Re: [PATCH] cgroup/cpuset: check effective mems for a movable-only configuration

[Waiman Long]

On 10/22/26 12:39 AM, Gregory Price (Meta) wrote:
> check_insane_mems_config() warns and enables cpusets_insane_config_key
> when a cpuset's tasks can allocate only from movable memory.
>
> In cgroup v2 the request may name a memoryless node, which has no zones
> and cpusets reports it as a movable-only nodemask.
>
> For example, with node 2 a CPU-only node:
>
>    # echo +cpuset > /sys/fs/cgroup/cgroup.subtree_control
>    # mkdir /sys/fs/cgroup/test
>    # echo 2 > /sys/fs/cgroup/test/cpuset.mems
>    # cat /sys/fs/cgroup/test/cpuset.mems.effective
>      0-1
>    # dmesg
>      Unsupported (movable nodes only) cpuset configuration detected (nmask=2)!
>
> This is a bit confusing, especially consider mems.effective ends up with
> a valid nodemask by-design (cpuset.mems is treated as-if it were empty).
>
> Check each new effective mask in update_nodemasks_hier() instead, as the
> hotplug path already does. The allocator's bail-out tests the task's
> mems_allowed, which comes from the effective mask.
>
> A request naming a movable-only not present in the parent's nodemask
> no longer warns (since, again, the cpuset.mems gets overridden and
> treated as empty). However, it will warn if the parent's mems include
> the node later.
>
> For Example, with node 3 a movable-only node:
>
>    # echo 0 > /sys/fs/cgroup/test/cpuset.mems
>    # echo 3 > /sys/fs/cgroup/test/child/cpuset.mems
>    # cat /sys/fs/cgroup/test/child/cpuset.mems.effective
>      0
>    # dmesg
>      { empty }
>    # echo 0,3 > /sys/fs/cgroup/test/cpuset.mems
>    # cat /sys/fs/cgroup/test/child/cpuset.mems.effective
>      3
>    # dmesg
>      Unsupported (movable nodes only) cpuset configuration detected (nmask=3)!
>
> Fixes: 8ca1b5a49885 ("mm/page_alloc: detect allocation forbidden by cpuset and bail out early")
> Signed-off-by: Gregory Price (Meta) <gourry@xxxxxxxxxx>
> ---
>   kernel/cgroup/cpuset.c | 4 ++--
>   1 file changed, 2 insertions(+), 2 deletions(-)
>
> diff --git a/kernel/cgroup/cpuset.c b/kernel/cgroup/cpuset.c
> index 8ee239d5c6b48..70fbece1b9399 100644
> --- a/kernel/cgroup/cpuset.c
> +++ b/kernel/cgroup/cpuset.c
> @@ -2795,6 +2795,8 @@ static void update_nodemasks_hier(struct cpuset *cs, nodemask_t *new_mems)
>   			continue;
>   		rcu_read_unlock();
>   
> +		check_insane_mems_config(new_mems);
> +
>   		spin_lock_irq(&callback_lock);
>   		cp->effective_mems = *new_mems;
>   		spin_unlock_irq(&callback_lock);
> @@ -2848,8 +2850,6 @@ static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
>   	if (retval < 0)
>   		return retval;
>   
> -	check_insane_mems_config(&trialcs->mems_allowed);
> -
>   	spin_lock_irq(&callback_lock);
>   	cs->mems_allowed = trialcs->mems_allowed;
>   	spin_unlock_irq(&callback_lock);
Reviewed-by: Waiman Long <longman@xxxxxxxxxx>