Re: [PATCH next] cpuset: fix warning when attaching tasks with offline CPUs

[Waiman Long <llong@redhat.com>]

On 7/13/25 11:23 PM, Chen Ridong wrote:
> From: Chen Ridong <chenridong@huawei.com>
>
> A kernel warning was observed in the cpuset migration path:
>
>      WARNING: CPU: 3 PID: 123 at kernel/cgroup/cpuset.c:3130
>      cgroup_migrate_execute+0x8df/0xf30
>      Call Trace:
>       cgroup_transfer_tasks+0x2f3/0x3b0
>       cpuset_migrate_tasks_workfn+0x146/0x3b0
>       process_one_work+0x5ba/0xda0
>       worker_thread+0x788/0x1220
>
> The issue can be reliably reproduced with:
>
>      # Setup test cpuset
>      mkdir /sys/fs/cgroup/cpuset/test
>      echo 2-3 > /sys/fs/cgroup/cpuset/test/cpuset.cpus
>      echo 0 > /sys/fs/cgroup/cpuset/test/cpuset.mems
>
>      # Start test process
>      sleep 100 &
>      pid=$!
>      echo $pid > /sys/fs/cgroup/cpuset/test/cgroup.procs
>      taskset -p 0xC $pid  # Bind to CPUs 2-3
>
>      # Take CPUs offline
>      echo 0 > /sys/devices/system/cpu/cpu3/online
>      echo 0 > /sys/devices/system/cpu/cpu2/online
>
> Root cause analysis:
> When tasks are migrated to top_cpuset due to CPUs going offline,
> cpuset_attach_task() sets the CPU affinity using cpus_attach which
> is initialized from cpu_possible_mask. This mask may include offline
> CPUs. When __set_cpus_allowed_ptr() computes the intersection between:
> 1. cpus_attach (possible CPUs, may include offline)
> 2. p->user_cpus_ptr (original user-set mask)
> The resulting new_mask may contain only offline CPUs, causing the
> operation to fail.
>
> The fix changes cpus_attach initialization to use cpu_active_mask
> instead of cpu_possible_mask, ensuring we only consider online CPUs
> when setting the new affinity. This prevents the scenario where
> the intersection would result in an invalid CPU set.
>
> Fixes: da019032819a ("sched: Enforce user requested affinity")
> Reported-by: Yang Lijin <yanglijin@huawei.com>
> Signed-off-by: Chen Ridong <chenridong@huawei.com>
> ---
>   kernel/cgroup/cpuset.c | 2 +-
>   1 file changed, 1 insertion(+), 1 deletion(-)
>
> diff --git a/kernel/cgroup/cpuset.c b/kernel/cgroup/cpuset.c
> index f74d04429a29..5401adbdffa6 100644
> --- a/kernel/cgroup/cpuset.c
> +++ b/kernel/cgroup/cpuset.c
> @@ -3121,7 +3121,7 @@ static void cpuset_attach_task(struct cpuset *cs, struct task_struct *task)
>   	if (cs != &top_cpuset)
>   		guarantee_active_cpus(task, cpus_attach);
>   	else
> -		cpumask_andnot(cpus_attach, task_cpu_possible_mask(task),
> +		cpumask_andnot(cpus_attach, cpu_active_mask,
>   			       subpartitions_cpus);
>   	/*
>   	 * can_attach beforehand should guarantee that this doesn't

Offline CPUs are explicitly included for tasks in top_cpuset. Can you 
try the following patch to see if it works?

Thanks,
Longman

diff --git a/kernel/cgroup/cpuset.c b/kernel/cgroup/cpuset.c
index 3bc4301466f3..acd70120228c 100644
--- a/kernel/cgroup/cpuset.c
+++ b/kernel/cgroup/cpuset.c
@@ -3114,6 +3114,10 @@ static void cpuset_cancel_attach(struct 
cgroup_taskset *tset)
  static cpumask_var_t cpus_attach;
  static nodemask_t cpuset_attach_nodemask_to;

+/*
+ * Note that tasks in the top cpuset won't get update to their cpumasks 
when
+ * a hotplug event happens. So we include offline CPUs as well.
+ */
  static void cpuset_attach_task(struct cpuset *cs, struct task_struct 
*task)
  {
         lockdep_assert_held(&cpuset_mutex);
@@ -3127,7 +3131,16 @@ static void cpuset_attach_task(struct cpuset *cs, 
struct task_struct *task)
          * can_attach beforehand should guarantee that this doesn't
          * fail.  TODO: have a better way to handle failure here
          */
-       WARN_ON_ONCE(set_cpus_allowed_ptr(task, cpus_attach));
+       if (unlikely(set_cpus_allowed_ptr(task, cpus_attach))) {
+               /*
+                * Since offline CPUs are included for top_cpuset,
+                * set_cpus_allowed_ptr() can fail if user_cpus_ptr contains
+                * only offline CPUs. Take out the offline CPUs and retry.
+                */
+               if (cs == &top_cpuset)
+                       cpumask_and(cpus_attach, cpus_attach, 
cpu_active_mask);
+               WARN_ON_ONCE(set_cpus_allowed_ptr(task, cpus_attach));
+       }

         cpuset_change_task_nodemask(task, &cpuset_attach_nodemask_to);
         cpuset1_update_task_spread_flags(cs, task);