[v2] mm: memcg: Use larger batches for proactive reclaim

  Before 388536ac291 ("mm:vmscan: fix inaccurate reclaim during proactive
reclaim") we passed the number of pages for the reclaim request directly
to try_to_free_mem_cgroup_pages, which could lead to significant
overreclaim. After 0388536ac291 the number of pages was limited to a
maximum 32 (SWAP_CLUSTER_MAX) to reduce the amount of overreclaim.
However such a small batch size caused a regression in reclaim
performance due to many more reclaim start/stop cycles inside
memory_reclaim.

Reclaim tries to balance nr_to_reclaim fidelity with fairness across
nodes and cgroups over which the pages are spread. As such, the bigger
the request, the bigger the absolute overreclaim error. Historic
in-kernel users of reclaim have used fixed, small sized requests to
approach an appropriate reclaim rate over time. When we reclaim a user
request of arbitrary size, use decaying batch sizes to manage error while
maintaining reasonable throughput.

root - full reclaim       pages/sec   time (sec)
pre-0388536ac291      :    68047        10.46
post-0388536ac291     :    13742        inf
(reclaim-reclaimed)/4 :    67352        10.51

/uid_0 - 1G reclaim       pages/sec   time (sec)  overreclaim (MiB)
pre-0388536ac291      :    258822       1.12            107.8
post-0388536ac291     :    105174       2.49            3.5
(reclaim-reclaimed)/4 :    233396       1.12            -7.4

/uid_0 - full reclaim     pages/sec   time (sec)
pre-0388536ac291      :    72334        7.09
post-0388536ac291     :    38105        14.45
(reclaim-reclaimed)/4 :    72914        6.96

Fixes: 0388536ac291 ("mm:vmscan: fix inaccurate reclaim during proactive reclaim")
Signed-off-by: T.J. Mercier <tjmercier@google.com>

---
v2: Simplify the request size calculation per Johannes Weiner and Michal Koutný

 mm/memcontrol.c | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)
  

On Fri, Feb 2, 2024 at 2:10 PM T.J. Mercier <tjmercier@google.com> wrote:
>
> Before 388536ac291 ("mm:vmscan: fix inaccurate reclaim during proactive
> reclaim") we passed the number of pages for the reclaim request directly
> to try_to_free_mem_cgroup_pages, which could lead to significant
> overreclaim. After 0388536ac291 the number of pages was limited to a
> maximum 32 (SWAP_CLUSTER_MAX) to reduce the amount of overreclaim.
> However such a small batch size caused a regression in reclaim
> performance due to many more reclaim start/stop cycles inside
> memory_reclaim.
>
> Reclaim tries to balance nr_to_reclaim fidelity with fairness across
> nodes and cgroups over which the pages are spread. As such, the bigger
> the request, the bigger the absolute overreclaim error. Historic
> in-kernel users of reclaim have used fixed, small sized requests to
> approach an appropriate reclaim rate over time. When we reclaim a user
> request of arbitrary size, use decaying batch sizes to manage error while
> maintaining reasonable throughput.
>
> root - full reclaim       pages/sec   time (sec)
> pre-0388536ac291      :    68047        10.46
> post-0388536ac291     :    13742        inf
> (reclaim-reclaimed)/4 :    67352        10.51
>
> /uid_0 - 1G reclaim       pages/sec   time (sec)  overreclaim (MiB)
> pre-0388536ac291      :    258822       1.12            107.8
> post-0388536ac291     :    105174       2.49            3.5
> (reclaim-reclaimed)/4 :    233396       1.12            -7.4
>
> /uid_0 - full reclaim     pages/sec   time (sec)
> pre-0388536ac291      :    72334        7.09
> post-0388536ac291     :    38105        14.45
> (reclaim-reclaimed)/4 :    72914        6.96
>
> Fixes: 0388536ac291 ("mm:vmscan: fix inaccurate reclaim during proactive reclaim")
> Signed-off-by: T.J. Mercier <tjmercier@google.com>

LGTM with a nit below:
Reviewed-by: Yosry Ahmed <yosryahmed@google.com>

>
> ---
> v2: Simplify the request size calculation per Johannes Weiner and Michal Koutný
>
>  mm/memcontrol.c | 5 ++++-
>  1 file changed, 4 insertions(+), 1 deletion(-)
>
> diff --git a/mm/memcontrol.c b/mm/memcontrol.c
> index 46d8d02114cf..e6f921555e07 100644
> --- a/mm/memcontrol.c
> +++ b/mm/memcontrol.c
> @@ -6965,6 +6965,9 @@ static ssize_t memory_reclaim(struct kernfs_open_file *of, char *buf,
>         while (nr_reclaimed < nr_to_reclaim) {
>                 unsigned long reclaimed;
>
> +               /* Will converge on zero, but reclaim enforces a minimum */
> +               unsigned long batch_size = (nr_to_reclaim - nr_reclaimed) / 4;
> +
>                 if (signal_pending(current))
>                         return -EINTR;
>
> @@ -6977,7 +6980,7 @@ static ssize_t memory_reclaim(struct kernfs_open_file *of, char *buf,
>                         lru_add_drain_all();
>
>                 reclaimed = try_to_free_mem_cgroup_pages(memcg,
> -                                       min(nr_to_reclaim - nr_reclaimed, SWAP_CLUSTER_MAX),
> +                                       batch_size,
>                                         GFP_KERNEL, reclaim_options);

I think the above two lines should now fit into one.
  

On Fri, Feb 2, 2024 at 2:14 PM Yosry Ahmed <yosryahmed@google.com> wrote:
>
> On Fri, Feb 2, 2024 at 2:10 PM T.J. Mercier <tjmercier@google.com> wrote:
> >
> > Before 388536ac291 ("mm:vmscan: fix inaccurate reclaim during proactive
> > reclaim") we passed the number of pages for the reclaim request directly
> > to try_to_free_mem_cgroup_pages, which could lead to significant
> > overreclaim. After 0388536ac291 the number of pages was limited to a
> > maximum 32 (SWAP_CLUSTER_MAX) to reduce the amount of overreclaim.
> > However such a small batch size caused a regression in reclaim
> > performance due to many more reclaim start/stop cycles inside
> > memory_reclaim.
> >
> > Reclaim tries to balance nr_to_reclaim fidelity with fairness across
> > nodes and cgroups over which the pages are spread. As such, the bigger
> > the request, the bigger the absolute overreclaim error. Historic
> > in-kernel users of reclaim have used fixed, small sized requests to
> > approach an appropriate reclaim rate over time. When we reclaim a user
> > request of arbitrary size, use decaying batch sizes to manage error while
> > maintaining reasonable throughput.
> >
> > root - full reclaim       pages/sec   time (sec)
> > pre-0388536ac291      :    68047        10.46
> > post-0388536ac291     :    13742        inf
> > (reclaim-reclaimed)/4 :    67352        10.51
> >
> > /uid_0 - 1G reclaim       pages/sec   time (sec)  overreclaim (MiB)
> > pre-0388536ac291      :    258822       1.12            107.8
> > post-0388536ac291     :    105174       2.49            3.5
> > (reclaim-reclaimed)/4 :    233396       1.12            -7.4
> >
> > /uid_0 - full reclaim     pages/sec   time (sec)
> > pre-0388536ac291      :    72334        7.09
> > post-0388536ac291     :    38105        14.45
> > (reclaim-reclaimed)/4 :    72914        6.96
> >
> > Fixes: 0388536ac291 ("mm:vmscan: fix inaccurate reclaim during proactive reclaim")
> > Signed-off-by: T.J. Mercier <tjmercier@google.com>
>
> LGTM with a nit below:
> Reviewed-by: Yosry Ahmed <yosryahmed@google.com>

Thanks

> >
> > ---
> > v2: Simplify the request size calculation per Johannes Weiner and Michal Koutný
> >
> >  mm/memcontrol.c | 5 ++++-
> >  1 file changed, 4 insertions(+), 1 deletion(-)
> >
> > diff --git a/mm/memcontrol.c b/mm/memcontrol.c
> > index 46d8d02114cf..e6f921555e07 100644
> > --- a/mm/memcontrol.c
> > +++ b/mm/memcontrol.c
> > @@ -6965,6 +6965,9 @@ static ssize_t memory_reclaim(struct kernfs_open_file *of, char *buf,
> >         while (nr_reclaimed < nr_to_reclaim) {
> >                 unsigned long reclaimed;
> >
> > +               /* Will converge on zero, but reclaim enforces a minimum */
> > +               unsigned long batch_size = (nr_to_reclaim - nr_reclaimed) / 4;
> > +
> >                 if (signal_pending(current))
> >                         return -EINTR;
> >
> > @@ -6977,7 +6980,7 @@ static ssize_t memory_reclaim(struct kernfs_open_file *of, char *buf,
> >                         lru_add_drain_all();
> >
> >                 reclaimed = try_to_free_mem_cgroup_pages(memcg,
> > -                                       min(nr_to_reclaim - nr_reclaimed, SWAP_CLUSTER_MAX),
> > +                                       batch_size,
> >                                         GFP_KERNEL, reclaim_options);
>
> I think the above two lines should now fit into one.

It goes out to 81 characters. I wasn't brave enough, even though the
80 char limit is no more. :)

This takes it out to 100 but gets rid of batch_size if folks are ok with it:

                reclaimed = try_to_free_mem_cgroup_pages(memcg,
-                                       min(nr_to_reclaim -
nr_reclaimed, SWAP_CLUSTER_MAX),
+                                       /* Will converge on zero, but
reclaim enforces a minimum */
+                                       (nr_to_reclaim - nr_reclaimed) / 4,
                                        GFP_KERNEL, reclaim_options);
  

> > > diff --git a/mm/memcontrol.c b/mm/memcontrol.c
> > > index 46d8d02114cf..e6f921555e07 100644
> > > --- a/mm/memcontrol.c
> > > +++ b/mm/memcontrol.c
> > > @@ -6965,6 +6965,9 @@ static ssize_t memory_reclaim(struct kernfs_open_file *of, char *buf,
> > >         while (nr_reclaimed < nr_to_reclaim) {
> > >                 unsigned long reclaimed;
> > >
> > > +               /* Will converge on zero, but reclaim enforces a minimum */
> > > +               unsigned long batch_size = (nr_to_reclaim - nr_reclaimed) / 4;
> > > +

I think it's clearer with no blank lines between declarations. Perhaps
add these two lines right above the declaration of "reclaimed"?

> > >                 if (signal_pending(current))
> > >                         return -EINTR;
> > >
> > > @@ -6977,7 +6980,7 @@ static ssize_t memory_reclaim(struct kernfs_open_file *of, char *buf,
> > >                         lru_add_drain_all();
> > >
> > >                 reclaimed = try_to_free_mem_cgroup_pages(memcg,
> > > -                                       min(nr_to_reclaim - nr_reclaimed, SWAP_CLUSTER_MAX),
> > > +                                       batch_size,
> > >                                         GFP_KERNEL, reclaim_options);
> >
> > I think the above two lines should now fit into one.
>
> It goes out to 81 characters. I wasn't brave enough, even though the
> 80 char limit is no more. :)

Oh okay, I would leave it as-is or rename batch_size to something
slightly shorter. Not a big deal either way. Going to 81 chars is
probably fine too.
  

On Fri, Feb 02, 2024 at 02:13:20PM -0800, Yosry Ahmed wrote:
> On Fri, Feb 2, 2024 at 2:10 PM T.J. Mercier <tjmercier@google.com> wrote:
> > @@ -6965,6 +6965,9 @@ static ssize_t memory_reclaim(struct kernfs_open_file *of, char *buf,
> >         while (nr_reclaimed < nr_to_reclaim) {
> >                 unsigned long reclaimed;
> >
> > +               /* Will converge on zero, but reclaim enforces a minimum */
> > +               unsigned long batch_size = (nr_to_reclaim - nr_reclaimed) / 4;
> > +
> >                 if (signal_pending(current))
> >                         return -EINTR;
> >
> > @@ -6977,7 +6980,7 @@ static ssize_t memory_reclaim(struct kernfs_open_file *of, char *buf,
> >                         lru_add_drain_all();
> >
> >                 reclaimed = try_to_free_mem_cgroup_pages(memcg,
> > -                                       min(nr_to_reclaim - nr_reclaimed, SWAP_CLUSTER_MAX),
> > +                                       batch_size,
> >                                         GFP_KERNEL, reclaim_options);
> 
> I think the above two lines should now fit into one.

Yeah might as well compact that again. The newline in the declarations
is a bit unusual for this codebase as well, and puts the comment sort
of away from the "reclaim" it refers to. This?

		/* Will converge on zero, but reclaim enforces a minimum */
		batch_size = (nr_to_reclaim - nr_reclaimed) / 4;

		reclaimed = try_to_free_mem_cgroup_pages(memcg, batch_size,
					GFP_KERNEL, reclaim_options);

But agreed, it's all just nitpickety nickpicking. :)

Acked-by: Johannes Weiner <hannes@cmpxchg.org>
  

On Fri, Feb 2, 2024 at 2:41 PM Johannes Weiner <hannes@cmpxchg.org> wrote:
>
> On Fri, Feb 02, 2024 at 02:13:20PM -0800, Yosry Ahmed wrote:
> > On Fri, Feb 2, 2024 at 2:10 PM T.J. Mercier <tjmercier@google.com> wrote:
> > > @@ -6965,6 +6965,9 @@ static ssize_t memory_reclaim(struct kernfs_open_file *of, char *buf,
> > >         while (nr_reclaimed < nr_to_reclaim) {
> > >                 unsigned long reclaimed;
> > >
> > > +               /* Will converge on zero, but reclaim enforces a minimum */
> > > +               unsigned long batch_size = (nr_to_reclaim - nr_reclaimed) / 4;
> > > +
> > >                 if (signal_pending(current))
> > >                         return -EINTR;
> > >
> > > @@ -6977,7 +6980,7 @@ static ssize_t memory_reclaim(struct kernfs_open_file *of, char *buf,
> > >                         lru_add_drain_all();
> > >
> > >                 reclaimed = try_to_free_mem_cgroup_pages(memcg,
> > > -                                       min(nr_to_reclaim - nr_reclaimed, SWAP_CLUSTER_MAX),
> > > +                                       batch_size,
> > >                                         GFP_KERNEL, reclaim_options);
> >
> > I think the above two lines should now fit into one.
>
> Yeah might as well compact that again. The newline in the declarations
> is a bit unusual for this codebase as well, and puts the comment sort
> of away from the "reclaim" it refers to. This?
>
>                 /* Will converge on zero, but reclaim enforces a minimum */
>                 batch_size = (nr_to_reclaim - nr_reclaimed) / 4;
>
>                 reclaimed = try_to_free_mem_cgroup_pages(memcg, batch_size,
>                                         GFP_KERNEL, reclaim_options);
>
> But agreed, it's all just nitpickety nickpicking. :)
>
> Acked-by: Johannes Weiner <hannes@cmpxchg.org>

-std=gnu11 to the rescue

+               /* Will converge on zero, but reclaim enforces a minimum */
+               unsigned long batch_size = (nr_to_reclaim - nr_reclaimed) / 4;
                reclaimed = try_to_free_mem_cgroup_pages(memcg,
-                                       min(nr_to_reclaim -
nr_reclaimed, SWAP_CLUSTER_MAX),
+                                       batch_size,
                                        GFP_KERNEL, reclaim_options);