Fix race between attach_task and cpuset_exit
Srivatsa Vaddagiri [Tue, 8 May 2007 07:27:25 +0000 (00:27 -0700)]
Currently cpuset_exit() changes the exiting task's ->cpuset pointer w/o
taking task_lock().  This can lead to ugly races between attach_task and
cpuset_exit.  Details of the races are described at
http://lkml.org/lkml/2007/3/24/132.

Patch below closes those races.

Signed-off-by: Srivatsa Vaddagiri <vatsa@in.ibm.com>
Cc: Paul Jackson <pj@sgi.com>
Cc: Balbir Singh <balbir@in.ibm.com>
Cc: Paul Menage <menage@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

kernel/cpuset.c

index d240349..bde714d 100644 (file)
@@ -2200,10 +2200,6 @@ void cpuset_fork(struct task_struct *child)
  * it is holding that mutex while calling check_for_release(),
  * which calls kmalloc(), so can't be called holding callback_mutex().
  *
- * We don't need to task_lock() this reference to tsk->cpuset,
- * because tsk is already marked PF_EXITING, so attach_task() won't
- * mess with it, or task is a failed fork, never visible to attach_task.
- *
  * the_top_cpuset_hack:
  *
  *    Set the exiting tasks cpuset to the root cpuset (top_cpuset).
@@ -2242,8 +2238,10 @@ void cpuset_exit(struct task_struct *tsk)
 {
        struct cpuset *cs;
 
+       task_lock(current);
        cs = tsk->cpuset;
        tsk->cpuset = &top_cpuset;      /* the_top_cpuset_hack - see above */
+       task_unlock(current);
 
        if (notify_on_release(cs)) {
                char *pathbuf = NULL;