cgroup: fix cgroup_rmdir() vs close(eventfd) race
authorLi Zefan <lizefan@huawei.com>
Mon, 18 Feb 2013 10:56:14 +0000 (18:56 +0800)
committerTejun Heo <tj@kernel.org>
Mon, 18 Feb 2013 17:17:24 +0000 (09:17 -0800)
commit 205a872bd6f9a9a09ef035ef1e90185a8245cc58 ("cgroup: fix lockdep
warning for event_control") solved a deadlock by introducing a new
bug.

Move cgrp->event_list to a temporary list doesn't mean you can traverse
this list locklessly, because at the same time cgroup_event_wake() can
be called and remove the event from the list. The result of this race
is disastrous.

We adopt the way how kvm irqfd code implements race-free event removal,
which is now described in the comments in cgroup_event_wake().

v3:
- call eventfd_signal() no matter it's eventfd close or cgroup removal
that removes the cgroup event.

Acked-by: Kirill A. Shutemov <kirill@shutemov.name>
Signed-off-by: Li Zefan <lizefan@huawei.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
kernel/cgroup.c

index 5d4c92e..feda814 100644 (file)
@@ -3786,8 +3786,13 @@ static void cgroup_event_remove(struct work_struct *work)
                        remove);
        struct cgroup *cgrp = event->cgrp;
 
+       remove_wait_queue(event->wqh, &event->wait);
+
        event->cft->unregister_event(cgrp, event->cft, event->eventfd);
 
+       /* Notify userspace the event is going away. */
+       eventfd_signal(event->eventfd, 1);
+
        eventfd_ctx_put(event->eventfd);
        kfree(event);
        dput(cgrp->dentry);
@@ -3807,15 +3812,25 @@ static int cgroup_event_wake(wait_queue_t *wait, unsigned mode,
        unsigned long flags = (unsigned long)key;
 
        if (flags & POLLHUP) {
-               __remove_wait_queue(event->wqh, &event->wait);
-               spin_lock(&cgrp->event_list_lock);
-               list_del_init(&event->list);
-               spin_unlock(&cgrp->event_list_lock);
                /*
-                * We are in atomic context, but cgroup_event_remove() may
-                * sleep, so we have to call it in workqueue.
+                * If the event has been detached at cgroup removal, we
+                * can simply return knowing the other side will cleanup
+                * for us.
+                *
+                * We can't race against event freeing since the other
+                * side will require wqh->lock via remove_wait_queue(),
+                * which we hold.
                 */
-               schedule_work(&event->remove);
+               spin_lock(&cgrp->event_list_lock);
+               if (!list_empty(&event->list)) {
+                       list_del_init(&event->list);
+                       /*
+                        * We are in atomic context, but cgroup_event_remove()
+                        * may sleep, so we have to call it in workqueue.
+                        */
+                       schedule_work(&event->remove);
+               }
+               spin_unlock(&cgrp->event_list_lock);
        }
 
        return 0;
@@ -4375,20 +4390,14 @@ static int cgroup_destroy_locked(struct cgroup *cgrp)
        /*
         * Unregister events and notify userspace.
         * Notify userspace about cgroup removing only after rmdir of cgroup
-        * directory to avoid race between userspace and kernelspace. Use
-        * a temporary list to avoid a deadlock with cgroup_event_wake(). Since
-        * cgroup_event_wake() is called with the wait queue head locked,
-        * remove_wait_queue() cannot be called while holding event_list_lock.
+        * directory to avoid race between userspace and kernelspace.
         */
        spin_lock(&cgrp->event_list_lock);
-       list_splice_init(&cgrp->event_list, &tmp_list);
-       spin_unlock(&cgrp->event_list_lock);
-       list_for_each_entry_safe(event, tmp, &tmp_list, list) {
+       list_for_each_entry_safe(event, tmp, &cgrp->event_list, list) {
                list_del_init(&event->list);
-               remove_wait_queue(event->wqh, &event->wait);
-               eventfd_signal(event->eventfd, 1);
                schedule_work(&event->remove);
        }
+       spin_unlock(&cgrp->event_list_lock);
 
        return 0;
 }