cpuidle: fix the menu governor to boost IO performance

Fix the menu idle governor which balances power savings, energy efficiency
and performance impact.

The reason for a reworked governor is that there have been serious
performance issues reported with the existing code on Nehalem server
systems.

To show this I'm sure Andrew wants to see benchmark results:
(benchmark is "fio", "no cstates" is using "idle=poll")

		no cstates	current linux	new algorithm
1 disk		107 Mb/s	85 Mb/s		105 Mb/s
2 disks		215 Mb/s	123 Mb/s	209 Mb/s
12 disks	590 Mb/s	320 Mb/s	585 Mb/s

In various power benchmark measurements, no degredation was found by our
measurement&diagnostics team.  Obviously a small percentage more power was
used in the "fio" benchmark, due to the much higher performance.

While it would be a novel idea to describe the new algorithm in this
commit message, I cheaped out and described it in comments in the code
instead.

[changes since first post: spelling fixes from akpm, review feedback,
folded menu-tng into menu.c]

Signed-off-by: Arjan van de Ven <arjan@linux.intel.com>
Cc: Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>
Cc: Len Brown <lenb@kernel.org>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Yanmin Zhang <yanmin_zhang@linux.intel.com>
Acked-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

1 files changed, 212 insertions, 39 deletions

diff --git a/drivers/cpuidle/governors/menu.c b/drivers/cpuidle/governors/menu.c
index f1df59f59a37..9f3d77532ab9 100644
--- a/drivers/cpuidle/governors/menu.c
+++ b/drivers/cpuidle/governors/menu.c
@@ -2,8 +2,12 @@
  * menu.c - the menu idle governor
  *
  * Copyright (C) 2006-2007 Adam Belay <abelay@novell.com>
+ * Copyright (C) 2009 Intel Corporation
+ * Author:
+ *        Arjan van de Ven <arjan@linux.intel.com>
  *
- * This code is licenced under the GPL.
+ * This code is licenced under the GPL version 2 as described
+ * in the COPYING file that acompanies the Linux Kernel.
  */
 
 #include <linux/kernel.h>
@@ -13,20 +17,153 @@
 #include <linux/ktime.h>
 #include <linux/hrtimer.h>
 #include <linux/tick.h>
+#include <linux/sched.h>
 
-#define BREAK_FUZZ      4       /* 4 us */
-#define PRED_HISTORY_PCT        50
+#define BUCKETS 12
+#define RESOLUTION 1024
+#define DECAY 4
+#define MAX_INTERESTING 50000
+
+/*
+ * Concepts and ideas behind the menu governor
+ *
+ * For the menu governor, there are 3 decision factors for picking a C
+ * state:
+ * 1) Energy break even point
+ * 2) Performance impact
+ * 3) Latency tolerance (from pmqos infrastructure)
+ * These these three factors are treated independently.
+ *
+ * Energy break even point
+ * -----------------------
+ * C state entry and exit have an energy cost, and a certain amount of time in
+ * the  C state is required to actually break even on this cost. CPUIDLE
+ * provides us this duration in the "target_residency" field. So all that we
+ * need is a good prediction of how long we'll be idle. Like the traditional
+ * menu governor, we start with the actual known "next timer event" time.
+ *
+ * Since there are other source of wakeups (interrupts for example) than
+ * the next timer event, this estimation is rather optimistic. To get a
+ * more realistic estimate, a correction factor is applied to the estimate,
+ * that is based on historic behavior. For example, if in the past the actual
+ * duration always was 50% of the next timer tick, the correction factor will
+ * be 0.5.
+ *
+ * menu uses a running average for this correction factor, however it uses a
+ * set of factors, not just a single factor. This stems from the realization
+ * that the ratio is dependent on the order of magnitude of the expected
+ * duration; if we expect 500 milliseconds of idle time the likelihood of
+ * getting an interrupt very early is much higher than if we expect 50 micro
+ * seconds of idle time. A second independent factor that has big impact on
+ * the actual factor is if there is (disk) IO outstanding or not.
+ * (as a special twist, we consider every sleep longer than 50 milliseconds
+ * as perfect; there are no power gains for sleeping longer than this)
+ *
+ * For these two reasons we keep an array of 12 independent factors, that gets
+ * indexed based on the magnitude of the expected duration as well as the
+ * "is IO outstanding" property.
+ *
+ * Limiting Performance Impact
+ * ---------------------------
+ * C states, especially those with large exit latencies, can have a real
+ * noticable impact on workloads, which is not acceptable for most sysadmins,
+ * and in addition, less performance has a power price of its own.
+ *
+ * As a general rule of thumb, menu assumes that the following heuristic
+ * holds:
+ *     The busier the system, the less impact of C states is acceptable
+ *
+ * This rule-of-thumb is implemented using a performance-multiplier:
+ * If the exit latency times the performance multiplier is longer than
+ * the predicted duration, the C state is not considered a candidate
+ * for selection due to a too high performance impact. So the higher
+ * this multiplier is, the longer we need to be idle to pick a deep C
+ * state, and thus the less likely a busy CPU will hit such a deep
+ * C state.
+ *
+ * Two factors are used in determing this multiplier:
+ * a value of 10 is added for each point of "per cpu load average" we have.
+ * a value of 5 points is added for each process that is waiting for
+ * IO on this CPU.
+ * (these values are experimentally determined)
+ *
+ * The load average factor gives a longer term (few seconds) input to the
+ * decision, while the iowait value gives a cpu local instantanious input.
+ * The iowait factor may look low, but realize that this is also already
+ * represented in the system load average.
+ *
+ */
 
 struct menu_device {
         int             last_state_idx;
 
         unsigned int    expected_us;
-        unsigned int    predicted_us;
-        unsigned int    current_predicted_us;
-        unsigned int    last_measured_us;
-        unsigned int    elapsed_us;
+        u64             predicted_us;
+        unsigned int    measured_us;
+        unsigned int    exit_us;
+        unsigned int    bucket;
+        u64             correction_factor[BUCKETS];
 };
 
+
+#define LOAD_INT(x) ((x) >> FSHIFT)
+#define LOAD_FRAC(x) LOAD_INT(((x) & (FIXED_1-1)) * 100)
+
+static int get_loadavg(void)
+{
+        unsigned long this = this_cpu_load();
+
+
+        return LOAD_INT(this) * 10 + LOAD_FRAC(this) / 10;
+}
+
+static inline int which_bucket(unsigned int duration)
+{
+        int bucket = 0;
+
+        /*
+         * We keep two groups of stats; one with no
+         * IO pending, one without.
+         * This allows us to calculate
+         * E(duration)|iowait
+         */
+        if (nr_iowait_cpu())
+                bucket = BUCKETS/2;
+
+        if (duration < 10)
+                return bucket;
+        if (duration < 100)
+                return bucket + 1;
+        if (duration < 1000)
+                return bucket + 2;
+        if (duration < 10000)
+                return bucket + 3;
+        if (duration < 100000)
+                return bucket + 4;
+        return bucket + 5;
+}
+
+/*
+ * Return a multiplier for the exit latency that is intended
+ * to take performance requirements into account.
+ * The more performance critical we estimate the system
+ * to be, the higher this multiplier, and thus the higher
+ * the barrier to go to an expensive C state.
+ */
+static inline int performance_multiplier(void)
+{
+        int mult = 1;
+
+        /* for higher loadavg, we are more reluctant */
+
+        mult += 2 * get_loadavg();
+
+        /* for IO wait tasks (per cpu!) we add 5x each */
+        mult += 10 * nr_iowait_cpu();
+
+        return mult;
+}
+
 static DEFINE_PER_CPU(struct menu_device, menu_devices);
 
 /**
@@ -38,37 +175,59 @@ static int menu_select(struct cpuidle_device *dev)
         struct menu_device *data = &__get_cpu_var(menu_devices);
         int latency_req = pm_qos_requirement(PM_QOS_CPU_DMA_LATENCY);
         int i;
+        int multiplier;
+
+        data->last_state_idx = 0;
+        data->exit_us = 0;
 
         /* Special case when user has set very strict latency requirement */
-        if (unlikely(latency_req == 0)) {
-                data->last_state_idx = 0;
+        if (unlikely(latency_req == 0))
                 return 0;
-        }
 
-        /* determine the expected residency time */
+        /* determine the expected residency time, round up */
         data->expected_us =
-                (u32) ktime_to_ns(tick_nohz_get_sleep_length()) / 1000;
+            DIV_ROUND_UP((u32)ktime_to_ns(tick_nohz_get_sleep_length()), 1000);
+
+
+        data->bucket = which_bucket(data->expected_us);
+
+        multiplier = performance_multiplier();
+
+        /*
+         * if the correction factor is 0 (eg first time init or cpu hotplug
+         * etc), we actually want to start out with a unity factor.
+         */
+        if (data->correction_factor[data->bucket] == 0)
+                data->correction_factor[data->bucket] = RESOLUTION * DECAY;
+
+        /* Make sure to round up for half microseconds */
+        data->predicted_us = DIV_ROUND_CLOSEST(
+                data->expected_us * data->correction_factor[data->bucket],
+                RESOLUTION * DECAY);
+
+        /*
+         * We want to default to C1 (hlt), not to busy polling
+         * unless the timer is happening really really soon.
+         */
+        if (data->expected_us > 5)
+                data->last_state_idx = CPUIDLE_DRIVER_STATE_START;
 
-        /* Recalculate predicted_us based on prediction_history_pct */
-        data->predicted_us *= PRED_HISTORY_PCT;
-        data->predicted_us += (100 - PRED_HISTORY_PCT) *
-                                data->current_predicted_us;
-        data->predicted_us /= 100;
 
         /* find the deepest idle state that satisfies our constraints */
-        for (i = CPUIDLE_DRIVER_STATE_START + 1; i < dev->state_count; i++) {
+        for (i = CPUIDLE_DRIVER_STATE_START; i < dev->state_count; i++) {
                 struct cpuidle_state *s = &dev->states[i];
 
-                if (s->target_residency > data->expected_us)
-                        break;
                 if (s->target_residency > data->predicted_us)
                         break;
                 if (s->exit_latency > latency_req)
                         break;
+                if (s->exit_latency * multiplier > data->predicted_us)
+                        break;
+                data->exit_us = s->exit_latency;
+                data->last_state_idx = i;
         }
 
-        data->last_state_idx = i - 1;
-        return i - 1;
+        return data->last_state_idx;
 }
 
 /**
@@ -85,35 +244,49 @@ static void menu_reflect(struct cpuidle_device *dev)
         unsigned int last_idle_us = cpuidle_get_last_residency(dev);
         struct cpuidle_state *target = &dev->states[last_idx];
         unsigned int measured_us;
+        u64 new_factor;
 
         /*
          * Ugh, this idle state doesn't support residency measurements, so we
          * are basically lost in the dark.  As a compromise, assume we slept
-         * for one full standard timer tick.  However, be aware that this
-         * could potentially result in a suboptimal state transition.
+         * for the whole expected time.
          */
         if (unlikely(!(target->flags & CPUIDLE_FLAG_TIME_VALID)))
-                last_idle_us = USEC_PER_SEC / HZ;
+                last_idle_us = data->expected_us;
+
+
+        measured_us = last_idle_us;
 
         /*
-         * measured_us and elapsed_us are the cumulative idle time, since the
-         * last time we were woken out of idle by an interrupt.
+         * We correct for the exit latency; we are assuming here that the
+         * exit latency happens after the event that we're interested in.
          */
-        if (data->elapsed_us <= data->elapsed_us + last_idle_us)
-                measured_us = data->elapsed_us + last_idle_us;
+        if (measured_us > data->exit_us)
+                measured_us -= data->exit_us;
+
+
+        /* update our correction ratio */
+
+        new_factor = data->correction_factor[data->bucket]
+                        * (DECAY - 1) / DECAY;
+
+        if (data->expected_us > 0 && data->measured_us < MAX_INTERESTING)
+                new_factor += RESOLUTION * measured_us / data->expected_us;
         else
-                measured_us = -1;
+                /*
+                 * we were idle so long that we count it as a perfect
+                 * prediction
+                 */
+                new_factor += RESOLUTION;
 
-        /* Predict time until next break event */
-        data->current_predicted_us = max(measured_us, data->last_measured_us);
+        /*
+         * We don't want 0 as factor; we always want at least
+         * a tiny bit of estimated time.
+         */
+        if (new_factor == 0)
+                new_factor = 1;
 
-        if (last_idle_us + BREAK_FUZZ <
-            data->expected_us - target->exit_latency) {
-                data->last_measured_us = measured_us;
-                data->elapsed_us = 0;
-        } else {
-                data->elapsed_us = measured_us;
-        }
+        data->correction_factor[data->bucket] = new_factor;
 }
 
 /**