[PATCH] mm: workingset: clarify eviction order and distance calculation

[PATCH] mm: workingset: clarify eviction order and distance calculation

[Oscar Salvador]

The premise of the refault distance is that it can be seen as a deficit
of the inactive list space, so that if the inactive list would have had
(R - E) more slots, the page would not have been evicted but promoted
to the active list instead.

However, the way the code is ordered right now set us to be off by one,
so the real number of slots would be (R - E) + 1.
I stumbled upon this when trying to understand the code and it puzzled me
that the comments did not match what the code did.

This it not an issue at all since evictions and refaults tend to happen
in a number large enough that being off-by-one does not have any impact
- and since the compiler and CPUs are free to rearrange the execution
sequence anyway.
But as Johannes says, it is better to re-arrange the code in the proper
order since otherwise would be misleading to somebody who is actively
reading and trying to understand the logic of the code - like it
happened to me.

Signed-off-by: Oscar Salvador <osalvador@suse.de>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
---
 mm/workingset.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/mm/workingset.c b/mm/workingset.c
index 10e96de945b3..0201aa1ff320 100644
--- a/mm/workingset.c
+++ b/mm/workingset.c
@@ -263,10 +263,10 @@ void *workingset_eviction(struct page *page, struct mem_cgroup *target_memcg)
 	VM_BUG_ON_PAGE(!PageLocked(page), page);
 
 	lruvec = mem_cgroup_lruvec(target_memcg, pgdat);
-	workingset_age_nonresident(lruvec, thp_nr_pages(page));
 	/* XXX: target_memcg can be NULL, go through lruvec */
 	memcgid = mem_cgroup_id(lruvec_memcg(lruvec));
 	eviction = atomic_long_read(&lruvec->nonresident_age);
+	workingset_age_nonresident(lruvec, thp_nr_pages(page));
 	return pack_shadow(memcgid, pgdat, eviction, PageWorkingset(page));
 }
 
-- 
2.26.2

Re: [PATCH] mm: workingset: clarify eviction order and distance calculation

[Vlastimil Babka]

On 2/1/21 7:06 AM, Oscar Salvador wrote:
> The premise of the refault distance is that it can be seen as a deficit
> of the inactive list space, so that if the inactive list would have had
> (R - E) more slots, the page would not have been evicted but promoted
> to the active list instead.
> 
> However, the way the code is ordered right now set us to be off by one,
> so the real number of slots would be (R - E) + 1.
> I stumbled upon this when trying to understand the code and it puzzled me
> that the comments did not match what the code did.
> 
> This it not an issue at all since evictions and refaults tend to happen
> in a number large enough that being off-by-one does not have any impact
> - and since the compiler and CPUs are free to rearrange the execution
> sequence anyway.
> But as Johannes says, it is better to re-arrange the code in the proper
> order since otherwise would be misleading to somebody who is actively
> reading and trying to understand the logic of the code - like it
> happened to me.
> 
> Signed-off-by: Oscar Salvador <osalvador@suse.de>
> Acked-by: Johannes Weiner <hannes@cmpxchg.org>

Acked-by: Vlastimil Babka <vbabka@suse.cz>

> ---
>  mm/workingset.c | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
> 
> diff --git a/mm/workingset.c b/mm/workingset.c
> index 10e96de945b3..0201aa1ff320 100644
> --- a/mm/workingset.c
> +++ b/mm/workingset.c
> @@ -263,10 +263,10 @@ void *workingset_eviction(struct page *page, struct mem_cgroup *target_memcg)
>  	VM_BUG_ON_PAGE(!PageLocked(page), page);
>  
>  	lruvec = mem_cgroup_lruvec(target_memcg, pgdat);
> -	workingset_age_nonresident(lruvec, thp_nr_pages(page));
>  	/* XXX: target_memcg can be NULL, go through lruvec */
>  	memcgid = mem_cgroup_id(lruvec_memcg(lruvec));
>  	eviction = atomic_long_read(&lruvec->nonresident_age);
> +	workingset_age_nonresident(lruvec, thp_nr_pages(page));
>  	return pack_shadow(memcgid, pgdat, eviction, PageWorkingset(page));
>  }
>  
> 

Re: [PATCH] mm: workingset: clarify eviction order and distance calculation

[Oscar Salvador]

On Mon, Feb 01, 2021 at 07:06:51AM +0100, Oscar Salvador wrote:
> The premise of the refault distance is that it can be seen as a deficit
> of the inactive list space, so that if the inactive list would have had
> (R - E) more slots, the page would not have been evicted but promoted
> to the active list instead.
> 
> However, the way the code is ordered right now set us to be off by one,
> so the real number of slots would be (R - E) + 1.
> I stumbled upon this when trying to understand the code and it puzzled me
> that the comments did not match what the code did.
> 
> This it not an issue at all since evictions and refaults tend to happen
> in a number large enough that being off-by-one does not have any impact
> - and since the compiler and CPUs are free to rearrange the execution
> sequence anyway.
> But as Johannes says, it is better to re-arrange the code in the proper
> order since otherwise would be misleading to somebody who is actively
> reading and trying to understand the logic of the code - like it
> happened to me.
> 
> Signed-off-by: Oscar Salvador <osalvador@suse.de>
> Acked-by: Johannes Weiner <hannes@cmpxchg.org>

Hi Andrew,

is this on your radar?

Thanks!

-- 
Oscar Salvador
SUSE L3

Re: [PATCH] mm: workingset: clarify eviction order and distance calculation

[Oscar Salvador]

On Thu, Feb 11, 2021 at 10:26:45PM +0100, Oscar Salvador wrote:
> Hi Andrew,
> 
> is this on your radar?

Please, disregard this, I was obviously blind as I did not spot it
in mmotm.

> 
> Thanks!
> 
> -- 
> Oscar Salvador
> SUSE L3
> 

-- 
Oscar Salvador
SUSE L3