]> git.openfabrics.org - ~shefty/rdma-dev.git/commitdiff
cpuidle: Quickly notice prediction failure in general case
authorYouquan Song <youquan.song@intel.com>
Fri, 26 Oct 2012 10:26:50 +0000 (12:26 +0200)
committerRafael J. Wysocki <rafael.j.wysocki@intel.com>
Wed, 14 Nov 2012 23:34:20 +0000 (00:34 +0100)
The prediction for future is difficult and when the cpuidle governor prediction
fails and govenor possibly choose the shallower C-state than it should. How to
quickly notice and find the failure becomes important for power saving.

The patch extends to general case that prediction logic get a small predicted
residency, so it choose a shallow C-state though the expected residency is large
. Once the prediction will be fail, the CPU will keep staying at shallow C-state
for a long time. Acutally, the CPU has change enter into deep C-state.
So when the expected residency is long enough but governor choose a shallow
C-state, an timer will be added in order to monitor if the prediction failure.

When C-state is waken up prior to the adding timer, the timer will be cancelled
initiatively. When the timer is triggered and menu governor will quickly notice
prediction failure and re-evaluates deeper C-states possibility.

Signed-off-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Youquan Song <youquan.song@intel.com>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
drivers/cpuidle/governors/menu.c

index 37c0ff6c805c5a42283eaa4d439dfd76a9b47bae..43a54fd6bfa2f31aacbd7a6ad5f22b5e1df0be15 100644 (file)
@@ -34,7 +34,7 @@
 static DEFINE_PER_CPU(struct hrtimer, menu_hrtimer);
 static DEFINE_PER_CPU(int, hrtimer_status);
 /* menu hrtimer mode */
-enum {MENU_HRTIMER_STOP, MENU_HRTIMER_REPEAT};
+enum {MENU_HRTIMER_STOP, MENU_HRTIMER_REPEAT, MENU_HRTIMER_GENERAL};
 
 /*
  * Concepts and ideas behind the menu governor
@@ -116,6 +116,13 @@ enum {MENU_HRTIMER_STOP, MENU_HRTIMER_REPEAT};
  *
  */
 
+/*
+ * The C-state residency is so long that is is worthwhile to exit
+ * from the shallow C-state and re-enter into a deeper C-state.
+ */
+static unsigned int perfect_cstate_ms __read_mostly = 30;
+module_param(perfect_cstate_ms, uint, 0000);
+
 struct menu_device {
        int             last_state_idx;
        int             needs_update;
@@ -216,6 +223,16 @@ EXPORT_SYMBOL_GPL(menu_hrtimer_cancel);
 static enum hrtimer_restart menu_hrtimer_notify(struct hrtimer *hrtimer)
 {
        int cpu = smp_processor_id();
+       struct menu_device *data = &per_cpu(menu_devices, cpu);
+
+       /* In general case, the expected residency is much larger than
+        *  deepest C-state target residency, but prediction logic still
+        *  predicts a small predicted residency, so the prediction
+        *  history is totally broken if the timer is triggered.
+        *  So reset the correction factor.
+        */
+       if (per_cpu(hrtimer_status, cpu) == MENU_HRTIMER_GENERAL)
+               data->correction_factor[data->bucket] = RESOLUTION * DECAY;
 
        per_cpu(hrtimer_status, cpu) = MENU_HRTIMER_STOP;
 
@@ -353,6 +370,7 @@ static int menu_select(struct cpuidle_driver *drv, struct cpuidle_device *dev)
        /* not deepest C-state chosen for low predicted residency */
        if (low_predicted) {
                unsigned int timer_us = 0;
+               unsigned int perfect_us = 0;
 
                /*
                 * Set a timer to detect whether this sleep is much
@@ -363,12 +381,26 @@ static int menu_select(struct cpuidle_driver *drv, struct cpuidle_device *dev)
                 */
                timer_us = 2 * (data->predicted_us + MAX_DEVIATION);
 
+               perfect_us = perfect_cstate_ms * 1000;
+
                if (repeat && (4 * timer_us < data->expected_us)) {
                        hrtimer_start(hrtmr, ns_to_ktime(1000 * timer_us),
                                HRTIMER_MODE_REL_PINNED);
                        /* In repeat case, menu hrtimer is started */
                        per_cpu(hrtimer_status, cpu) = MENU_HRTIMER_REPEAT;
+               } else if (perfect_us < data->expected_us) {
+                       /*
+                        * The next timer is long. This could be because
+                        * we did not make a useful prediction.
+                        * In that case, it makes sense to re-enter
+                        * into a deeper C-state after some time.
+                        */
+                       hrtimer_start(hrtmr, ns_to_ktime(1000 * timer_us),
+                               HRTIMER_MODE_REL_PINNED);
+                       /* In general case, menu hrtimer is started */
+                       per_cpu(hrtimer_status, cpu) = MENU_HRTIMER_GENERAL;
                }
+
        }
 
        return data->last_state_idx;