Skip to contentSkip to conversation

cgroup: Fix potential race between cgroup_exit and migrate path

PATCH In review Neeraj Upadhyay 2 participants 5 messages Aug 30, 2017 – Sep 4, 2017 cgroup

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 cgroup: Fix potential race between cgroup_exit and migrate path

There is a potential race between cgroup_exit() and the migration path. This race happens because cgroup_exit path reads the css_set and does cg_list empty check outside of css_set lock. This can potentially race with the migrate path trying to move the tasks to a different css_set. For instance, below is the interleaved sequence of events, where race is observed:

cpuset_hotplug_workfn()
cgroup_transfer_tasks()

    cgroup_migrate()
      cgroup_migrate_execute()
          css_set_move_task()
            list_del_init(&task->cg_list);
              <TASK EXIT>
                cgroup_exit()
                  cset = task_css_set(tsk);
                  if (!list_empty(&tsk->cg_list))
                    <TASK NOT DISSOCIATED FROM ITS CSS_SET>
            list_add_tail(&task->cg_list, use_mg_tasks

In above sequence, as cgroup_exit() read the cg_list for the task as empty, it didn't disassociate it from its current css_set, and was moved to new css_set instance css_set_move_task() called from cpuset_hotplug_workfn() path. This eventually can result in use after free scenarios, while accessing the same task_struct again, like in following sequence:

kernfs_seq_start()
cgroup_seqfile_start()
cgroup_pidlist_start()
css_task_iter_next()
__put_task_struct()
<NULL pointer dereference>

Fix this problem, by moving the css_set and cg_list fetch in cgroup_exit() inside css_set lock.

---
Hi,

We observed this issue for cgroup code corresponding to stable v4.4.85 snapshot 3144d81 ("cgroup, kthread: close race window where new kthreads can be migrated to non-root cgroups"). Can you please tell us, if there are any patches in latest code, which fixes these issue?

 kernel/cgroup/cgroup.c | 20 ++++++++++++--------
 1 file changed, 12 insertions(+), 8 deletions(-)
kernel/cgroup/cgroup.c +12 -8
index df2e0f1..f746b70 100644
@@ -692,10 +692,10 @@ static void css_set_move_task(struct task_struct *task,
 
 	if (to_cset) {
 		/*
-		 * We are synchronized through cgroup_threadgroup_rwsem
-		 * against PF_EXITING setting such that we can't race
-		 * against cgroup_exit() changing the css_set to
-		 * init_css_set and dropping the old one.
+		 * We are synchronized through css_set_lock against
+		 * PF_EXITING setting such that we can't race against
+		 * cgroup_exit() disassociating the task from the
+		 * css_set.
 		 */
 		WARN_ON_ONCE(task->flags & PF_EXITING);
 
@@ -4934,20 +4934,24 @@ void cgroup_exit(struct task_struct *tsk)
 	int i;
 
 	/*
-	 * Unlink from @tsk from its css_set.  As migration path can't race
-	 * with us, we can check css_set and cg_list without synchronization.
+	 * Avoid potential race with the migrate path.
+	 */
+	spin_lock_irq(&css_set_lock);
+
+	/*
+	 * Unlink from @tsk from its css_set.
 	 */
 	cset = task_css_set(tsk);
 
 	if (!list_empty(&tsk->cg_list)) {
-		spin_lock_irq(&css_set_lock);
 		css_set_move_task(tsk, cset, NULL, false);
 		cset->nr_tasks--;
-		spin_unlock_irq(&css_set_lock);
 	} else {
 		get_css_set(cset);
 	}
 
+	spin_unlock_irq(&css_set_lock);
+
 	/* see cgroup_post_fork() for details */
 	do_each_subsys_mask(ss, i, have_exit_callback) {
 		ss->exit(tsk);
Signature · 2 lines
QUALCOMM INDIA, on behalf of Qualcomm Innovation Center, Inc. is a
member of the Code Aurora Forum, hosted by The Linux Foundation
maintainer
#1 lore raw

Hello, Neeraj.

Quoted 36 lines · Neeraj Upadhyay

There is a potential race between cgroup_exit() and the migration path. This race happens because cgroup_exit path reads the css_set and does cg_list empty check outside of css_set lock. This can potentially race with the migrate path trying to move the tasks to a different css_set. For instance, below is the interleaved sequence of events, where race is observed:

cpuset_hotplug_workfn()
cgroup_transfer_tasks()

    cgroup_migrate()
      cgroup_migrate_execute()
          css_set_move_task()
            list_del_init(&task->cg_list);
              <TASK EXIT>
                cgroup_exit()
                  cset = task_css_set(tsk);
                  if (!list_empty(&tsk->cg_list))
                    <TASK NOT DISSOCIATED FROM ITS CSS_SET>
            list_add_tail(&task->cg_list, use_mg_tasks

In above sequence, as cgroup_exit() read the cg_list for the task as empty, it didn't disassociate it from its current css_set, and was moved to new css_set instance css_set_move_task() called from cpuset_hotplug_workfn() path. This eventually can result in use after free scenarios, while accessing the same task_struct again, like in following sequence:

kernfs_seq_start()
cgroup_seqfile_start()
cgroup_pidlist_start()
css_task_iter_next()
__put_task_struct()
<NULL pointer dereference>

Quoted 2 lines

Fix this problem, by moving the css_set and cg_list fetch in cgroup_exit() inside css_set lock.

Hmm... I haven't really thought through but could the problem be that css_set_move_task() is temporarily making ->cg_list empty? The use_task_css_set_links optimization can't handle that.

Would something like the following fix the issue? Thanks.

kernel/cgroup/cgroup.c +1 -1
index df2e0f1..cd85ca0 100644
@@ -683,7 +683,7 @@ static void css_set_move_task(struct task_struct *task,
 			if (it->task_pos == &task->cg_list)
 				css_task_iter_advance(it);
 
-		list_del_init(&task->cg_list);
+		list_del(&task->cg_list);
 		if (!css_set_populated(from_cset))
 			css_set_update_populated(from_cset, false);
 	} else {
maintainer
#2 lore raw
Quoted 7 lines · Tejun Heo

Would something like the following fix the issue? Thanks.

kernel/cgroup/cgroup.c +0 -0
index df2e0f1..cd85ca0 100644

kernel/cgroup/cgroup.c static void css_set_move_task(struct task_struct *task, · line 689

Quoted 8 lines
kernel/cgroup/cgroup.c +1 -1
@@ -683,7 +683,7 @@ static void css_set_move_task(struct task_struct *task,
 			if (it->task_pos == &task->cg_list)
 				css_task_iter_advance(it);
 
-		list_del_init(&task->cg_list);
+		list_del(&task->cg_list);
 		if (!css_set_populated(from_cset))
 			css_set_update_populated(from_cset, false);
 	} else {

Oops, more like the following.

kernel/cgroup/cgroup.c +3 -1
index df2e0f1..6f34025 100644
@@ -683,7 +683,7 @@ static void css_set_move_task(struct task_struct *task,
 			if (it->task_pos == &task->cg_list)
 				css_task_iter_advance(it);
 
-		list_del_init(&task->cg_list);
+		list_del(&task->cg_list);
 		if (!css_set_populated(from_cset))
 			css_set_update_populated(from_cset, false);
 	} else {
@@ -702,6 +702,8 @@ static void css_set_move_task(struct task_struct *task,
 		rcu_assign_pointer(task->cgroups, to_cset);
 		list_add_tail(&task->cg_list, use_mg_tasks ? &to_cset->mg_tasks :
 							     &to_cset->tasks);
+	} else {
+		INIT_LIST_HEAD(&task->cg_list);
 	}
 }
 
maintainer
#3 lore raw
Quoted 16 lines · Tejun Heo

Oops, more like the following.

kernel/cgroup/cgroup.c +1 -1
index df2e0f1..6f34025 100644
@@ -683,7 +683,7 @@ static void css_set_move_task(struct task_struct *task,
 			if (it->task_pos == &task->cg_list)
 				css_task_iter_advance(it);
 
-		list_del_init(&task->cg_list);
+		list_del(&task->cg_list);
 		if (!css_set_populated(from_cset))
 			css_set_update_populated(from_cset, false);
 	} else {

kernel/cgroup/cgroup.c static void css_set_move_task(struct task_struct *task, · line 708

Quoted 7 lines
kernel/cgroup/cgroup.c +2 -0
@@ -702,6 +702,8 @@ static void css_set_move_task(struct task_struct *task,
 		rcu_assign_pointer(task->cgroups, to_cset);
 		list_add_tail(&task->cg_list, use_mg_tasks ? &to_cset->mg_tasks :
 							     &to_cset->tasks);
+	} else {
+		INIT_LIST_HEAD(&task->cg_list);
 	}
 }

On the third thought, I don't think this can happen either because now migration is strongly synchronized against exits. Please take a look at the changes around cgroup_threadgroup_rwsem.

Thanks.

Signature · 1 line
tejun
#4 lore raw
Quoted 28 lines · Tejun Heo
Quoted 16 lines · Tejun Heo

Oops, more like the following.

kernel/cgroup/cgroup.c +1 -1
index df2e0f1..6f34025 100644
@@ -683,7 +683,7 @@ static void css_set_move_task(struct task_struct *task,
  			if (it->task_pos == &task->cg_list)
  				css_task_iter_advance(it);
 
-		list_del_init(&task->cg_list);
+		list_del(&task->cg_list);
  		if (!css_set_populated(from_cset))
  			css_set_update_populated(from_cset, false);
  	} else {

kernel/cgroup/cgroup.c static void css_set_move_task(struct task_struct *task, · line 708

Quoted 7 lines
kernel/cgroup/cgroup.c +2 -0
@@ -702,6 +702,8 @@ static void css_set_move_task(struct task_struct *task,
  		rcu_assign_pointer(task->cgroups, to_cset);
  		list_add_tail(&task->cg_list, use_mg_tasks ? &to_cset->mg_tasks :
  							     &to_cset->tasks);
+	} else {
+		INIT_LIST_HEAD(&task->cg_list);
  	}
  }

On the third thought, I don't think this can happen either because now migration is strongly synchronized against exits. Please take a look at the changes around cgroup_threadgroup_rwsem.

Quoted 2 lines

Thanks.

Thank you for the suggestion; found below fix, which is not present in v4.4.86 stable code base. Please let me know in case I am missing something:

eedd0f4 cgroupns: Close race between cgroup_post_fork and copy_cgroup_ns

Thanks.

Signature · 2 lines
QUALCOMM INDIA, on behalf of Qualcomm Innovation Center, Inc. is a
member of the Code Aurora Forum, hosted by The Linux Foundation