I had a sudden surge of inspiration and decided to follow up on this
one a bit: 

quoted text
> 
> 5. Second core
> 
> Disabling the second core makes the noise go away. This might be a subset of 1., as I've been told that a stopped core enters C1.
> 

I have now found a very hacky workaround that is slightly better than
disabling C3 altogether; making C3 exclusive to one core at a time. It
seems to kill the noise and the system now spends 50% in C3, instead of
0%.

I read somewhere that the Intel Duo:s have an extra low power mode that
is entered if both cores are suspended at the same time (disabling the
shared cache or something like that I guess). So perhaps that mode is
the culprit. I'm hoping there's a way to disable just that feature. Do
any of you know a way? Venkatesh perhaps?

Here's the patch I'm currently running at least:

diff --git a/drivers/cpuidle/governors/menu.c b/drivers/cpuidle/governors/menu.c
index 78d77c5..9a859fb 100644
--- a/drivers/cpuidle/governors/menu.c
+++ b/drivers/cpuidle/governors/menu.c
@@ -16,6 +16,8 @@
 
 #define BREAK_FUZZ     4       /* 4 us */
 
+static atomic_t cur_deep_cpu = ATOMIC_INIT(-1);
+
 struct menu_device {
        int             last_state_idx;
 
@@ -50,6 +52,13 @@ static int menu_select(struct cpuidle_device *dev)
                        break;
                if (s->exit_latency > pm_qos_requirement(PM_QOS_CPU_DMA_LATENCY))
                        break;
+
+               if (s->flags & CPUIDLE_FLAG_DEEP) {
+                       rmb();
+                       if (atomic_cmpxchg(&cur_deep_cpu, -1, dev->cpu) != -1)
+                               break;
+                       wmb();
+               }
        }
 
        data->last_state_idx = i - 1;
@@ -80,6 +89,11 @@ static void menu_reflect(struct cpuidle_device *dev)
        if (!(target->flags & CPUIDLE_FLAG_TIME_VALID))
                measured_us = USEC_PER_SEC / HZ;
 
+       if (atomic_read(&cur_deep_cpu) == dev->cpu) {
+               atomic_set(&cur_deep_cpu, -1);
+               wmb();
+       }
+
        /* Predict time remaining until next break event */
        if (measured_us + BREAK_FUZZ < data->expected_us - target->exit_latency) {
                data->predicted_us = max(measured_us, data->last_measured_us);


-- 
     -- Pierre Ossman

  Linux kernel, MMC maintainer        http://www.kernel.org
  PulseAudio, core developer          http://pulseaudio.org
  rdesktop, core developer          http://www.rdesktop.org
--
unsubscribe notice
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/