[PATCH 2/5] OProfile update

John Levon (levon@movementarian.org)
Mon, 26 May 2003 05:27:29 +0100


The code that attempts to reset last_task and in_kernel has a race
against samples appearing during the handling of the buffers, that
causes a small number of mis-attribution of samples. Closing the
window is non-obvious, and not worth it, so we just make it smaller.
Even without the patch, there seem to be few such "bad" samples
because its effects are mitigated on a switch into userspace or
a task switch.

diff -Naur -X dontdiff linux-cvs/drivers/oprofile/buffer_sync.c linux-me/drivers/oprofile/buffer_sync.c
--- linux-cvs/drivers/oprofile/buffer_sync.c 2003-05-26 03:20:20.000000000 +0100
+++ linux-me/drivers/oprofile/buffer_sync.c 2003-05-26 04:25:15.000000000 +0100
@@ -366,12 +377,26 @@
}


-/* compute number of filled slots in cpu_buffer queue */
-static unsigned long nr_filled_slots(struct oprofile_cpu_buffer * b)
+/* "acquire" as many cpu buffer slots as we can */
+static unsigned long get_slots(struct oprofile_cpu_buffer * b)
{
unsigned long head = b->head_pos;
unsigned long tail = b->tail_pos;

+ /*
+ * Subtle. This resets the persistent last_task
+ * and in_kernel values used for switching notes.
+ * BUT, there is a small window between reading
+ * head_pos, and this call, that means samples
+ * can appear at the new head position, but not
+ * be prefixed with the notes for switching
+ * kernel mode or a task switch. This small hole
+ * can lead to mis-attribution or samples where
+ * we don't know if it's in the kernel or not,
+ * at the start of an event buffer.
+ */
+ cpu_buffer_reset(b);
+
if (head >= tail)
return head - tail;

@@ -408,9 +433,9 @@

/* Remember, only we can modify tail_pos */

- unsigned long const available_elements = nr_filled_slots(cpu_buf);
+ unsigned long const available = get_slots(cpu_buf);

- for (i=0; i < available_elements; ++i) {
+ for (i=0; i < available; ++i) {
struct op_sample * s = &cpu_buf->buffer[cpu_buf->tail_pos];

if (is_ctx_switch(s->eip)) {
@@ -435,8 +460,6 @@
increment_tail(cpu_buf);
}
release_mm(mm);
-
- cpu_buffer_reset(cpu_buf);
}


-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/