Re: [PATCH] cgroup/cpuset: Enable memory migration for cpuset v2

[Johannes Weiner <hannes@cmpxchg.org>]

On Wed, Aug 11, 2021 at 12:30:35PM -0400, Waiman Long wrote:
> When a user changes cpuset.cpus, each task in a v2 cpuset will be moved
> to one of the new cpus if it is not there already. For memory, however,
> they won't be migrated to the new nodes when cpuset.mems changes. This is
> an inconsistency in behavior.
> 
> In cpuset v1, there is a memory_migrate control file to enable such
> behavior by setting the CS_MEMORY_MIGRATE flag. Make it the default
> for cpuset v2 so that we have a consistent set of behavior for both
> cpus and memory.
> 
> There is certainly a cost to make memory migration the default, but it
> is a one time cost that shouldn't really matter as long as cpuset.mems
> isn't changed frequenty.  Update the cgroup-v2.rst file to document the
> new behavior and recommend against changing cpuset.mems frequently.
> 
> Since there won't be any concurrent access to the newly allocated cpuset
> structure in cpuset_css_alloc(), we can use the cheaper non-atomic
> __set_bit() instead of the more expensive atomic set_bit().
> 
> Signed-off-by: Waiman Long <longman@redhat.com>
> ---
>  Documentation/admin-guide/cgroup-v2.rst | 6 ++++++
>  kernel/cgroup/cpuset.c                  | 6 +++++-
>  2 files changed, 11 insertions(+), 1 deletion(-)
> 
> diff --git a/Documentation/admin-guide/cgroup-v2.rst b/Documentation/admin-guide/cgroup-v2.rst
> index 5c7377b5bd3e..47c832ad1322 100644
> --- a/Documentation/admin-guide/cgroup-v2.rst
> +++ b/Documentation/admin-guide/cgroup-v2.rst
> @@ -2056,6 +2056,12 @@ Cpuset Interface Files
>  	The value of "cpuset.mems" stays constant until the next update
>  	and won't be affected by any memory nodes hotplug events.
>  
> +	Setting a non-empty value to "cpuset.mems" causes memory of
> +	tasks within the cgroup to be migrated to the designated nodes if
> +	they are currently using memory outside of the designated nodes.
> +	There is a cost for this migration.  So "cpuset.mems" shouldn't
> +	be changed frequently.

The migration skips over pages that are (temporarily) off the LRU for
reclaim, compaction etc. so it can leave random pages behind.

In practice it's probably fine, but it probably makes sense to say
it's advisable to set this config once before the workload launches
for best results, and not rely too much on changing things around
post-hoc, due to cost you pointed out but also due to reliability.

Otherwise no objection from me.