[PATCH] mm, page_alloc: only use per-cpu allocator for irq-safe requests -fix

[PATCH] mm, page_alloc: only use per-cpu allocator for irq-safe requests -fix

[Mel Gorman]

preempt_enable_no_resched() was used based on review feedback that had no
strong objection at the time. It avoided introducing a preemption point
where one didn't exist before which was marginal at best.

However, it is hazardous to the RT tree according to Thomas Gleixner
and is a violation of its expected use according to Peter Zijlstra. In
Peter's own words "the only acceptable use of preempt_enable_no_resched()
is if the next statement is a schedule() variant".

The impact of using preempt_enable in this particular
fast path is negligible. This is a fix to the mmotm patch
mm-page_alloc-only-use-per-cpu-allocator-for-irq-safe-requests.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 eaecb4b145e6..2a36dad03dac 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c
@@ -2520,7 +2520,7 @@ void free_hot_cold_page(struct page *page, bool cold)
 	}
 
 out:
-	preempt_enable_no_resched();
+	preempt_enable();
 }
 
 /*
@@ -2686,7 +2686,7 @@ static struct page *rmqueue_pcplist(struct zone *preferred_zone,
 		__count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
 		zone_statistics(preferred_zone, zone);
 	}
-	preempt_enable_no_resched();
+	preempt_enable();
 	return page;
 }
 

[PATCH] mm, page_alloc: only use per-cpu allocator for irq-safe requests -fix

[Mel Gorman]

preempt_enable_no_resched() was used based on review feedback that had no
strong objection at the time. It avoided introducing a preemption point
where one didn't exist before which was marginal at best.

However, it is hazardous to the RT tree according to Thomas Gleixner
and is a violation of its expected use according to Peter Zijlstra. In
Peter's own words "the only acceptable use of preempt_enable_no_resched()
is if the next statement is a schedule() variant".

The impact of using preempt_enable in this particular
fast path is negligible. This is a fix to the mmotm patch
mm-page_alloc-only-use-per-cpu-allocator-for-irq-safe-requests.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 eaecb4b145e6..2a36dad03dac 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c
@@ -2520,7 +2520,7 @@ void free_hot_cold_page(struct page *page, bool cold)
 	}
 
 out:
-	preempt_enable_no_resched();
+	preempt_enable();
 }
 
 /*
@@ -2686,7 +2686,7 @@ static struct page *rmqueue_pcplist(struct zone *preferred_zone,
 		__count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
 		zone_statistics(preferred_zone, zone);
 	}
-	preempt_enable_no_resched();
+	preempt_enable();
 	return page;
 }
 

--
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, page_alloc: only use per-cpu allocator for irq-safe requests -fix

[Thomas Gleixner]

On Wed, 8 Feb 2017, Mel Gorman wrote:

> preempt_enable_no_resched() was used based on review feedback that had no
> strong objection at the time. It avoided introducing a preemption point
> where one didn't exist before which was marginal at best.

Actually local_irq_enable() _IS_ a preemption point, indirect but still:

   local_irq_disable()
   ....
--> HW interrupt is raised
   ....
   local_irq_enable()

   handle_irq()
	set_need_resched()
   ret_from_irq()
     preempt()

while with preempt_disable that looks like this:

   preempt_disable()
   ....
--> HW interrupt is raised
   handle_irq()
	set_need_resched()
   ret_from_irq()
   ....
   preempt_enable()
      preempt()

Now if you use preempt_enable_no_resched() then you miss the preemption and
depending on the actual code path you might run something which takes ages
without hitting a preemption point after that.

It's not only a problem for RT. It's also in mainline a violation of the
preemption mechanism.

Thanks,

	tglx

Re: [PATCH] mm, page_alloc: only use per-cpu allocator for irq-safe requests -fix

[Thomas Gleixner]

On Wed, 8 Feb 2017, Mel Gorman wrote:

> preempt_enable_no_resched() was used based on review feedback that had no
> strong objection at the time. It avoided introducing a preemption point
> where one didn't exist before which was marginal at best.

Actually local_irq_enable() _IS_ a preemption point, indirect but still:

   local_irq_disable()
   ....
--> HW interrupt is raised
   ....
   local_irq_enable()

   handle_irq()
	set_need_resched()
   ret_from_irq()
     preempt()

while with preempt_disable that looks like this:

   preempt_disable()
   ....
--> HW interrupt is raised
   handle_irq()
	set_need_resched()
   ret_from_irq()
   ....
   preempt_enable()
      preempt()

Now if you use preempt_enable_no_resched() then you miss the preemption and
depending on the actual code path you might run something which takes ages
without hitting a preemption point after that.

It's not only a problem for RT. It's also in mainline a violation of the
preemption mechanism.

Thanks,

	tglx



--
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, page_alloc: only use per-cpu allocator for irq-safe requests -fix

[Mel Gorman]

On Wed, Feb 08, 2017 at 03:56:22PM +0100, Thomas Gleixner wrote:
> On Wed, 8 Feb 2017, Mel Gorman wrote:
> 
> > preempt_enable_no_resched() was used based on review feedback that had no
> > strong objection at the time. It avoided introducing a preemption point
> > where one didn't exist before which was marginal at best.
> 
> Actually local_irq_enable() _IS_ a preemption point, indirect but still:
> 
>    local_irq_disable()
>    ....
> --> HW interrupt is raised
>    ....
>    local_irq_enable()
> 
>    handle_irq()
> 	set_need_resched()
>    ret_from_irq()
>      preempt()
> 
> while with preempt_disable that looks like this:
> 
>    preempt_disable()
>    ....
> --> HW interrupt is raised
>    handle_irq()
> 	set_need_resched()
>    ret_from_irq()
>    ....
>    preempt_enable()
>       preempt()
> 
> Now if you use preempt_enable_no_resched() then you miss the preemption and
> depending on the actual code path you might run something which takes ages
> without hitting a preemption point after that.
> 

Thanks for the education, I had missed it. The changelog should have been
"fix a dumb mistake and stick to preempt_enable".  Assuming Andrew picks
this patch up, it'll be folded into the patch that introduced the problem
in the first place and will the broken usage will never hit mainline.

> It's not only a problem for RT. It's also in mainline a violation of the
> preemption mechanism.
> 

Understood, thanks.

-- 
Mel Gorman
SUSE Labs

Re: [PATCH] mm, page_alloc: only use per-cpu allocator for irq-safe requests -fix

[Mel Gorman]

On Wed, Feb 08, 2017 at 03:56:22PM +0100, Thomas Gleixner wrote:
> On Wed, 8 Feb 2017, Mel Gorman wrote:
> 
> > preempt_enable_no_resched() was used based on review feedback that had no
> > strong objection at the time. It avoided introducing a preemption point
> > where one didn't exist before which was marginal at best.
> 
> Actually local_irq_enable() _IS_ a preemption point, indirect but still:
> 
>    local_irq_disable()
>    ....
> --> HW interrupt is raised
>    ....
>    local_irq_enable()
> 
>    handle_irq()
> 	set_need_resched()
>    ret_from_irq()
>      preempt()
> 
> while with preempt_disable that looks like this:
> 
>    preempt_disable()
>    ....
> --> HW interrupt is raised
>    handle_irq()
> 	set_need_resched()
>    ret_from_irq()
>    ....
>    preempt_enable()
>       preempt()
> 
> Now if you use preempt_enable_no_resched() then you miss the preemption and
> depending on the actual code path you might run something which takes ages
> without hitting a preemption point after that.
> 

Thanks for the education, I had missed it. The changelog should have been
"fix a dumb mistake and stick to preempt_enable".  Assuming Andrew picks
this patch up, it'll be folded into the patch that introduced the problem
in the first place and will the broken usage will never hit mainline.

> It's not only a problem for RT. It's also in mainline a violation of the
> preemption mechanism.
> 

Understood, thanks.

-- 
Mel Gorman
SUSE Labs

--
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>