG_readahead flag.
Theoretically this would incur slight performance loss as the
application might have to ramp its readahead back up again.  Even though
such problem happens, it might be hidden by something else since
migration is typically triggered by compaction and NUMA balancing, any
of which should be more noticeable.

Migrate the flag after end_page_writeback() since it may clear
PG_reclaim flag, which is the same bit as PG_readahead, for the new
page.

Cc: Matthew Wilcox <willy@infradead.org>
Cc: Michal Hocko <mhocko@suse.com>
Signed-off-by: Yang Shi <yang.shi@linux.alibaba.com>
---
I didn't experience any real problem, found by visual inspection. And, this was
discussed in thread: https://lore.kernel.org/linux-mm/185ce762-f25d-a013-6daa-8c288f1ff791@linux.alibaba.com/T/#m1977ce1de513401b7d09d6fa14fcffe849580aae

 mm/migrate.c | 8 ++++++++
 1 file changed, 8 insertions(+)

Comments

Andrew Morton Feb. 14, 2020, 2:55 a.m. UTC | #1
On Fri, 14 Feb 2020 08:29:45 +0800 Yang Shi <yang.shi@linux.alibaba.com> wrote:

> Currently migration code doesn't migrate PG_readahead flag.
> Theoretically this would incur slight performance loss as the
> application might have to ramp its readahead back up again.  Even though
> such problem happens, it might be hidden by something else since
> migration is typically triggered by compaction and NUMA balancing, any
> of which should be more noticeable.
> 
> Migrate the flag after end_page_writeback() since it may clear
> PG_reclaim flag, which is the same bit as PG_readahead, for the new
> page.
> 
> --- a/mm/migrate.c
> +++ b/mm/migrate.c
> @@ -647,6 +647,14 @@ void migrate_page_states(struct page *newpage, struct page *page)
>  	if (PageWriteback(newpage))
>  		end_page_writeback(newpage);
>  
> +	/*
> +	 * PG_readahead share the same bit with PG_reclaim, the above
> +	 * end_page_writeback() may clear PG_readahead mistakenly, so set
> +	 * the bit after that.
> +	 */
> +	if (PageReadahead(page))
> +		SetPageReadahead(newpage);
> +
>  	copy_page_owner(page, newpage);
>  

Why not

  	if (PageWriteback(newpage)) {
  		end_page_writeback(newpage);
		/*
		 * PG_readahead share the same bit with PG_reclaim, the above
		 * end_page_writeback() may clear PG_readahead mistakenly, so
		 * set the bit after that.
		 */
		if (PageReadahead(page))
			SetPageReadahead(newpage);
	}

?
Yang Shi Feb. 14, 2020, 3:58 a.m. UTC | #2
On 2/13/20 6:55 PM, Andrew Morton wrote:
> On Fri, 14 Feb 2020 08:29:45 +0800 Yang Shi <yang.shi@linux.alibaba.com> wrote:
>
>> Currently migration code doesn't migrate PG_readahead flag.
>> Theoretically this would incur slight performance loss as the
>> application might have to ramp its readahead back up again.  Even though
>> such problem happens, it might be hidden by something else since
>> migration is typically triggered by compaction and NUMA balancing, any
>> of which should be more noticeable.
>>
>> Migrate the flag after end_page_writeback() since it may clear
>> PG_reclaim flag, which is the same bit as PG_readahead, for the new
>> page.
>>
>> --- a/mm/migrate.c
>> +++ b/mm/migrate.c
>> @@ -647,6 +647,14 @@ void migrate_page_states(struct page *newpage, struct page *page)
>>   	if (PageWriteback(newpage))
>>   		end_page_writeback(newpage);
>>   
>> +	/*
>> +	 * PG_readahead share the same bit with PG_reclaim, the above
>> +	 * end_page_writeback() may clear PG_readahead mistakenly, so set
>> +	 * the bit after that.
>> +	 */
>> +	if (PageReadahead(page))
>> +		SetPageReadahead(newpage);
>> +
>>   	copy_page_owner(page, newpage);
>>   
> Why not

The newpage may not have writeback set, migrating readahead flag should 
not depend on it.

>
>    	if (PageWriteback(newpage)) {
>    		end_page_writeback(newpage);
> 		/*
> 		 * PG_readahead share the same bit with PG_reclaim, the above
> 		 * end_page_writeback() may clear PG_readahead mistakenly, so
> 		 * set the bit after that.
> 		 */
> 		if (PageReadahead(page))
> 			SetPageReadahead(newpage);
> 	}
>
> ?
Matthew Wilcox Feb. 14, 2020, 4:07 a.m. UTC | #3
On Thu, Feb 13, 2020 at 07:58:40PM -0800, Yang Shi wrote:
> On 2/13/20 6:55 PM, Andrew Morton wrote:
> > On Fri, 14 Feb 2020 08:29:45 +0800 Yang Shi <yang.shi@linux.alibaba.com> wrote:
> > > Currently migration code doesn't migrate PG_readahead flag.
> > > Theoretically this would incur slight performance loss as the
> > > application might have to ramp its readahead back up again.  Even though
> > > such problem happens, it might be hidden by something else since
> > > migration is typically triggered by compaction and NUMA balancing, any
> > > of which should be more noticeable.
> > > 
> > > Migrate the flag after end_page_writeback() since it may clear
> > > PG_reclaim flag, which is the same bit as PG_readahead, for the new
> > > page.
> > > 
> > > --- a/mm/migrate.c
> > > +++ b/mm/migrate.c
> > > @@ -647,6 +647,14 @@ void migrate_page_states(struct page *newpage, struct page *page)
> > >   	if (PageWriteback(newpage))
> > >   		end_page_writeback(newpage);
> > > +	/*
> > > +	 * PG_readahead share the same bit with PG_reclaim, the above
> > > +	 * end_page_writeback() may clear PG_readahead mistakenly, so set
> > > +	 * the bit after that.
> > > +	 */
> > > +	if (PageReadahead(page))
> > > +		SetPageReadahead(newpage);
> > > +
> > >   	copy_page_owner(page, newpage);
> > Why not
> 
> The newpage may not have writeback set, migrating readahead flag should not
> depend on it.

Indeed, if the page has writeback set, then the page does not have the
readahead flag set; it has the reclaim flag set.  The original patch is
correct, afaict.

> >    	if (PageWriteback(newpage)) {
> >    		end_page_writeback(newpage);
> > 		/*
> > 		 * PG_readahead share the same bit with PG_reclaim, the above
> > 		 * end_page_writeback() may clear PG_readahead mistakenly, so
> > 		 * set the bit after that.
> > 		 */
> > 		if (PageReadahead(page))
> > 			SetPageReadahead(newpage);
> > 	}
> > 
> > ?
>
Michal Hocko Feb. 14, 2020, 8:34 a.m. UTC | #4
On Fri 14-02-20 08:29:45, Yang Shi wrote:
> Currently migration code doesn't migrate PG_readahead flag.
> Theoretically this would incur slight performance loss as the
> application might have to ramp its readahead back up again.  Even though
> such problem happens, it might be hidden by something else since
> migration is typically triggered by compaction and NUMA balancing, any
> of which should be more noticeable.
> 
> Migrate the flag after end_page_writeback() since it may clear
> PG_reclaim flag, which is the same bit as PG_readahead, for the new
> page.

Looks like an omission. The readahead flag has been added later (2.6.23)
while the migration predates 2.6.17.

> Cc: Matthew Wilcox <willy@infradead.org>
> Cc: Michal Hocko <mhocko@suse.com>
> Signed-off-by: Yang Shi <yang.shi@linux.alibaba.com>

Acked-by: Michal Hocko <mhocko@suse.com>

> ---
> I didn't experience any real problem, found by visual inspection. And, this was
> discussed in thread: https://lore.kernel.org/linux-mm/185ce762-f25d-a013-6daa-8c288f1ff791@linux.alibaba.com/T/#m1977ce1de513401b7d09d6fa14fcffe849580aae
> 
>  mm/migrate.c | 8 ++++++++
>  1 file changed, 8 insertions(+)
> 
> diff --git a/mm/migrate.c b/mm/migrate.c
> index edf42ed..f3c492d 100644
> --- a/mm/migrate.c
> +++ b/mm/migrate.c
> @@ -647,6 +647,14 @@ void migrate_page_states(struct page *newpage, struct page *page)
>  	if (PageWriteback(newpage))
>  		end_page_writeback(newpage);
>  
> +	/*
> +	 * PG_readahead share the same bit with PG_reclaim, the above
> +	 * end_page_writeback() may clear PG_readahead mistakenly, so set
> +	 * the bit after that.
> +	 */
> +	if (PageReadahead(page))
> +		SetPageReadahead(newpage);
> +
>  	copy_page_owner(page, newpage);
>  
>  	mem_cgroup_migrate(page, newpage);
> -- 
> 1.8.3.1
David Hildenbrand Feb. 14, 2020, 10:04 a.m. UTC | #5
On 14.02.20 01:29, Yang Shi wrote:
> Currently migration code doesn't migrate PG_readahead flag.
> Theoretically this would incur slight performance loss as the
> application might have to ramp its readahead back up again.  Even though
> such problem happens, it might be hidden by something else since
> migration is typically triggered by compaction and NUMA balancing, any
> of which should be more noticeable.
> 
> Migrate the flag after end_page_writeback() since it may clear
> PG_reclaim flag, which is the same bit as PG_readahead, for the new
> page.
> 
> Cc: Matthew Wilcox <willy@infradead.org>
> Cc: Michal Hocko <mhocko@suse.com>
> Signed-off-by: Yang Shi <yang.shi@linux.alibaba.com>
> ---
> I didn't experience any real problem, found by visual inspection. And, this was
> discussed in thread: https://lore.kernel.org/linux-mm/185ce762-f25d-a013-6daa-8c288f1ff791@linux.alibaba.com/T/#m1977ce1de513401b7d09d6fa14fcffe849580aae
> 
>  mm/migrate.c | 8 ++++++++
>  1 file changed, 8 insertions(+)
> 
> diff --git a/mm/migrate.c b/mm/migrate.c
> index edf42ed..f3c492d 100644
> --- a/mm/migrate.c
> +++ b/mm/migrate.c
> @@ -647,6 +647,14 @@ void migrate_page_states(struct page *newpage, struct page *page)
>  	if (PageWriteback(newpage))
>  		end_page_writeback(newpage);
>  
> +	/*
> +	 * PG_readahead share the same bit with PG_reclaim, the above
> +	 * end_page_writeback() may clear PG_readahead mistakenly, so set
> +	 * the bit after that.
> +	 */
> +	if (PageReadahead(page))
> +		SetPageReadahead(newpage);
> +
>  	copy_page_owner(page, newpage);
>  
>  	mem_cgroup_migrate(page, newpage);
> 

Looks good to me!

Reviewed-by: David Hildenbrand <david@redhat.com>

Patch
diff mbox series

diff --git a/mm/migrate.c b/mm/migrate.c
index edf42ed..f3c492d 100644
--- a/mm/migrate.c
+++ b/mm/migrate.c
@@ -647,6 +647,14 @@  void migrate_page_states(struct page *newpage, struct page *page)
 	if (PageWriteback(newpage))
 		end_page_writeback(newpage);
 
+	/*
+	 * PG_readahead share the same bit with PG_reclaim, the above
+	 * end_page_writeback() may clear PG_readahead mistakenly, so set
+	 * the bit after that.
+	 */
+	if (PageReadahead(page))
+		SetPageReadahead(newpage);
+
 	copy_page_owner(page, newpage);
 
 	mem_cgroup_migrate(page, newpage);