[PATCH v2 5/6] KVM-GST: adjust scheduler cpu power

[Glauber Costa]

This is a first proposal for using steal time information
to influence the scheduler. There are a lot of optimizations
and fine grained adjustments to be done, but it is working reasonably
so far for me (mostly)

With this patch (and some host pinnings to demonstrate the situation),
two vcpus with very different steal time (Say 80 % vs 1 %) will not get
an even distribution of processes. This is a situation that can naturally
arise, specially in overcommited scenarios. Previosly, the guest scheduler
would wrongly think that all cpus have the same ability to run processes,
lowering the overall throughput.

Signed-off-by: Glauber Costa <glommer@xxxxxxxxxx>
CC: Rik van Riel <riel@xxxxxxxxxx>
CC: Jeremy Fitzhardinge <jeremy.fitzhardinge@xxxxxxxxxx>
CC: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
CC: Avi Kivity <avi@xxxxxxxxxx>
---
 arch/x86/Kconfig        |   12 +++++++
 kernel/sched.c          |   85 ++++++++++++++++++++++++++++++++++++++++-------
 kernel/sched_features.h |    4 +-
 3 files changed, 87 insertions(+), 14 deletions(-)

diff --git a/arch/x86/Kconfig b/arch/x86/Kconfig
index 3ed5ad9..8f7a666 100644
--- a/arch/x86/Kconfig
+++ b/arch/x86/Kconfig
@@ -515,6 +515,18 @@ menuconfig PARAVIRT_GUEST
 
 if PARAVIRT_GUEST
 
+config PARAVIRT_TIME_ACCOUNTING
+	bool "Paravirtual steal time accounting"
+	select PARAVIRT
+	default n
+	---help---
+	  Select this option to enable fine granularity task steal time 
+	  accounting. Time spent executing other tasks in parallel with
+          the current vCPU is discounted from the vCPU power. To account for
+	  that, there can be a small performance impact.
+
+	  If in doubt, say N here.
+
 source "arch/x86/xen/Kconfig"
 
 config KVM_CLOCK
diff --git a/kernel/sched.c b/kernel/sched.c
index 7765e9d..40df0d8 100644
--- a/kernel/sched.c
+++ b/kernel/sched.c
@@ -524,6 +524,9 @@ struct rq {
 	u64 prev_irq_time;
 #endif
 
+#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
+	u64 prev_steal_time;
+#endif
 	/* calc_load related fields */
 	unsigned long calc_load_update;
 	long calc_load_active;
@@ -1780,6 +1783,54 @@ static void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
 	dec_nr_running(rq);
 }
 
+#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
+static DEFINE_PER_CPU(u64, cpu_steal_time);
+
+#ifndef CONFIG_64BIT
+static DEFINE_PER_CPU(seqcount_t, steal_time_seq);
+
+static inline void steal_time_write_begin(void)
+{
+	__this_cpu_inc(steal_time_seq.sequence);
+	smp_wmb();
+}
+
+static inline void steal_time_write_end(void)
+{
+	smp_wmb();
+	__this_cpu_inc(steal_time_seq.sequence);
+}
+
+static inline u64 steal_time_read(int cpu)
+{
+	u64 steal_time;
+	unsigned seq;
+
+	do {
+		seq = read_seqcount_begin(&per_cpu(steal_time_seq, cpu));
+		steal_time = per_cpu(cpu_steal_time, cpu);
+	} while (read_seqcount_retry(&per_cpu(steal_time_seq, cpu), seq));
+
+	return steal_time;
+}
+#else /* CONFIG_64BIT */
+static inline void steal_time_write_begin(void)
+{
+}
+
+static inline void steal_time_write_end(void)
+{
+}
+
+static inline u64 steal_time_read(int cpu)
+{
+	return per_cpu(cpu_steal_time, cpu);
+}
+
+#endif /* CONFIG_64BIT */
+
+#endif
+
 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
 
 /*
@@ -1888,10 +1939,13 @@ void account_system_vtime(struct task_struct *curr)
 }
 EXPORT_SYMBOL_GPL(account_system_vtime);
 
+#endif /* CONFIG_IRQ_TIME_ACCOUNTING */
+
 static void update_rq_clock_task(struct rq *rq, s64 delta)
 {
-	s64 irq_delta;
+	s64 irq_delta = 0, steal = 0;
 
+#ifdef CONFIG_IRQ_TIME_ACCOUNTING
 	irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
 
 	/*
@@ -1914,20 +1968,22 @@ static void update_rq_clock_task(struct rq *rq, s64 delta)
 
 	rq->prev_irq_time += irq_delta;
 	delta -= irq_delta;
-	rq->clock_task += delta;
-
-	if (irq_delta && sched_feat(NONIRQ_POWER))
-		sched_rt_avg_update(rq, irq_delta);
-}
+#endif
+#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
+	steal = steal_time_read(cpu_of(rq)) - rq->prev_steal_time;
+  
+	if (steal > delta)
+		steal = delta;
+	rq->prev_steal_time += steal;
 
-#else /* CONFIG_IRQ_TIME_ACCOUNTING */
+	delta -= steal;
+#endif
 
-static void update_rq_clock_task(struct rq *rq, s64 delta)
-{
 	rq->clock_task += delta;
-}
 
-#endif /* CONFIG_IRQ_TIME_ACCOUNTING */
+	if ((irq_delta + steal) && sched_feat(NONTASK_POWER))
+		sched_rt_avg_update(rq, irq_delta + steal);
+}
 
 #include "sched_idletask.c"
 #include "sched_fair.c"
@@ -3536,6 +3592,11 @@ static int touch_steal_time(int is_idle)
 
 	if (st) {
 		account_steal_time(st);
+#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
+		steal_time_write_begin();
+		__this_cpu_add(cpu_steal_time, steal);
+		steal_time_write_end();
+#endif
 		return 1;
 	}
 	return 0;
diff --git a/kernel/sched_features.h b/kernel/sched_features.h
index 68e69ac..194fc6d 100644
--- a/kernel/sched_features.h
+++ b/kernel/sched_features.h
@@ -61,6 +61,6 @@ SCHED_FEAT(LB_BIAS, 1)
 SCHED_FEAT(OWNER_SPIN, 1)
 
 /*
- * Decrement CPU power based on irq activity
+ * Decrement CPU power based on time not spent running tasks
  */
-SCHED_FEAT(NONIRQ_POWER, 1)
+SCHED_FEAT(NONTASK_POWER, 1)
-- 
1.7.2.3

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/