Re: [PATCH-next 2/6] cgroup/cpuset: Properly compute effective xcpus in update_parent_effective_cpumask()

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

Hi Longman,

在 2026/10/22 11:16, Waiman Long 写道:
> Since the introduction of cpuset.cpus.exclusive in commit
> e2ffe502ba45 ("cgroup/cpuset: Add cpuset.cpus.exclusive for v2"), a
> compute_effective_exclusive_cpumask() helper which is later renamed to
> compute_excpus() is added to compute the effective xcpus of a partition
> root. However update_parent_effective_cpumask() is not updated in all
> proper places to use the new helper to check if a given cpuset should
> remain a valid partition root. This can cause a cpuset to incorrectly
> remain a valid partition root even if no effective xcpus is left.
> 
> Fix that problem by using compute_excpus() to compute the effective xcpus
> into delmask in all cases except partcmd_disable and partcmd_invalidate
> and use it for tasks_nocpu_error() check as well as checking its
> emptiness. This enables us to catch invalid partition root more
> accurately in update_parent_effective_cpumask().
> 

I re-tested the repro from the link below on top of this series
(applied on cgroup/for-7.4-fixes), and it still fails:

  https://lore.kernel.org/all/20261010075947.333549-3-guopeng.zhang@linux.dev/

  cd /sys/fs/cgroup
  echo +cpuset > cgroup.subtree_control
  mkdir A
  echo 0-7 > A/cpuset.cpus
  echo 0-7 > A/cpuset.cpus.exclusive
  echo root > A/cpuset.cpus.partition
  cat A/cpuset.cpus.partition
  echo 1 > A/cpuset.cpus.exclusive
  cat A/cpuset.cpus.partition

The last command still shows "root invalid (Parent unable to distribute
cpu downstream)".

In this case compute_excpus() returns the stale user mask unchanged.
user_xcpus() is gone from update_parent_effective_cpumask(), but the
dependency just moved one level down: the new code calls
compute_excpus(cs, tmp->delmask), and its first line is

  cpumask_and(excpus, user_xcpus(cs), parent->effective_xcpus)

Here user_xcpus(cs) is cs->exclusive_cpus, which still holds the old
0-7 when compute_excpus() runs: update_exclusive_cpumask() publishes
cs->exclusive_cpus only after partition_cpus_change(). Nothing corrects
it either - the sibling trim is skipped as exclusive_cpus is non-empty,
and for a direct child of the root cpuset the intersection with
parent->effective_xcpus is a no-op. So the stale 0-7 mask still reaches
tasks_nocpu_error() and PERR_NOCPUS is set again. Please point it out
if I got the analysis wrong.

I have rebased my fix on top of this series and the partition recovers
correctly with it applied. Should the fix be folded into this patch,
or is there a better way to fix it? Happy to hear your thoughts.

Also, is the test case below worth adding here as well? Note that it
fails until the fix is in:

    # An invalid local partition that took all the CPUs from its
    # populated parent should recover after its cpuset.cpus.exclusive
    # list is changed to a valid one.
    "C0-7:X0-7:P1  .      .      .     X1      .      .      .     0 A1:1|XA1:1 A1:P1 ."

Thanks,
Guopeng

> Some test case are added to test_cpuset_prs.sh for invalidating a
> partition root due to its effective xcpus becoming empty as well as
> making sure that an invalid partition can only be revived when exclusive
> CPUs are available agin.
> 
> Fixes: e2ffe502ba45 ("cgroup/cpuset: Add cpuset.cpus.exclusive for v2")
> Signed-off-by: Waiman Long <longman@redhat.com>
> ---
>  kernel/cgroup/cpuset.c                        | 36 +++++++++----------
>  .../selftests/cgroup/test_cpuset_prs.sh       |  6 ++++
>  2 files changed, 22 insertions(+), 20 deletions(-)
> 
> diff --git a/kernel/cgroup/cpuset.c b/kernel/cgroup/cpuset.c
> index 7ead8c3b7974..db6e4663f4be 100644
> --- a/kernel/cgroup/cpuset.c
> +++ b/kernel/cgroup/cpuset.c
> @@ -1798,7 +1798,6 @@ static int update_parent_effective_cpumask(struct cpuset *cs, int cmd,
>  	int deleting;	/* Deleting cpus from parent's effective_cpus	*/
>  	int old_prs, new_prs;
>  	int part_error = PERR_NONE;	/* Partition error? */
> -	struct cpumask *xcpus = user_xcpus(cs);
>  	int parent_prs = parent->partition_root_state;
>  	bool nocpu;
>  
> @@ -1841,30 +1840,29 @@ static int update_parent_effective_cpumask(struct cpuset *cs, int cmd,
>  	if (!newmask && xcpus_empty(cs))
>  		return PERR_CPUSEMPTY;
>  
> -	nocpu = tasks_nocpu_error(parent, cs, xcpus);
> +	/*
> +	 * Except for partcmd_disable, use delmask to compute the current
> +	 * effective xcpus value and use it for tasks_nocpu_error check.
> +	 */
> +	if (cmd != partcmd_disable) {
> +		compute_excpus(cs, tmp->delmask);
> +		nocpu = tasks_nocpu_error(parent, cs, tmp->delmask);
> +	}
>  
>  	if ((cmd == partcmd_enable) || (cmd == partcmd_enablei)) {
> -		/*
> -		 * Need to call compute_excpus() in case
> -		 * exclusive_cpus not set. Sibling conflict should only happen
> -		 * if exclusive_cpus isn't set.
> -		 */
> -		xcpus = tmp->delmask;
> -		if (compute_excpus(cs, xcpus))
> -			WARN_ON_ONCE(!cpumask_empty(cs->exclusive_cpus));
>  		new_prs = (cmd == partcmd_enable) ? PRS_ROOT : PRS_ISOLATED;
>  
>  		/*
>  		 * Enabling partition root is not allowed if its
>  		 * effective_xcpus is empty.
>  		 */
> -		if (cpumask_empty(xcpus))
> +		if (cpumask_empty(tmp->delmask))
>  			return PERR_INVCPUS;
>  
> -		if (prstate_housekeeping_conflict(new_prs, parent_prs, xcpus, NULL))
> +		if (prstate_housekeeping_conflict(new_prs, parent_prs, tmp->delmask, NULL))
>  			return PERR_HKEEPING;
>  
> -		if (tasks_nocpu_error(parent, cs, xcpus))
> +		if (nocpu)
>  			return PERR_NOCPUS;
>  
>  		/*
> @@ -1876,9 +1874,8 @@ static int update_parent_effective_cpumask(struct cpuset *cs, int cmd,
>  		 *
>  		 * Warn if it is not the case.
>  		 */
> -		cpumask_and(tmp->new_cpus, xcpus, cpu_active_mask);
> +		cpumask_and(tmp->new_cpus, tmp->delmask, cpu_active_mask);
>  		WARN_ON_ONCE(!cpumask_subset(tmp->new_cpus, parent->effective_cpus));
> -
>  		deleting = true;
>  	} else if (cmd == partcmd_disable) {
>  		/*
> @@ -1973,7 +1970,7 @@ static int update_parent_effective_cpumask(struct cpuset *cs, int cmd,
>  		 * A partition error happens when parent has tasks and all
>  		 * its effective CPUs will have to be distributed out.
>  		 */
> -		if (nocpu) {
> +		if (nocpu || cpumask_empty(tmp->delmask)) {
>  			part_error = PERR_NOCPUS;
>  			if (is_partition_valid(cs))
>  				adding = cpumask_and(tmp->addmask,
> @@ -1986,8 +1983,7 @@ static int update_parent_effective_cpumask(struct cpuset *cs, int cmd,
>  				cpumask_copy(tmp->addmask, cs->effective_xcpus);
>  				adding = true;
>  			}
> -		} else if (is_partition_invalid(cs) && !cpumask_empty(xcpus) &&
> -			   cpumask_subset(xcpus, parent->effective_xcpus)) {
> +		} else if (is_partition_invalid(cs)) {
>  			struct cgroup_subsys_state *css;
>  			struct cpuset *child;
>  			bool exclusive = true;
> @@ -1998,7 +1994,7 @@ static int update_parent_effective_cpumask(struct cpuset *cs, int cmd,
>  			 * check.
>  			 */
>  			if (prstate_housekeeping_conflict(cs->partition_root_state,
> -							  parent_prs, xcpus, NULL)) {
> +							  parent_prs, tmp->delmask, NULL)) {
>  				part_error = PERR_HKEEPING;
>  				goto write_error;
>  			}
> @@ -2014,7 +2010,7 @@ static int update_parent_effective_cpumask(struct cpuset *cs, int cmd,
>  			rcu_read_unlock();
>  			if (exclusive)
>  				deleting = cpumask_and(tmp->delmask,
> -						xcpus, parent->effective_cpus);
> +						tmp->delmask, parent->effective_cpus);
>  			else
>  				part_error = PERR_NOTEXCL;
>  		}
> diff --git a/tools/testing/selftests/cgroup/test_cpuset_prs.sh b/tools/testing/selftests/cgroup/test_cpuset_prs.sh
> index d04aad697b96..47791591002f 100755
> --- a/tools/testing/selftests/cgroup/test_cpuset_prs.sh
> +++ b/tools/testing/selftests/cgroup/test_cpuset_prs.sh
> @@ -359,6 +359,8 @@ TEST_MATRIX=(
>  								       A1:P0|A2:P1|A3:P2|B1:P1 2-4|3"
>  	"  C1-6:P1 C3-6:P1 C5-6:P1 .      .      P2      .      .    0 A1:1-2|A2:3-6|A3:5-6 \
>  								       A1:P1|A2:P2|A3:P-1 3-6"
> +	"  C1-6:P2 C3-6:P2 C3-4:P2 .      .     C5-6     .      .    0 A1:1-4|A2:5-6|A3:5-6 \
> +								       A1:P2|A2:P2|A3:P-2 1-6"
>  	" C1-4:X2-4 C2-4:X3-4:P1 C3-4:X4:P2 \
>  				   .      .      X5      .      .    0 A1:1-4|A2:2-4|A3:3-4 \
>  								       A1:P0|A2:P-1|A3:P-2 ."
> @@ -427,6 +429,10 @@ TEST_MATRIX=(
>  	# cpuset.cpus is allowed, but those extra CPUs are ignored.
>  	"  C2-3:P1   C3:P1  .      .      .     C2-4    .      .     0 A1:|A2:2-3 A1:P1|A2:P1"
>  
> +	# A child cannot recover until its exclusive CPUs are available again.
> +	" CX1-4:P1 CX3-4:P1 .     .     X1-2     .      .      .     0 A1:1-2|A2:1-2|XA2: A1:P1|A2:P-1"
> +	" CX1-4:P1 CX3-4:P1 .     .  X1-2:X1-4   .      .      .     0 A1:1-2|A2:3-4|XA2:3-4 A1:P1|A2:P1"
> +
>  	# Taking away all CPUs from parent or itself if there are tasks
>  	# will make the partition invalid.
>  	"  C2-3:P1   C3:P1  .      .      T     C2-3    .      .     0 A1:2-3|A2:2-3 A1:P1|A2:P-1"