ring-buffer: Make write slow path out of line
authorSteven Rostedt <srostedt@redhat.com>
Fri, 8 Oct 2010 17:51:48 +0000 (13:51 -0400)
committerSteven Rostedt <rostedt@goodmis.org>
Tue, 19 Oct 2010 17:22:36 +0000 (13:22 -0400)
Gcc inlines the slow path of the ring buffer write which can
hurt performance. This patch simply forces the slow path function
rb_move_tail() to always be a function.

The ring_buffer_benchmark module with reader_disabled=1 shows that
this patch changes the time to record an event from 135 ns to
132 ns. (3 ns or 2.22% improvement)

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

index bca96377fd4e8667df513dfd91ea56931e35578f..0b88df849a599773f02ee6847983d694f4640070 100644 (file)
@@ -1823,7 +1823,10 @@ rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
        local_sub(length, &tail_page->write);
 }
 
-static struct ring_buffer_event *
+/*
+ * This is the slow path, force gcc not to inline it.
+ */
+static noinline struct ring_buffer_event *
 rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
             unsigned long length, unsigned long tail,
             struct buffer_page *tail_page, u64 *ts)
@@ -1943,7 +1946,7 @@ __rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
        tail = write - length;
 
        /* See if we shot pass the end of this buffer page */
-       if (write > BUF_PAGE_SIZE)
+       if (unlikely(write > BUF_PAGE_SIZE))
                return rb_move_tail(cpu_buffer, length, tail,
                                    tail_page, ts);