tracing: Fix preempt count leak

While running my ftrace stress test, this showed up:

BUG: sleeping function called from invalid context at mm/mmap.c:233
...
note: cat[3293] exited with preempt_count 1

The bug was introduced by commit 91e86e5
("tracing: Fix recursive user stack trace")

Cc: <stable@kernel.org>
Signed-off-by: Li Zefan <lizf@cn.fujitsu.com>
LKML-Reference: <4D0089AC.1020802@cn.fujitsu.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>

kernel/trace/trace.c

@@ -1313,12 +1313,10 @@ ftrace_trace_userstack(struct ring_buffer *buffer, unsigned long flags, int pc)
 
 	__this_cpu_inc(user_stack_count);
 
-
-
 	event = trace_buffer_lock_reserve(buffer, TRACE_USER_STACK,
 					  sizeof(*entry), flags, pc);
 	if (!event)
-		return;
+		goto out_drop_count;
 	entry	= ring_buffer_event_data(event);
 
 	entry->tgid		= current->tgid;
@@ -1333,8 +1331,8 @@ ftrace_trace_userstack(struct ring_buffer *buffer, unsigned long flags, int pc)
 	if (!filter_check_discard(call, entry, buffer, event))
 		ring_buffer_unlock_commit(buffer, event);
 
+ out_drop_count:
 	__this_cpu_dec(user_stack_count);
-
  out:
 	preempt_enable();
 }