On 2 July 2012 11:11, Peter Zijlstra [off-list ref] wrote:

On Wed, 2012-06-20 at 17:19 +0200, Vincent Guittot wrote:

quoted

+#ifdef CONFIG_OF
+struct cpu_efficiency {
+       const char *compatible;
+       unsigned long efficiency;
+};
+
+/*
+ * Table of relative efficiency of each processors
+ * The efficiency value must fit in 20bit. The final
+ * cpu_scale value must be in the range
+ * 0 < cpu_scale < 2*SCHED_POWER_SCALE.

This wants a why.. I suspects its to do with keeping capacity on 1.

Yes, that's it. Now, Regarding the div_round_closest that is used in
the scheduler to compute the capacity, It should rather stay in the
range
0 < cpu_scale < 3*SCHED_POWER_SCALE/2.

quoted

+ * Processors that are not defined in the table,
+ * use the default SCHED_POWER_SCALE value for cpu_scale.
+ */
+struct cpu_efficiency table_efficiency[] = {
+       {"arm,cortex-a15", 3891},
+       {"arm,cortex-a7",  2048},
+       {NULL, },
+};
+
+struct cpu_capacity {
+       unsigned long hwid;
+       unsigned long capacity;
+};
+
+struct cpu_capacity *cpu_capacity;
+
+unsigned long middle_capacity = 1;

It would be very nice to not have to learn to read device-tree nonsense
to work on the scheduler, how about something like this:?

/*
 * Iterate all cpus and set the efficiency (as per table_efficiency)
 * also calculate the middle efficiency:
 *   (max{eff_i} - min{eff_i}) / 2
 * This is later used to scale the cpu_power field such that an
 * 'average' cpu is of middle power. Also see the comments near
 * table_efficiency[] and update_cpu_power().
 */

ok

quoted

+static void __init parse_dt_topology(void)
+{
+       struct cpu_efficiency *cpu_eff;
+       struct device_node *cn = NULL;
+       unsigned long min_capacity = (unsigned long)(-1);
+       unsigned long max_capacity = 0;
+       unsigned long capacity = 0;
+       int alloc_size, cpu = 0;
+
+       alloc_size = nr_cpu_ids * sizeof(struct cpu_capacity);
+       cpu_capacity = (struct cpu_capacity *)kzalloc(alloc_size, GFP_NOWAIT);
+
+       while ((cn = of_find_node_by_type(cn, "cpu"))) {
+               const u32 *rate, *reg;
+               int len;
+
+               if (cpu >= num_possible_cpus())
+                       break;
+
+               for (cpu_eff = table_efficiency; cpu_eff->compatible; cpu_eff++)
+                       if (of_device_is_compatible(cn, cpu_eff->compatible))
+                               break;
+
+               if (cpu_eff->compatible == NULL)
+                       continue;
+
+               rate = of_get_property(cn, "clock-frequency", &len);
+               if (!rate || len != 4) {
+                       pr_err("%s missing clock-frequency property\n",
+                               cn->full_name);
+                       continue;
+               }
+
+               reg = of_get_property(cn, "reg", &len);
+               if (!reg || len != 4) {
+                       pr_err("%s missing reg property\n", cn->full_name);
+                       continue;
+               }
+
+               capacity = ((be32_to_cpup(rate)) >> 20) * cpu_eff->efficiency;
+
+               /* Save min capacity of the system */
+               if (capacity < min_capacity)
+                       min_capacity = capacity;
+
+               /* Save max capacity of the system */
+               if (capacity > max_capacity)
+                       max_capacity = capacity;
+
+               cpu_capacity[cpu].capacity = capacity;
+               cpu_capacity[cpu++].hwid = be32_to_cpup(reg);
+       }
+
+       if (cpu < num_possible_cpus())
+               cpu_capacity[cpu].hwid = (unsigned long)(-1);
+
+       middle_capacity = (min_capacity + max_capacity) >> 11;
+}
+
+void update_cpu_power(unsigned int cpu, unsigned long hwid)
+{
+       unsigned int idx = 0;
+
+       /* look for the cpu's hwid in the cpu capacity table */

This smells like an O(n^2) loop.. ARM has only small cpu counts so this
isn't an immediate issue, would still be nice to make a note of it
though.

Yes, This function is called for each cpu. I will add a comment about
that and also about the fact that the complete sequence is done only
once.
I will also add an optimization for system with identical CPUs and DT
information

quoted

+       for (idx = 0; idx < num_possible_cpus(); idx++) {
+               if (cpu_capacity[idx].hwid == hwid)
+                       break;
+
+               if (cpu_capacity[idx].hwid == -1)
+                       return;
+       }
+
+       if (idx == num_possible_cpus())
+               return;
+
+       set_power_scale(cpu, cpu_capacity[idx].capacity / middle_capacity);

OK, but there's no guarantee here you'll stay within that
[1,2*SCHED_POWER_SCALE-1] range. This might want a comment and or
runtime verification so that when people extend the table_efficiency[]
wrongly we'll get notice, humm?

I will add more comments but we can't be higher than 2047. The max
value for a cpu_power will be :
max / ( (min + max) / 2^11 ) which is equal to
max / (min + max) * 2^11 so the cpu_scale is smaller than 2^11  as min
is never equal to 0

We can have a cpu_power of 0 is a CPU has 2048 times more capacity
than another one in the system but I'm not sure that it's a realistic
use case

quoted

+       printk(KERN_INFO "CPU%u: update cpu_power %lu\n",
+               cpu, arch_scale_freq_power(NULL, cpu));
+}