On Fri, May 09, 2014 at 12:12:19AM +0100, Sebastian Capella wrote:

Quoting Lorenzo Pieralisi (2014-05-06 11:04:40)

quoted

diff --git a/drivers/cpuidle/of_idle_states.c b/drivers/cpuidle/of_idle_states.c
new file mode 100644
index 0000000..360b7ad
--- /dev/null
+++ b/drivers/cpuidle/of_idle_states.c

@@ -0,0 +1,293 @@

...

quoted

+static int __init add_state_node(cpumask_t *cpumask,
+                                struct device_node *state_node)
+{
+       struct state_elem *el;
+       u32 tmp, val = 0;
+
+       pr_debug(" * %s...\n", state_node->full_name);
+
+       if (!state_cpus_valid(cpumask, state_node))
+               return -EINVAL;
+       /*
+        * Parse just the properties required to sort the states.
+        * Since we are missing a value defining the energy
+        * efficiency of a state, for now the sorting code uses
+        *
+        * min-residency-us+exit-latency-us
+        *
+        * as sorting rank.
+        */
+       if (of_property_read_u32(state_node, "min-residency-us",
+                                &tmp)) {
+               pr_debug(" * %s missing min-residency-us property\n",
+                        state_node->full_name);
+               return -EINVAL;
+       }
+
+       val += tmp;
+
+       if (of_property_read_u32(state_node, "exit-latency-us",
+                                &tmp)) {
+               pr_debug(" * %s missing exit-latency-us property\n",
+                            state_node->full_name);
+               return -EINVAL;
+       }
+
+       val += tmp;

Sorry if i'm rehashing old stuff, but I prefer not to use the
min-residency + exit-latency to sort.  I saw Rob's comment suggesting it
and your reply.  I'm not sure when it was decided.

Would it be possible to sort instead based on the order in the
cpus->cpu-idle-states?  If not, my preference would be to either use
index like you had before, or specify another sort order / rank value.

I think there's potential for us to create lower power states that
have lower min-residencies (reduced power consumption in the state,
allowing us to more quickly recover the higher entrance cost)
with higher exit latencies in such a way that the formula would not
sort as we expect.  Having a separate value would allow us to control
the sorting in those cases.

Ok, so adding to my previous comment, would exit_latency by itself be
enough ? Can we think of a system where ranking by exit_latency is wrong ?

If yes, we need a power rank, and this patch is wrong too:

http://marc.info/?l=linux-kernel&m=139924292401056&w=2

If answer is not, I can just rely on exit_latency to sort the states.

Lorenzo