Saw need_resched() warnings when swapping on large swapfile (TBs)
because page allocation in swap_cgroup_prepare() took too long.
We already cond_resched when freeing page in swap_cgroup_swapoff().
Do the same for the page allocation.
Signed-off-by: Yu Zhao <[email protected]>
---
mm/swap_cgroup.c | 3 +++
1 file changed, 3 insertions(+)
diff --git a/mm/swap_cgroup.c b/mm/swap_cgroup.c
index ac6318a064d3..3405b4ee1757 100644
--- a/mm/swap_cgroup.c
+++ b/mm/swap_cgroup.c
@@ -48,6 +48,9 @@ static int swap_cgroup_prepare(int type)
if (!page)
goto not_enough_page;
ctrl->map[idx] = page;
+
+ if (!(idx % SWAP_CLUSTER_MAX))
+ cond_resched();
}
return 0;
not_enough_page:
--
2.13.0.219.gdb65acc882-goog
On Thu 01-06-17 12:56:35, Yu Zhao wrote:
> Saw need_resched() warnings when swapping on large swapfile (TBs)
> because page allocation in swap_cgroup_prepare() took too long.
Hmm, but the page allocator makes sure to cond_resched for sleeping
allocations. I guess what you mean is something different. It is not the
allocation which took too look but there are too many of them and none
of them sleeps because there is enough memory and the allocator doesn't
sleep in that case. Right?
> We already cond_resched when freeing page in swap_cgroup_swapoff().
> Do the same for the page allocation.
>
> Signed-off-by: Yu Zhao <[email protected]>
The patch itself makes sense to me, the changelog could see some
clarification but other than that
Acked-by: Michal Hocko <[email protected]>
> ---
> mm/swap_cgroup.c | 3 +++
> 1 file changed, 3 insertions(+)
>
> diff --git a/mm/swap_cgroup.c b/mm/swap_cgroup.c
> index ac6318a064d3..3405b4ee1757 100644
> --- a/mm/swap_cgroup.c
> +++ b/mm/swap_cgroup.c
> @@ -48,6 +48,9 @@ static int swap_cgroup_prepare(int type)
> if (!page)
> goto not_enough_page;
> ctrl->map[idx] = page;
> +
> + if (!(idx % SWAP_CLUSTER_MAX))
> + cond_resched();
> }
> return 0;
> not_enough_page:
> --
> 2.13.0.219.gdb65acc882-goog
--
Michal Hocko
SUSE Labs
On Thu, Jun 01, 2017 at 12:56:35PM -0700, Yu Zhao wrote:
> Saw need_resched() warnings when swapping on large swapfile (TBs)
> because page allocation in swap_cgroup_prepare() took too long.
>
> We already cond_resched when freeing page in swap_cgroup_swapoff().
> Do the same for the page allocation.
>
> Signed-off-by: Yu Zhao <[email protected]>
Acked-by: Vladimir Davydov <[email protected]>
> ---
> mm/swap_cgroup.c | 3 +++
> 1 file changed, 3 insertions(+)
>
> diff --git a/mm/swap_cgroup.c b/mm/swap_cgroup.c
> index ac6318a064d3..3405b4ee1757 100644
> --- a/mm/swap_cgroup.c
> +++ b/mm/swap_cgroup.c
> @@ -48,6 +48,9 @@ static int swap_cgroup_prepare(int type)
> if (!page)
> goto not_enough_page;
> ctrl->map[idx] = page;
> +
> + if (!(idx % SWAP_CLUSTER_MAX))
> + cond_resched();
> }
> return 0;
> not_enough_page:
We should probably do the same on the error path.
On Fri, Jun 02, 2017 at 10:18:57AM +0200, Michal Hocko wrote:
> On Thu 01-06-17 12:56:35, Yu Zhao wrote:
> > Saw need_resched() warnings when swapping on large swapfile (TBs)
> > because page allocation in swap_cgroup_prepare() took too long.
>
> Hmm, but the page allocator makes sure to cond_resched for sleeping
> allocations. I guess what you mean is something different. It is not the
> allocation which took too look but there are too many of them and none
> of them sleeps because there is enough memory and the allocator doesn't
> sleep in that case. Right?
>
> > We already cond_resched when freeing page in swap_cgroup_swapoff().
> > Do the same for the page allocation.
> >
> > Signed-off-by: Yu Zhao <[email protected]>
>
> The patch itself makes sense to me, the changelog could see some
> clarification but other than that
> Acked-by: Michal Hocko <[email protected]>
Thanks, I'll clarify the problem in the commit message and resend the
patch.
Saw need_resched() warnings when swapping on large swapfile (TBs)
because continuously allocating many pages in swap_cgroup_prepare()
took too long.
We already cond_resched when freeing page in swap_cgroup_swapoff().
Do the same for the page allocation.
Signed-off-by: Yu Zhao <[email protected]>
Acked-by: Michal Hocko <[email protected]>
Acked-by: Vladimir Davydov <[email protected]>
---
Changelog since v1:
* clarify the problem in the commit message
mm/swap_cgroup.c | 3 +++
1 file changed, 3 insertions(+)
diff --git a/mm/swap_cgroup.c b/mm/swap_cgroup.c
index ac6318a064d3..3405b4ee1757 100644
--- a/mm/swap_cgroup.c
+++ b/mm/swap_cgroup.c
@@ -48,6 +48,9 @@ static int swap_cgroup_prepare(int type)
if (!page)
goto not_enough_page;
ctrl->map[idx] = page;
+
+ if (!(idx % SWAP_CLUSTER_MAX))
+ cond_resched();
}
return 0;
not_enough_page:
--
2.13.0.506.g27d5fe0cd-goog
[CC Andrew]
On Sun 04-06-17 13:01:09, Yu Zhao wrote:
> Saw need_resched() warnings when swapping on large swapfile (TBs)
> because continuously allocating many pages in swap_cgroup_prepare()
> took too long.
>
> We already cond_resched when freeing page in swap_cgroup_swapoff().
> Do the same for the page allocation.
>
> Signed-off-by: Yu Zhao <[email protected]>
> Acked-by: Michal Hocko <[email protected]>
> Acked-by: Vladimir Davydov <[email protected]>
> ---
> Changelog since v1:
> * clarify the problem in the commit message
>
> mm/swap_cgroup.c | 3 +++
> 1 file changed, 3 insertions(+)
>
> diff --git a/mm/swap_cgroup.c b/mm/swap_cgroup.c
> index ac6318a064d3..3405b4ee1757 100644
> --- a/mm/swap_cgroup.c
> +++ b/mm/swap_cgroup.c
> @@ -48,6 +48,9 @@ static int swap_cgroup_prepare(int type)
> if (!page)
> goto not_enough_page;
> ctrl->map[idx] = page;
> +
> + if (!(idx % SWAP_CLUSTER_MAX))
> + cond_resched();
> }
> return 0;
> not_enough_page:
> --
> 2.13.0.506.g27d5fe0cd-goog
--
Michal Hocko
SUSE Labs