Re: [PATCH-next v4 6/6] cgroup/cpuset: Fix lockdep_is_cpuset_held()

From: Waiman Long

Date: Mon Nov 16 2026 - 23:06:50 EST


On 11/16/26 2:24 AM, Ridong Chen wrote:


On 11/16/2026 9:23 AM, Waiman Long wrote:
Commit a84097e625f2 ("cgroup/cpuset: Call housekeeping_update()
without holding cpus_read_lock") changes lockdep_is_cpuset_held()
to return true if either cpuset_mutex or cpuset_top_mutex is held.
lockdep_is_cpuset_held() is used by housekeeping_dereference_check()
as one of the conditions that rcu_read_lock() is not needed to
access HK_TYPE_DOMAIN cpumask that can be changed at run time.
In reality, it should return true only if cpuset_top_mutex is held as
housekeeping_update() is called with only cpuset_top_mutex held.

If only cpuset_mutex is held, it is possible that there can be a parallel
housekeeping_update() call leading to a potential use-after-free situation
when HK_TYPE_DOMAIN housekeeping cpumask is being accessed. It also
means that rcu_read_lock will be required if only cpuset_mutex is being
held like when calling from hotplug, scheduler or cgroup.

Fixes: a84097e625f2 ("cgroup/cpuset: Call housekeeping_update() without holding cpus_read_lock")
Signed-off-by: Waiman Long <longman@xxxxxxxxxx>
---
  kernel/cgroup/cpuset-v1.c | 16 ++++++++++------
  kernel/cgroup/cpuset.c    |  9 +++++++--
  2 files changed, 17 insertions(+), 8 deletions(-)

diff --git a/kernel/cgroup/cpuset-v1.c b/kernel/cgroup/cpuset-v1.c
index 7c7f2887dfaf..6bf6da639fec 100644
--- a/kernel/cgroup/cpuset-v1.c
+++ b/kernel/cgroup/cpuset-v1.c
@@ -686,6 +686,7 @@ int cpuset1_generate_sched_domains(cpumask_var_t **domains,
      int csn;        /* how many cpuset ptrs in csa so far */
      int i, j;        /* indices for partition finding loops */
      cpumask_var_t *doms;    /* resulting partition; i.e. sched domains */
+    cpumask_var_t hk_cpus;    /* HK_TYPE_DOMAIN housekeeping CPUs */
      struct sched_domain_attr *dattr;  /* attributes for custom domains */
      int ndoms = 0;        /* number of sched domains in result */
      int nslot;        /* next empty doms[] struct cpumask slot */
@@ -698,6 +699,11 @@ int cpuset1_generate_sched_domains(cpumask_var_t **domains,
      dattr = NULL;
      csa = NULL;
  +    alloc_cpumask_var(&hk_cpus, GFP_KERNEL);
+    scoped_guard(rcu) {
+        cpumask_copy(hk_cpus, housekeeping_cpumask(HK_TYPE_DOMAIN));
+    }
+

The return value of alloc_cpumask_var() isn't checked. With
CONFIG_CPUMASK_OFFSTACK=y, hk_cpus is NULL on failure. Maybe goto done
like the other allocation failures here?

Also, rebuild_sched_domains() can be called from outside cpuset, e.g.
arch_topology and the energy model, with only cpus_read_lock and
cpuset_mutex held. partition_sched_domains_locked() reads
HK_TYPE_DOMAIN in asym_cpu_capacity_scan() and in its fallback path.
Will these trigger a lockdep warning after this change? If so, do they
need an RCU read lock as well?


Yes, I realize that some additional scheduler functions will need to be updated as well for this to work. I am planning to have an alternative solution to work around this problem. I am going to drop this patch and send out another patch soon.

Cheers,
Longman

      /* Special case for the 99% of systems with one, full, sched domain */
      if (is_sched_load_balance(&top_cpuset)) {
          ndoms = 1;
@@ -710,9 +716,7 @@ int cpuset1_generate_sched_domains(cpumask_var_t **domains,
              *dattr = SD_ATTR_INIT;
              update_domain_attr_tree(dattr, &top_cpuset);
          }
-        cpumask_and(doms[0], top_cpuset.effective_cpus,
-                housekeeping_cpumask(HK_TYPE_DOMAIN));
-
+        cpumask_and(doms[0], top_cpuset.effective_cpus, hk_cpus);
          goto done;
      }
  @@ -736,8 +740,7 @@ int cpuset1_generate_sched_domains(cpumask_var_t **domains,
           */
          if (!cpumask_empty(cp->cpus_allowed) &&
              !(is_sched_load_balance(cp) &&
-              cpumask_intersects(cp->cpus_allowed,
-                     housekeeping_cpumask(HK_TYPE_DOMAIN))))
+              cpumask_intersects(cp->cpus_allowed, hk_cpus)))
              continue;
            if (is_sched_load_balance(cp) &&
@@ -794,7 +797,7 @@ int cpuset1_generate_sched_domains(cpumask_var_t **domains,
                          *(dattr + nslot) = SD_ATTR_INIT;
                  }
                  cpumask_or(dp, dp, csa[j]->effective_cpus);
-                cpumask_and(dp, dp, housekeeping_cpumask(HK_TYPE_DOMAIN));
+                cpumask_and(dp, dp, hk_cpus);
                  if (dattr)
                      update_domain_attr_tree(dattr + nslot, csa[j]);
              }
@@ -806,6 +809,7 @@ int cpuset1_generate_sched_domains(cpumask_var_t **domains,
    done:
      kfree(csa);
+    free_cpumask_var(hk_cpus);
        /*
       * Fallback to the default domain if kmalloc() failed.
diff --git a/kernel/cgroup/cpuset.c b/kernel/cgroup/cpuset.c
index 64cbcbdbf6da..d18439e13bce 100644
--- a/kernel/cgroup/cpuset.c
+++ b/kernel/cgroup/cpuset.c
@@ -343,10 +343,13 @@ void cpuset_full_unlock(void)
  }
    #ifdef CONFIG_LOCKDEP
+/*
+ * Holding cpuset_top_mutex will guarantee that a parallel
+ * housekeeping_update() call modifying housekeeping cpumasks is not possible.
+ */
  bool lockdep_is_cpuset_held(void)
  {
-    return lockdep_is_held(&cpuset_mutex) ||
-           lockdep_is_held(&cpuset_top_mutex);
+    return lockdep_is_held(&cpuset_top_mutex);
  }
  #endif
  @@ -1416,8 +1419,10 @@ static bool prstate_housekeeping_conflict(int new_prs, int parent_prs,
          }
      }
  +    rcu_read_lock();
      cpumask_and(full_hk_cpus, housekeeping_cpumask(HK_TYPE_KERNEL_NOISE),
              housekeeping_cpumask(HK_TYPE_DOMAIN));
+    rcu_read_unlock();
      cpumask_andnot(full_hk_cpus, full_hk_cpus, isolated_cpus);
      cpumask_and(full_hk_cpus, full_hk_cpus, cpu_active_mask);
      res = !cpumask_weight_andnot(full_hk_cpus, new_cpus);