sched/walt: Improve the scheduler

This change is for general scheduler improvement.

Change-Id: I4da8fd848f9cd43d510ac2ae63605f051e723775
Signed-off-by: Pavankumar Kondeti <pkondeti@codeaurora.org>

include/linux/sched.h

@@ -618,6 +618,7 @@ struct walt_task_struct {
 	int				boost;
 	bool				wake_up_idle;
 	bool				misfit;
+	bool				rtg_high_prio;
 	bool				low_latency;
 	u64				boost_period;
 	u64				boost_expires;

include/trace/events/sched.h

@@ -927,6 +927,7 @@ TRACE_EVENT(sched_cpu_util,
 		__field(int,		isolated)
 		__field(int,		reserved)
 		__field(int,		high_irq_load)
+		__field(unsigned int,	nr_rtg_high_prio_tasks)
 	),
 
 	TP_fast_assign(
@@ -943,14 +944,16 @@ TRACE_EVENT(sched_cpu_util,
 		__entry->isolated           = cpu_isolated(cpu);
 		__entry->reserved           = is_reserved(cpu);
 		__entry->high_irq_load      = sched_cpu_high_irqload(cpu);
+		__entry->nr_rtg_high_prio_tasks = walt_nr_rtg_high_prio(cpu);
 	),
 
-	TP_printk("cpu=%d nr_running=%d cpu_util=%ld cpu_util_cum=%ld capacity_curr=%u capacity=%u capacity_orig=%u idle_state=%d irqload=%llu online=%u, isolated=%u, reserved=%u, high_irq_load=%u",
+	TP_printk("cpu=%d nr_running=%d cpu_util=%ld cpu_util_cum=%ld capacity_curr=%u capacity=%u capacity_orig=%u idle_state=%d irqload=%llu online=%u, isolated=%u, reserved=%u, high_irq_load=%u nr_rtg_hp=%u",
 		__entry->cpu, __entry->nr_running, __entry->cpu_util,
 		__entry->cpu_util_cum, __entry->capacity_curr,
 		__entry->capacity, __entry->capacity_orig,
 		__entry->idle_state, __entry->irqload, __entry->online,
-		__entry->isolated, __entry->reserved, __entry->high_irq_load)
+		__entry->isolated, __entry->reserved, __entry->high_irq_load,
+		__entry->nr_rtg_high_prio_tasks)
 );
 
 TRACE_EVENT(sched_compute_energy,

kernel/sched/fair.c

@@ -3937,6 +3937,12 @@ static inline void walt_adjust_cpus_for_packing(struct task_struct *p,
 	if (prefer_spread_on_idle(*best_idle_cpu))
 		fbt_env->need_idle |= 2;
 
+	if (task_rtg_high_prio(p) && walt_nr_rtg_high_prio(*target_cpu) > 0) {
+		*target_cpu = -1;
+		return;
+	}
+
+
 	if (fbt_env->need_idle || task_placement_boost_enabled(p) ||
 		fbt_env->boosted || shallowest_idle_cstate <= 0) {
 		*target_cpu = -1;
@@ -6549,6 +6555,8 @@ static void walt_find_best_target(struct sched_domain *sd, cpumask_t *cpus,
 	int unisolated_candidate = -1;
 	int order_index = fbt_env->order_index, end_index = fbt_env->end_index;
 	int cluster;
+	unsigned int target_nr_rtg_high_prio = UINT_MAX;
+	bool rtg_high_prio_task = task_rtg_high_prio(p);
 
 	/* Find start CPU based on boost value */
 	start_cpu = fbt_env->start_cpu;
@@ -6696,11 +6704,30 @@ static void walt_find_best_target(struct sched_domain *sd, cpumask_t *cpus,
 			if (p->state == TASK_RUNNING)
 				continue;
 
-			/* Favor CPUs with maximum spare capacity */
-			if (spare_cap < target_max_spare_cap)
-				continue;
+			/*
+			 * Try to spread the rtg high prio tasks so that they
+			 * don't preempt each other. This is a optimisitc
+			 * check assuming rtg high prio can actually preempt
+			 * the current running task with the given vruntime
+			 * boost.
+			 */
+			if (rtg_high_prio_task)  {
+				if (walt_nr_rtg_high_prio(i) > target_nr_rtg_high_prio)
+					continue;
+
+				/* Favor CPUs with maximum spare capacity */
+				if (walt_nr_rtg_high_prio(i) == target_nr_rtg_high_prio &&
+						spare_cap < target_max_spare_cap)
+					continue;
+
+			} else {
+				/* Favor CPUs with maximum spare capacity */
+				if (spare_cap < target_max_spare_cap)
+					continue;
+			}
 
 			target_max_spare_cap = spare_cap;
+			target_nr_rtg_high_prio = walt_nr_rtg_high_prio(i);
 			target_cpu = i;
 		}
 

kernel/sched/sched.h

@@ -105,6 +105,7 @@ struct walt_sched_stats {
 	int nr_big_tasks;
 	u64 cumulative_runnable_avg_scaled;
 	u64 pred_demands_sum_scaled;
+	unsigned int nr_rtg_high_prio_tasks;
 };
 
 struct walt_task_group {

kernel/sched/walt.h

@@ -78,6 +78,10 @@ static inline void inc_rq_walt_stats(struct rq *rq, struct task_struct *p)
 	if (p->wts.misfit)
 		rq->wrq.walt_stats.nr_big_tasks++;
 
+	p->wts.rtg_high_prio = task_rtg_high_prio(p);
+	if (p->wts.rtg_high_prio)
+		rq->wrq.walt_stats.nr_rtg_high_prio_tasks++;
+
 	walt_inc_cumulative_runnable_avg(rq, p);
 }
 
@@ -86,6 +90,9 @@ static inline void dec_rq_walt_stats(struct rq *rq, struct task_struct *p)
 	if (p->wts.misfit)
 		rq->wrq.walt_stats.nr_big_tasks--;
 
+	if (p->wts.rtg_high_prio)
+		rq->wrq.walt_stats.nr_rtg_high_prio_tasks--;
+
 	BUG_ON(rq->wrq.walt_stats.nr_big_tasks < 0);
 
 	walt_dec_cumulative_runnable_avg(rq, p);
@@ -192,6 +199,11 @@ static inline void walt_try_to_wake_up(struct task_struct *p)
 	rcu_read_unlock();
 }
 
+static inline unsigned int walt_nr_rtg_high_prio(int cpu)
+{
+	return cpu_rq(cpu)->wrq.walt_stats.nr_rtg_high_prio_tasks;
+}
+
 #else /* CONFIG_SCHED_WALT */
 
 static inline void walt_sched_init_rq(struct rq *rq) { }