pcpu_spin_unlock and pcpu_spin_unlock_irqrestore both unlock
pcp->lock and then enable preemption. This lacks symmetry against
both the pcpu_spin helpers and differs from how local_unlock_* is
implemented. While this is harmless, it's unnecessary and it's generally
better to unwind locks and preemption state in the reverse order as
they were acquired.

This is a fix on top of the mm-unstable patch
mm-page_alloc-replace-local_lock-with-normal-spinlock-fix.patch

Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
---
 mm/page_alloc.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/mm/page_alloc.c b/mm/page_alloc.c
index 934d1b5a5449..d0141e51e613 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c
@@ -192,14 +192,14 @@ static DEFINE_MUTEX(pcp_batch_high_lock);
 
 #define pcpu_spin_unlock(member, ptr)					\
 ({									\
-	spin_unlock(&ptr->member);					\
 	pcpu_task_unpin();						\
+	spin_unlock(&ptr->member);					\
 })
 
 #define pcpu_spin_unlock_irqrestore(member, ptr, flags)			\
 ({									\
-	spin_unlock_irqrestore(&ptr->member, flags);			\
 	pcpu_task_unpin();						\
+	spin_unlock_irqrestore(&ptr->member, flags);			\
 })
 
 /* struct per_cpu_pages specific helpers. */

On 7/8/22 16:44, Mel Gorman wrote:
> pcpu_spin_unlock and pcpu_spin_unlock_irqrestore both unlock
> pcp->lock and then enable preemption. This lacks symmetry against
> both the pcpu_spin helpers and differs from how local_unlock_* is
> implemented. While this is harmless, it's unnecessary and it's generally
> better to unwind locks and preemption state in the reverse order as
> they were acquired.

Hm I'm confused, it seems it's done in reverse order (which I agree with)
before this -fix-fix, but not after it?

before, pcpu_spin_lock() (and variants) do pcpu_task_pin() and then
spin_lock() (or variant), and pcpu_spin_unlock() does spin_unlock() and then
pcpu_task_unpin(). That seems symmetrical, i.e. reverse order to me? And
seems to match what local_lock family does too.

> This is a fix on top of the mm-unstable patch
> mm-page_alloc-replace-local_lock-with-normal-spinlock-fix.patch
> 
> Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
> ---
>  mm/page_alloc.c | 4 ++--
>  1 file changed, 2 insertions(+), 2 deletions(-)
> 
> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> index 934d1b5a5449..d0141e51e613 100644
> --- a/mm/page_alloc.c
> +++ b/mm/page_alloc.c
> @@ -192,14 +192,14 @@ static DEFINE_MUTEX(pcp_batch_high_lock);
>  
>  #define pcpu_spin_unlock(member, ptr)					\
>  ({									\
> -	spin_unlock(&ptr->member);					\
>  	pcpu_task_unpin();						\
> +	spin_unlock(&ptr->member);					\
>  })
>  
>  #define pcpu_spin_unlock_irqrestore(member, ptr, flags)			\
>  ({									\
> -	spin_unlock_irqrestore(&ptr->member, flags);			\
>  	pcpu_task_unpin();						\
> +	spin_unlock_irqrestore(&ptr->member, flags);			\
>  })
>  
>  /* struct per_cpu_pages specific helpers. */

On Fri, Jul 08, 2022 at 04:54:47PM +0200, Vlastimil Babka wrote:
> On 7/8/22 16:44, Mel Gorman wrote:
> > pcpu_spin_unlock and pcpu_spin_unlock_irqrestore both unlock
> > pcp->lock and then enable preemption. This lacks symmetry against
> > both the pcpu_spin helpers and differs from how local_unlock_* is
> > implemented. While this is harmless, it's unnecessary and it's generally
> > better to unwind locks and preemption state in the reverse order as
> > they were acquired.
> 
> Hm I'm confused, it seems it's done in reverse order (which I agree with)
> before this -fix-fix, but not after it?
> 
> before, pcpu_spin_lock() (and variants) do pcpu_task_pin() and then
> spin_lock() (or variant), and pcpu_spin_unlock() does spin_unlock() and then
> pcpu_task_unpin(). That seems symmetrical, i.e. reverse order to me? And
> seems to match what local_lock family does too.
> 

You're not confused, I am. The patch and the changelog are outright brain
damage from excessive context switching and a sign that it's time for the
weekend to start.

Sorry for this absolute misfortune.

-- 
Mel Gorman
SUSE Labs