exit: fix race between wait_consider_task() and wait_task_zombie()

wait_consider_task() checks EXIT_ZOMBIE after EXIT_DEAD/EXIT_TRACE and
both checks can fail if we race with EXIT_ZOMBIE -> EXIT_DEAD/EXIT_TRACE
change in between, gcc needs to reload p->exit_state after
security_task_wait().  In this case ->notask_error will be wrongly
cleared and do_wait() can hang forever if it was the last eligible
child.

Many thanks to Arne who carefully investigated the problem.

Note: this bug is very old but it was pure theoretical until commit
b3ab03160dfa ("wait: completely ignore the EXIT_DEAD tasks").  Before
this commit "-O2" was probably enough to guarantee that compiler won't
read ->exit_state twice.

Signed-off-by: Oleg Nesterov <oleg@redhat.com>
Reported-by: Arne Goedeke <el@laramies.com>
Tested-by: Arne Goedeke <el@laramies.com>
Cc: <stable@vger.kernel.org>	[3.15+]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

diff --git a/kernel/exit.c b/kernel/exit.c
index 1ea4369..6806c55 100644 (file)
--- a/kernel/exit.c
+++ b/kernel/exit.c
@@ -1287,9 +1287,15 @@ static int wait_task_continued(struct wait_opts *wo, struct task_struct *p)
 static int wait_consider_task(struct wait_opts *wo, int ptrace,
                                struct task_struct *p)
 {
+       /*
+        * We can race with wait_task_zombie() from another thread.
+        * Ensure that EXIT_ZOMBIE -> EXIT_DEAD/EXIT_TRACE transition
+        * can't confuse the checks below.
+        */
+       int exit_state = ACCESS_ONCE(p->exit_state);
        int ret;
 
-       if (unlikely(p->exit_state == EXIT_DEAD))
+       if (unlikely(exit_state == EXIT_DEAD))
                return 0;
 
        ret = eligible_child(wo, p);
@@ -1310,7 +1316,7 @@ static int wait_consider_task(struct wait_opts *wo, int ptrace,
                return 0;
        }
 
-       if (unlikely(p->exit_state == EXIT_TRACE)) {
+       if (unlikely(exit_state == EXIT_TRACE)) {
                /*
                 * ptrace == 0 means we are the natural parent. In this case
                 * we should clear notask_error, debugger will notify us.
@@ -1337,7 +1343,7 @@ static int wait_consider_task(struct wait_opts *wo, int ptrace,
        }
 
        /* slay zombie? */
-       if (p->exit_state == EXIT_ZOMBIE) {
+       if (exit_state == EXIT_ZOMBIE) {
                /* we don't reap group leaders with subthreads */
                if (!delay_group_leader(p)) {
                        /*