Make sure "user->sigpending" count is in sync

The previous commit (45c18b0, aka "Fix
unlikely (but possible) race condition on task->user access") fixed a
potential oops due to __sigqueue_alloc() getting its "user" pointer out
of sync with switch_user(), and accessing a user pointer that had been
de-allocated on another CPU.

It still left another (much less serious) problem, where a concurrent
__sigqueue_alloc and swich_user could cause sigqueue_alloc to do signal
pending reference counting for a _different_ user than the one it then
actually ended up using.  No oops, but we'd end up with the wrong signal
accounting.

Another case of Oleg's eagle-eyes picking up the problem.

This is trivially fixed by just making sure we load whichever "user"
structure we decide to use (it doesn't matter _which_ one we pick, we
just need to pick one) just once.

Acked-by: Oleg Nesterov <[email protected]>
Cc: Andrew Morton <[email protected]>
Cc: Ingo Molnar <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>

kernel/signal.c

@@ -267,18 +267,25 @@ static struct sigqueue *__sigqueue_alloc(struct task_struct *t, gfp_t flags,
 					 int override_rlimit)
 {
 	struct sigqueue *q = NULL;
+	struct user_struct *user;
 
-	atomic_inc(&t->user->sigpending);
+	/*
+	 * In order to avoid problems with "switch_user()", we want to make
+	 * sure that the compiler doesn't re-load "t->user"
+	 */
+	user = t->user;
+	barrier();
+	atomic_inc(&user->sigpending);
 	if (override_rlimit ||
-	    atomic_read(&t->user->sigpending) <=
+	    atomic_read(&user->sigpending) <=
 			t->signal->rlim[RLIMIT_SIGPENDING].rlim_cur)
 		q = kmem_cache_alloc(sigqueue_cachep, flags);
 	if (unlikely(q == NULL)) {
-		atomic_dec(&t->user->sigpending);
+		atomic_dec(&user->sigpending);
 	} else {
 		INIT_LIST_HEAD(&q->list);
 		q->flags = 0;
-		q->user = get_uid(t->user);
+		q->user = get_uid(user);
 	}
 	return(q);
 }