cgroup: Fix deadlock in cpu hotplug path
authorPrateek Sood <prsood@codeaurora.org>
Tue, 19 Dec 2017 07:26:57 +0000 (12:56 +0530)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Sat, 3 Mar 2018 09:24:26 +0000 (10:24 +0100)
[ Upstream commit 116d2f7496c51b2e02e8e4ecdd2bdf5fb9d5a641 ]

Deadlock during cgroup migration from cpu hotplug path when a task T is
being moved from source to destination cgroup.

kworker/0:0
cpuset_hotplug_workfn()
   cpuset_hotplug_update_tasks()
      hotplug_update_tasks_legacy()
        remove_tasks_in_empty_cpuset()
          cgroup_transfer_tasks() // stuck in iterator loop
            cgroup_migrate()
              cgroup_migrate_add_task()

In cgroup_migrate_add_task() it checks for PF_EXITING flag of task T.
Task T will not migrate to destination cgroup. css_task_iter_start()
will keep pointing to task T in loop waiting for task T cg_list node
to be removed.

Task T
do_exit()
  exit_signals() // sets PF_EXITING
  exit_task_namespaces()
    switch_task_namespaces()
      free_nsproxy()
        put_mnt_ns()
          drop_collected_mounts()
            namespace_unlock()
              synchronize_rcu()
                _synchronize_rcu_expedited()
                  schedule_work() // on cpu0 low priority worker pool
                  wait_event() // waiting for work item to execute

Task T inserted a work item in the worklist of cpu0 low priority
worker pool. It is waiting for expedited grace period work item
to execute. This work item will only be executed once kworker/0:0
complete execution of cpuset_hotplug_workfn().

kworker/0:0 ==> Task T ==>kworker/0:0

In case of PF_EXITING task being migrated from source to destination
cgroup, migrate next available task in source cgroup.

Signed-off-by: Prateek Sood <prsood@codeaurora.org>
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Sasha Levin <alexander.levin@verizon.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
kernel/cgroup/cgroup-v1.c

index 024085d..a2c05d2 100644 (file)
@@ -123,7 +123,11 @@ int cgroup_transfer_tasks(struct cgroup *to, struct cgroup *from)
         */
        do {
                css_task_iter_start(&from->self, 0, &it);
-               task = css_task_iter_next(&it);
+
+               do {
+                       task = css_task_iter_next(&it);
+               } while (task && (task->flags & PF_EXITING));
+
                if (task)
                        get_task_struct(task);
                css_task_iter_end(&it);