Re: [PATCH-next v2 2/6] cgroup/cpuset: Consider all the exclusive CPUs when doing housekeeping check

[Guopeng Zhang <guopeng.zhang@linux.dev>]

在 2026/10/11 06:19, Waiman Long 写道:
> When prstate_housekeeping_conflict() is called to perform housekeeping
> check with cpumask changes, the exclusive CPUs owned by child partitions
> are excluded. If the current cpuset is an isolated partition with root
> partition children. It is possible the cpumask change will still leave
> active housekeeping CPUs but then no housekeeping CPUs will be left when
> the child partitions are disabled or invalidated.
> 
> To protect against this possibility, all the exclusive CPUs of the
> cpuset should be considered to be owned by the current cpuset when doing
> housekeeping check. Do that by calling prstate_housekeeping_conflict()
> in partition_cpus_change() and passing in the add_cpus and del_cpus
> parameters as if the cpuset owns all the exclusive CPUs.
> 
> With this change, the prstate_housekeeping_conflict() call in
> update_parent_effective_cpumask() with partcmd_update and newmask
> is now duplicative and can be removed. The housekeeping check in
> remote_cpus_update(), however, will still be useful as this function can
> be called directly from update_cpumasks_hier() without going through
> partition_cpus_change(). The update_parent_effective_cpumask() call
> from update_cpumasks_hier() is partcmd_update with no cpumask which
> still have the housekeeping check and so is covered.
> 
> In the case of partition state switch from isolated to root and vice
> versa, the set of exclusive CPUs are recomputed to include those
> transferred to child paritions as well.
> 
> Fixes: 4a74e418881f ("cgroup/cpuset: Check partition conflict with housekeeping setup")
> Signed-off-by: Waiman Long <longman@redhat.com>
> ---
>  kernel/cgroup/cpuset.c | 26 +++++++++++++++++---------
>  1 file changed, 17 insertions(+), 9 deletions(-)
> 
> diff --git a/kernel/cgroup/cpuset.c b/kernel/cgroup/cpuset.c
> index f3cebb277a68..1c0441fa3bea 100644
> --- a/kernel/cgroup/cpuset.c
> +++ b/kernel/cgroup/cpuset.c
> @@ -1907,14 +1907,6 @@ static int update_parent_effective_cpumask(struct cpuset *cs, int cmd,
>  					       parent->effective_xcpus);
>  		}
>  
> -		/*
> -		 * Check for housekeeping conflicts
> -		 */
> -		if (is_partition_valid(cs) &&
> -		    prstate_housekeeping_conflict(old_prs, parent_prs,
> -						  tmp->delmask, tmp->addmask))
> -			part_error = PERR_HKEEPING;
> -
>  		/*
>  		 * The new CPUs to be removed from parent's effective CPUs
>  		 * must be present.
> @@ -2398,6 +2390,21 @@ static void partition_cpus_change(struct cpuset *cs, struct cpuset *trialcs,
>  		return;
>  
>  	prs_err = validate_partition(cs, trialcs);
> +	if (!prs_err) {
> +		int parent_prs = is_remote_partition(cs)
> +			       ? PRS_ROOT : parent_cs(cs)->partition_root_state;
> +		/*
> +		 * Check for housekeeping CPUs conflict assuming that all the
> +		 * exclusive CPUs belong to the current cpuset.
> +		 */
> +		compute_excpus(trialcs, tmp->new_cpus);
> +		cpumask_andnot(tmp->addmask, tmp->new_cpus, cs->effective_xcpus);
> +		cpumask_andnot(tmp->delmask, cs->effective_xcpus, tmp->new_cpus);

I ran a few local tests on v2 and still found the following gaps in the
housekeeping checks.

With CPUs 0-15 online and nohz_full=2-15:

    # cd /sys/fs/cgroup
    # echo +cpuset > cgroup.subtree_control
    # mkdir A
    # echo 1 > A/cpuset.cpus
    # echo isolated > A/cpuset.cpus.partition
    # echo 0-1 > A/cpuset.cpus
    # cat A/cpuset.cpus.partition
    isolated
    # cat A/cpuset.cpus.exclusive.effective
    0-1
    # cat cpuset.cpus.isolated
    0-1

The check computes delmask={1}, although CPU 1 is still isolated after the
update. That lets it isolate CPU 0, the last remaining housekeeping CPU.

On a fresh boot with the same nohz_full setting:

    # cd /sys/fs/cgroup
    # echo +cpuset > cgroup.subtree_control
    # mkdir A B
    # echo 1 > A/cpuset.cpus
    # echo 1 > A/cpuset.cpus.exclusive
    # echo isolated > A/cpuset.cpus.partition
    # echo +cpuset > A/cgroup.subtree_control
    # mkdir A/C
    # echo 1 > A/C/cpuset.cpus
    # echo root > A/C/cpuset.cpus.partition
    # echo 0 > B/cpuset.cpus
    # echo isolated > B/cpuset.cpus.partition
    # echo 1-2 > A/cpuset.cpus
    # echo 1-2 > A/cpuset.cpus.exclusive
    # cat cpuset.cpus.isolated
    0,2
    # echo member > A/C/cpuset.cpus.partition
    # cat A/cpuset.cpus.partition
    isolated
    # cat cpuset.cpus.isolated
    0-2

CPU 1 is present in both masks and cancels out of addmask while the root
child owns it. With nohz_full=2-15 and B isolating CPU 0, CPU 1 was the last
remaining housekeeping CPU; disabling the child leaves none.
This also reproduces before the series, so it is an existing issue.

> +		if (prstate_housekeeping_conflict(cs->partition_root_state,
> +						  parent_prs, tmp->addmask,
> +						  tmp->delmask))
> +			prs_err = PERR_HKEEPING;
> +	}

An invalid root with an explicit exclusive mask can retain effective_xcpus
without owning those CPUs. Subtracting this mask skips checking them when
the partition becomes valid again.

With CPUs 0-15 online and isolcpus=domain,10-11, on a fresh boot:

    # cd /sys/fs/cgroup
    # echo +cpuset > cgroup.subtree_control
    # mkdir A
    # echo 10 > A/cpuset.cpus
    # echo 10 > A/cpuset.cpus.exclusive
    # echo root > A/cpuset.cpus.partition
    # cat A/cpuset.cpus.partition
    root invalid (partition config conflicts with housekeeping setup)
    # cat A/cpuset.cpus.exclusive.effective
    10
    # echo 10,12 > A/cpuset.cpus
    # cat A/cpuset.cpus.partition
    root
    # cat A/cpuset.cpus.effective
    10
    # cat cpuset.cpus.effective
    0-9,11-15

addmask is empty here, so boot-isolated CPU 10 is allocated to the now-valid
root without being checked.

Should an invalid partition be checked against the full mask it will
acquire, rather than these deltas?

Thanks,
Guopeng
>  	if (prs_err) {
>  		WRITE_ONCE(cs->prs_err, prs_err);
>  		trialcs->prs_err = prs_err;
> @@ -2897,8 +2904,9 @@ static int update_prstate(struct cpuset *cs, int new_prs)
>  		 * A change in load balance state only, no change in cpumasks.
>  		 * Need to update isolated_cpus.
>  		 */
> +		compute_excpus(cs, tmpmask.new_cpus);
>  		if (prstate_housekeeping_conflict(new_prs, parent_prs,
> -						  cs->effective_xcpus, NULL))
> +						  tmpmask.new_cpus, NULL))
>  			err = PERR_HKEEPING;
>  		else
>  			isolcpus_updated = true;