tracing: Break out of tracing_wait_pipe() before wait_pipe() is called
authorSteven Rostedt (Red Hat) <rostedt@goodmis.org>
Tue, 29 Apr 2014 20:07:28 +0000 (16:07 -0400)
committerSteven Rostedt <rostedt@goodmis.org>
Tue, 29 Apr 2014 20:07:28 +0000 (16:07 -0400)
When reading from trace_pipe, if tracing is off but nothing was read
it should block. If something is read and tracing is off, then EOF
is returned. If tracing is on and there's nothing to read, it will block.

But because the check of whether tracing is off and something was read
is done after the block on the pipe, it is hit or miss if the EOF is
returned or not leading to inconsistent behavior.

Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
kernel/trace/trace.c

index cb41e98cc64b0096897eafda5b5bd09f509c8fd3..e058c6091e4543f737c1645ff63a2f473d6e098e 100644 (file)
@@ -4237,15 +4237,6 @@ static int tracing_wait_pipe(struct file *filp)
                        return -EAGAIN;
                }
 
-               mutex_unlock(&iter->mutex);
-
-               iter->trace->wait_pipe(iter);
-
-               mutex_lock(&iter->mutex);
-
-               if (signal_pending(current))
-                       return -EINTR;
-
                /*
                 * We block until we read something and tracing is disabled.
                 * We still block if tracing is disabled, but we have never
@@ -4257,6 +4248,15 @@ static int tracing_wait_pipe(struct file *filp)
                 */
                if (!tracing_is_on() && iter->pos)
                        break;
+
+               mutex_unlock(&iter->mutex);
+
+               iter->trace->wait_pipe(iter);
+
+               mutex_lock(&iter->mutex);
+
+               if (signal_pending(current))
+                       return -EINTR;
        }
 
        return 1;