[RFC PATCH v2 1/6] sched/fair: Add util_guest for tasks

[David Dai]

For virtualization usecases, util_est and util_avg currently tracked
on the host aren't sufficient to accurately represent the workload on
vCPU threads, which results in poor frequency selection and performance.
For example, when a large workload migrates from a busy vCPU thread to
an idle vCPU thread, it incurs additional DVFS ramp-up latencies
as util accumulates.

Introduce a new "util_guest" member as an additional PELT signal that's
independently updated by the guest. When used, it's max aggregated to
provide a boost to both task_util and task_util_est.

Updating task_util and task_util_est will ensure:
-Better task placement decisions for vCPU threads on the host
-Correctly updating util_est.ewma during dequeue
-Additive util with other threads on the same runqueue for more
accurate frequency responses

Co-developed-by: Saravana Kannan <saravanak@xxxxxxxxxx>
Signed-off-by: Saravana Kannan <saravanak@xxxxxxxxxx>
Signed-off-by: David Dai <davidai@xxxxxxxxxx>
---
 include/linux/sched.h | 11 +++++++++++
 kernel/sched/core.c   | 18 +++++++++++++++++-
 kernel/sched/fair.c   | 15 +++++++++++++--
 3 files changed, 41 insertions(+), 3 deletions(-)

diff --git a/include/linux/sched.h b/include/linux/sched.h
index 63d242164b1a..d8c346fcdf52 100644
--- a/include/linux/sched.h
+++ b/include/linux/sched.h
@@ -445,6 +445,16 @@ struct util_est {
 #define UTIL_AVG_UNCHANGED		0x80000000
 } __attribute__((__aligned__(sizeof(u64))));
 
+/*
+ * For sched_setattr_nocheck() (kernel) only
+ *
+ * Allow vCPU threads to use UTIL_GUEST as a way to hint the scheduler with more
+ * accurate utilization info. This is useful when guest kernels have some way of
+ * tracking its own runqueue's utilization.
+ *
+ */
+#define SCHED_FLAG_UTIL_GUEST   0x20000000
+
 /*
  * The load/runnable/util_avg accumulates an infinite geometric series
  * (see __update_load_avg_cfs_rq() in kernel/sched/pelt.c).
@@ -499,6 +509,7 @@ struct sched_avg {
 	unsigned long			load_avg;
 	unsigned long			runnable_avg;
 	unsigned long			util_avg;
+	unsigned long			util_guest;
 	struct util_est			util_est;
 } ____cacheline_aligned;
 
diff --git a/kernel/sched/core.c b/kernel/sched/core.c
index 0d18c3969f90..7700ef5610c1 100644
--- a/kernel/sched/core.c
+++ b/kernel/sched/core.c
@@ -2024,6 +2024,16 @@ static inline void uclamp_post_fork(struct task_struct *p) { }
 static inline void init_uclamp(void) { }
 #endif /* CONFIG_UCLAMP_TASK */
 
+static void __setscheduler_task_util(struct task_struct *p,
+				  const struct sched_attr *attr)
+{
+
+	if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_GUEST)))
+		return;
+
+	p->se.avg.util_guest = attr->sched_util_min;
+}
+
 bool sched_task_on_rq(struct task_struct *p)
 {
 	return task_on_rq_queued(p);
@@ -7561,7 +7571,7 @@ static int __sched_setscheduler(struct task_struct *p,
 			return -EINVAL;
 	}
 
-	if (attr->sched_flags & ~(SCHED_FLAG_ALL | SCHED_FLAG_SUGOV))
+	if (attr->sched_flags & ~(SCHED_FLAG_ALL | SCHED_FLAG_SUGOV | SCHED_FLAG_UTIL_GUEST))
 		return -EINVAL;
 
 	/*
@@ -7583,6 +7593,9 @@ static int __sched_setscheduler(struct task_struct *p,
 		if (attr->sched_flags & SCHED_FLAG_SUGOV)
 			return -EINVAL;
 
+		if (attr->sched_flags & SCHED_FLAG_UTIL_GUEST)
+			return -EINVAL;
+
 		retval = security_task_setscheduler(p);
 		if (retval)
 			return retval;
@@ -7629,6 +7642,8 @@ static int __sched_setscheduler(struct task_struct *p,
 			goto change;
 		if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)
 			goto change;
+		if (attr->sched_flags & SCHED_FLAG_UTIL_GUEST)
+			goto change;
 
 		p->sched_reset_on_fork = reset_on_fork;
 		retval = 0;
@@ -7718,6 +7733,7 @@ static int __sched_setscheduler(struct task_struct *p,
 		__setscheduler_prio(p, newprio);
 	}
 	__setscheduler_uclamp(p, attr);
+	__setscheduler_task_util(p, attr);
 
 	if (queued) {
 		/*
diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index 6986ea31c984..998649554344 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -4276,14 +4276,16 @@ static int newidle_balance(struct rq *this_rq, struct rq_flags *rf);
 
 static inline unsigned long task_util(struct task_struct *p)
 {
-	return READ_ONCE(p->se.avg.util_avg);
+	return max(READ_ONCE(p->se.avg.util_avg),
+			READ_ONCE(p->se.avg.util_guest));
 }
 
 static inline unsigned long _task_util_est(struct task_struct *p)
 {
 	struct util_est ue = READ_ONCE(p->se.avg.util_est);
 
-	return max(ue.ewma, (ue.enqueued & ~UTIL_AVG_UNCHANGED));
+	return max_t(unsigned long, READ_ONCE(p->se.avg.util_guest),
+			max(ue.ewma, (ue.enqueued & ~UTIL_AVG_UNCHANGED)));
 }
 
 static inline unsigned long task_util_est(struct task_struct *p)
@@ -6242,6 +6244,15 @@ enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
 	 */
 	util_est_enqueue(&rq->cfs, p);
 
+	/*
+	 * The normal code path for host thread enqueue doesn't take into
+	 * account guest task migrations when updating cpufreq util.
+	 * So, always update the cpufreq when a vCPU thread has a
+	 * non-zero util_guest value.
+	 */
+	if (READ_ONCE(p->se.avg.util_guest))
+		cpufreq_update_util(rq, 0);
+
 	/*
 	 * If in_iowait is set, the code below may not trigger any cpufreq
 	 * utilization updates, so do it here explicitly with the IOWAIT flag
-- 
2.40.0.348.gf938b09366-goog