sched: Add separate load tracking histogram to predict loads

#endif

#define RAVG_HIST_SIZE_MAX  5

#define NUM_BUSY_BUCKETS 10

/* ravg represents frequency scaled cpu-demand of tasks */

struct ravg {

	 *

	 * 'prev_window' represents task's contribution to cpu busy time

	 * statistics (rq->prev_runnable_sum) in previous window

	 *

	 * 'pred_demand' represents task's current predicted cpu busy time

	 *

	 * 'busy_buckets' groups historical busy time into different buckets

	 * used for prediction

	 */

	u64 mark_start;

	u32 sum, demand;

#ifdef CONFIG_SCHED_FREQ_INPUT

	u32 curr_window, prev_window;

	u16 active_windows;

	u32 pred_demand;

	u8 busy_buckets[NUM_BUSY_BUCKETS];

#endif

};

struct sched_load {

	unsigned long prev_load;

	unsigned long new_task_load;

	unsigned long predicted_load;

};

#if defined(CONFIG_SCHED_FREQ_INPUT)

extern unsigned int sysctl_sched_restrict_cluster_spill;

#if defined(CONFIG_SCHED_FREQ_INPUT)

extern unsigned int sysctl_sched_new_task_windows;

extern unsigned int sysctl_sched_pred_alert_freq;

#endif

#endif

		__field(unsigned int,	cpus_allowed		)

#ifdef CONFIG_SCHED_HMP

		__field(unsigned int,	demand			)

#ifdef CONFIG_SCHED_FREQ_INPUT

		__field(unsigned int,	pred_demand		)

#endif

#endif

	),

		__entry->cpus_allowed	= cpus_allowed;

#ifdef CONFIG_SCHED_HMP

		__entry->demand		= p->ravg.demand;

#ifdef CONFIG_SCHED_FREQ_INPUT

		__entry->pred_demand	= p->ravg.pred_demand;

#endif

#endif

	),

	TP_printk("cpu=%d %s comm=%s pid=%d prio=%d nr_running=%u cpu_load=%lu rt_nr_running=%u affine=%x"

#ifdef CONFIG_SCHED_HMP

		 " demand=%u"

#ifdef CONFIG_SCHED_FREQ_INPUT

		 " pred_demand=%u"

#endif

#endif

			, __entry->cpu,

			__entry->enqueue ? "enqueue" : "dequeue",

			__entry->cpu_load, __entry->rt_nr_running, __entry->cpus_allowed

#ifdef CONFIG_SCHED_HMP

			, __entry->demand

#ifdef CONFIG_SCHED_FREQ_INPUT

			, __entry->pred_demand

#endif

#endif

			)

);

		__field(unsigned int,	sum			)

		__field(	 int,	cpu			)

#ifdef CONFIG_SCHED_FREQ_INPUT

		__field(unsigned int,	pred_demand		)

		__field(	u64,	cs			)

		__field(	u64,	ps			)

		__field(	u32,	curr_window		)

		__entry->sum            = p->ravg.sum;

		__entry->irqtime        = irqtime;

#ifdef CONFIG_SCHED_FREQ_INPUT

		__entry->pred_demand     = p->ravg.pred_demand;

		__entry->cs             = rq->curr_runnable_sum;

		__entry->ps             = rq->prev_runnable_sum;

		__entry->curr_window	= p->ravg.curr_window;

	TP_printk("wc %llu ws %llu delta %llu event %s cpu %d cur_freq %u cur_pid %d task %d (%s) ms %llu delta %llu demand %u sum %u irqtime %llu"

#ifdef CONFIG_SCHED_FREQ_INPUT

		" cs %llu ps %llu cur_window %u prev_window %u nt_cs %llu nt_ps %llu active_wins %u"

		" pred_demand %u cs %llu ps %llu cur_window %u prev_window %u nt_cs %llu nt_ps %llu active_wins %u"

#endif

		, __entry->wallclock, __entry->win_start, __entry->delta,

		task_event_names[__entry->evt], __entry->cpu,

		__entry->delta_m, __entry->demand,

		__entry->sum, __entry->irqtime

#ifdef CONFIG_SCHED_FREQ_INPUT

		, __entry->cs, __entry->ps, __entry->curr_window,

		  __entry->prev_window,

		, __entry->pred_demand, __entry->cs, __entry->ps,

		__entry->curr_window, __entry->prev_window,

		  __entry->nt_cs, __entry->nt_ps,

		  __entry->active_windows

#endif

		__field(	 int,	samples			)

		__field(enum task_event,	evt		)

		__field(unsigned int,	demand			)

#ifdef CONFIG_SCHED_FREQ_INPUT

		__field(unsigned int,	pred_demand		)

#endif

		__array(	 u32,	hist, RAVG_HIST_SIZE_MAX)

		__field(unsigned int,	nr_big_tasks		)

		__field(	 int,	cpu			)

		__entry->samples        = samples;

		__entry->evt            = evt;

		__entry->demand         = p->ravg.demand;

#ifdef CONFIG_SCHED_FREQ_INPUT

		__entry->pred_demand     = p->ravg.pred_demand;

#endif

		memcpy(__entry->hist, p->ravg.sum_history,

					RAVG_HIST_SIZE_MAX * sizeof(u32));

		__entry->nr_big_tasks   = rq->hmp_stats.nr_big_tasks;

		__entry->cpu            = rq->cpu;

	),

	TP_printk("%d (%s): runtime %u samples %d event %s demand %u (hist: %u %u %u %u %u) cpu %d nr_big %u",

	TP_printk("%d (%s): runtime %u samples %d event %s demand %u"

#ifdef CONFIG_SCHED_FREQ_INPUT

		" pred_demand %u"

#endif

		" (hist: %u %u %u %u %u) cpu %d nr_big %u",

		__entry->pid, __entry->comm,

		__entry->runtime, __entry->samples,

		task_event_names[__entry->evt],

		__entry->demand, __entry->hist[0],

		__entry->hist[1], __entry->hist[2], __entry->hist[3],

		__entry->demand,

#ifdef CONFIG_SCHED_FREQ_INPUT

		__entry->pred_demand,

#endif

		__entry->hist[0], __entry->hist[1],

		__entry->hist[2], __entry->hist[3],

		__entry->hist[4], __entry->cpu, __entry->nr_big_tasks)

);

#ifdef CONFIG_SCHED_FREQ_INPUT

TRACE_EVENT(sched_update_pred_demand,

	TP_PROTO(struct rq *rq, struct task_struct *p, u32 runtime, int pct,

		 unsigned int pred_demand),

	TP_ARGS(rq, p, runtime, pct, pred_demand),

	TP_STRUCT__entry(

		__array(	char,	comm,   TASK_COMM_LEN	)

		__field(       pid_t,	pid			)

		__field(unsigned int,	runtime			)

		__field(	 int,	pct			)

		__field(unsigned int,	pred_demand		)

		__array(	  u8,	bucket, NUM_BUSY_BUCKETS)

		__field(	 int,	cpu			)

	),

	TP_fast_assign(

		memcpy(__entry->comm, p->comm, TASK_COMM_LEN);

		__entry->pid            = p->pid;

		__entry->runtime        = runtime;

		__entry->pct            = pct;

		__entry->pred_demand     = pred_demand;

		memcpy(__entry->bucket, p->ravg.busy_buckets,

					NUM_BUSY_BUCKETS * sizeof(u8));

		__entry->cpu            = rq->cpu;

	),

	TP_printk("%d (%s): runtime %u pct %d cpu %d pred_demand %u (buckets: %u %u %u %u %u %u %u %u %u %u)",

		__entry->pid, __entry->comm,

		__entry->runtime, __entry->pct, __entry->cpu,

		__entry->pred_demand, __entry->bucket[0], __entry->bucket[1],

		__entry->bucket[2], __entry->bucket[3],__entry->bucket[4],

		__entry->bucket[5], __entry->bucket[6], __entry->bucket[7],

		__entry->bucket[8], __entry->bucket[9])

);

TRACE_EVENT(sched_migration_update_sum,

	TP_PROTO(struct rq *rq, struct task_struct *p),

TRACE_EVENT(sched_get_busy,

	TP_PROTO(int cpu, u64 load, u64 nload, int early),

	TP_PROTO(int cpu, u64 load, u64 nload, u64 pload, int early),

	TP_ARGS(cpu, load, nload, early),

	TP_ARGS(cpu, load, nload, pload, early),

	TP_STRUCT__entry(

		__field(	int,	cpu			)

		__field(	u64,	load			)

		__field(	u64,	nload			)

		__field(	u64,	pload			)

		__field(	int,	early			)

	),

		__entry->cpu		= cpu;

		__entry->load		= load;

		__entry->nload		= nload;

		__entry->pload		= pload;

		__entry->early		= early;

	),

	TP_printk("cpu %d load %lld new_task_load %lld early %d",

		__entry->cpu, __entry->load, __entry->nload, __entry->early)

	TP_printk("cpu %d load %lld new_task_load %lld predicted_load %lld early %d",

		__entry->cpu, __entry->load, __entry->nload,

		__entry->pload, __entry->early)

);

TRACE_EVENT(sched_freq_alert,

	TP_PROTO(int cpu, u64 old_load, u64 new_load),

	TP_PROTO(int cpu, int pd_notif, u64 old_load, u64 new_load,

		u64 old_pred, u64 new_pred),

	TP_ARGS(cpu, old_load, new_load),

	TP_ARGS(cpu, pd_notif, old_load, new_load, old_pred, new_pred),

	TP_STRUCT__entry(

		__field(	int,	cpu			)

		__field(	int,	pd_notif		)

		__field(	u64,	old_load		)

		__field(	u64,	new_load		)

		__field(	u64,	old_pred		)

		__field(	u64,	new_pred		)

	),

	TP_fast_assign(

		__entry->cpu		= cpu;

		__entry->pd_notif	= pd_notif;

		__entry->old_load	= old_load;

		__entry->new_load	= new_load;

		__entry->old_pred	= old_pred;

		__entry->new_pred	= new_pred;

	),

	TP_printk("cpu %d old_load=%llu new_load=%llu",

		__entry->cpu, __entry->old_load, __entry->new_load)

	TP_printk("cpu %d pd_notif=%d old_load=%llu new_load=%llu "

		"old_pred=%llu new_pred=%llu",

		__entry->cpu, __entry->pd_notif, __entry->old_load,

		__entry->new_load, __entry->old_pred,

		 __entry->new_pred)

);

#endif	/* CONFIG_SCHED_FREQ_INPUT */

__read_mostly unsigned int sysctl_sched_window_stats_policy =

	WINDOW_STATS_MAX_RECENT_AVG;

__read_mostly unsigned int sysctl_sched_new_task_windows = 5;

static __read_mostly unsigned int sched_account_wait_time = 1;

__read_mostly unsigned int sysctl_sched_account_wait_time = 1;

#ifdef CONFIG_SCHED_FREQ_INPUT

__read_mostly unsigned int sysctl_sched_new_task_windows = 5;

static __read_mostly unsigned int sched_migration_fixup = 1;

__read_mostly unsigned int sysctl_sched_migration_fixup = 1;

__read_mostly int sysctl_sched_freq_dec_notify = 10 * 1024 * 1024; /* - 10GHz */

static __read_mostly unsigned int sched_io_is_busy;

__read_mostly unsigned int sysctl_sched_pred_alert_freq = 10 * 1024 * 1024;

#endif	/* CONFIG_SCHED_FREQ_INPUT */

/* 1 -> use PELT based load stats, 0 -> use window-based load stats */

/* Temporarily disable window-stats activity on all cpus */

unsigned int __read_mostly sched_disable_window_stats;

/*

 * Major task runtime. If a task runs for more than sched_major_task_runtime

 * in a window, it's considered to be generating majority of workload

 * for this window. Prediction could be adjusted for such tasks.

 */

#ifdef CONFIG_SCHED_FREQ_INPUT

__read_mostly unsigned int sched_major_task_runtime = 10000000;

#endif

static unsigned int sync_cpu;

#define EXITING_TASK_MARKER	0xdeaddead

}

/* Should scheduler alert governor for changing frequency? */

static int send_notification(struct rq *rq)

static int send_notification(struct rq *rq, int check_pred)

{

	unsigned int cur_freq, freq_required;

	unsigned long flags;

	if (!sched_enable_hmp)

		return 0;

	if (check_pred) {

		u64 prev = rq->old_busy_time;

		u64 predicted = rq->hmp_stats.pred_demands_sum;

		if (rq->cluster->cur_freq == rq->cluster->max_freq)

			return 0;

		prev = max(prev, rq->old_estimated_time);

		if (prev > predicted)

			return 0;

		cur_freq = load_to_freq(rq, prev);

		freq_required = load_to_freq(rq, predicted);

		if (freq_required < cur_freq + sysctl_sched_pred_alert_freq)

			return 0;

	} else {

		cur_freq = load_to_freq(rq, rq->old_busy_time);

		freq_required = load_to_freq(rq, rq->prev_runnable_sum);

		if (nearly_same_freq(cur_freq, freq_required))

			return 0;

	}

	raw_spin_lock_irqsave(&rq->lock, flags);

	if (!rq->notifier_sent) {

}

/* Alert governor if there is a need to change frequency */

void check_for_freq_change(struct rq *rq)

void check_for_freq_change(struct rq *rq, bool check_pred)

{

	int cpu = cpu_of(rq);

	if (!send_notification(rq))

	if (!send_notification(rq, check_pred))

		return;

	trace_sched_freq_alert(cpu, rq->old_busy_time, rq->prev_runnable_sum);

	trace_sched_freq_alert(cpu, check_pred, rq->old_busy_time,

			rq->prev_runnable_sum, rq->old_estimated_time,

			rq->hmp_stats.pred_demands_sum);

	atomic_notifier_call_chain(

		&load_alert_notifier_head, 0,

	return p->ravg.active_windows < sysctl_sched_new_task_windows;

}

#define INC_STEP 8

#define DEC_STEP 2

#define CONSISTENT_THRES 16

#define INC_STEP_BIG 16

/*

 * bucket_increase - update the count of all buckets

 *

 * @buckets: array of buckets tracking busy time of a task

 * @idx: the index of bucket to be incremented

 *

 * Each time a complete window finishes, count of bucket that runtime

 * falls in (@idx) is incremented. Counts of all other buckets are

 * decayed. The rate of increase and decay could be different based

 * on current count in the bucket.

 */

static inline void bucket_increase(u8 *buckets, int idx)

{

	int i, step;

	for (i = 0; i < NUM_BUSY_BUCKETS; i++) {

		if (idx != i) {

			if (buckets[i] > DEC_STEP)

				buckets[i] -= DEC_STEP;

			else

				buckets[i] = 0;

		} else {

			step = buckets[i] >= CONSISTENT_THRES ?

						INC_STEP_BIG : INC_STEP;

			if (buckets[i] > U8_MAX - step)

				buckets[i] = U8_MAX;

			else

				buckets[i] += step;

		}

	}

}

static inline int busy_to_bucket(u32 normalized_rt)

{

	int bidx;

	bidx = mult_frac(normalized_rt, NUM_BUSY_BUCKETS, max_task_load());

	bidx = min(bidx, NUM_BUSY_BUCKETS - 1);

	/*

	 * Combine lowest two buckets. The lowest frequency falls into

	 * 2nd bucket and thus keep predicting lowest bucket is not

	 * useful.

	 */

	if (!bidx)

		bidx++;

	return bidx;

}

static inline u64

scale_load_to_freq(u64 load, unsigned int src_freq, unsigned int dst_freq)

{

	return div64_u64(load * (u64)src_freq, (u64)dst_freq);

}

#define HEAVY_TASK_SKIP 2

#define HEAVY_TASK_SKIP_LIMIT 4

/*

 * get_pred_busy - calculate predicted demand for a task on runqueue

 *

 * @rq: runqueue of task p

 * @p: task whose prediction is being updated

 * @start: starting bucket. returned prediction should not be lower than

 *         this bucket.

 * @runtime: runtime of the task. returned prediction should not be lower

 *           than this runtime.

 * Note: @start can be derived from @runtime. It's passed in only to

 * avoid duplicated calculation in some cases.

 *

 * A new predicted busy time is returned for task @p based on @runtime

 * passed in. The function searches through buckets that represent busy

 * time equal to or bigger than @runtime and attempts to find the bucket to

 * to use for prediction. Once found, it searches through historical busy

 * time and returns the latest that falls into the bucket. If no such busy

 * time exists, it returns the medium of that bucket.

 */

static u32 get_pred_busy(struct rq *rq, struct task_struct *p,

				int start, u32 runtime)

{

	int i;

	u8 *buckets = p->ravg.busy_buckets;

	u32 *hist = p->ravg.sum_history;

	u32 dmin, dmax;

	u64 cur_freq_runtime = 0;

	int first = NUM_BUSY_BUCKETS, final, skip_to;

	u32 ret = runtime;

	/* skip prediction for new tasks due to lack of history */

	if (unlikely(is_new_task(p)))

		goto out;

	/* find minimal bucket index to pick */

	for (i = start; i < NUM_BUSY_BUCKETS; i++) {

		if (buckets[i]) {

			first = i;

			break;

		}

	}

	/* if no higher buckets are filled, predict runtime */

	if (first >= NUM_BUSY_BUCKETS)

		goto out;

	/* compute the bucket for prediction */

	final = first;

	if (first < HEAVY_TASK_SKIP_LIMIT) {

		/* compute runtime at current CPU frequency */

		cur_freq_runtime = mult_frac(runtime, max_possible_efficiency,

					     rq->cluster->efficiency);

		cur_freq_runtime = scale_load_to_freq(cur_freq_runtime,

				max_possible_freq, rq->cluster->cur_freq);

		/*

		 * if the task runs for majority of the window, try to

		 * pick higher buckets.

		 */

		if (cur_freq_runtime >= sched_major_task_runtime) {

			int next = NUM_BUSY_BUCKETS;

			/*

			 * if there is a higher bucket that's consistently

			 * hit, don't jump beyond that.

			 */

			for (i = start + 1; i <= HEAVY_TASK_SKIP_LIMIT &&

			     i < NUM_BUSY_BUCKETS; i++) {

				if (buckets[i] > CONSISTENT_THRES) {

					next = i;

					break;

				}

			}

			skip_to = min(next, start + HEAVY_TASK_SKIP);

			/* don't jump beyond HEAVY_TASK_SKIP_LIMIT */

			skip_to = min(HEAVY_TASK_SKIP_LIMIT, skip_to);

			/* don't go below first non-empty bucket, if any */

			final = max(first, skip_to);

		}

	}

	/* determine demand range for the predicted bucket */

	if (final < 2) {

		/* lowest two buckets are combined */

		dmin = 0;

		final = 1;

	} else {

		dmin = mult_frac(final, max_task_load(), NUM_BUSY_BUCKETS);

	}

	dmax = mult_frac(final + 1, max_task_load(), NUM_BUSY_BUCKETS);

	/*

	 * search through runtime history and return first runtime that falls

	 * into the range of predicted bucket.

	 */

	for (i = 0; i < sched_ravg_hist_size; i++) {

		if (hist[i] >= dmin && hist[i] < dmax) {

			ret = hist[i];

			break;

		}

	}

	/* no historical runtime within bucket found, use average of the bin */

	if (ret < dmin)

		ret = (dmin + dmax) / 2;

	/*

	 * when updating in middle of a window, runtime could be higher

	 * than all recorded history. Always predict at least runtime.

	 */

	ret = max(runtime, ret);

out:

	trace_sched_update_pred_demand(rq, p, runtime,

		mult_frac((unsigned int)cur_freq_runtime, 100,

			  sched_ravg_window), ret);

	return ret;

}

static inline u32 calc_pred_demand(struct rq *rq, struct task_struct *p)

{

	if (p->ravg.pred_demand >= p->ravg.curr_window)

		return p->ravg.pred_demand;

	return get_pred_busy(rq, p, busy_to_bucket(p->ravg.curr_window),

			     p->ravg.curr_window);

}

/*

 * predictive demand of a task is calculated at the window roll-over.

 * if the task current window busy time exceeds the predicted

 * demand, update it here to reflect the task needs.

 */

void update_task_pred_demand(struct rq *rq, struct task_struct *p, int event)

{

	u32 new, old;

	if (is_idle_task(p) || exiting_task(p))

		return;

	if (event != PUT_PREV_TASK && event != TASK_UPDATE &&

			(!sched_freq_account_wait_time ||

			 (event != TASK_MIGRATE &&

			 event != PICK_NEXT_TASK)))

		return;

	new = calc_pred_demand(rq, p);

	old = p->ravg.pred_demand;

	if (old >= new)

		return;

	if (task_on_rq_queued(p) && (!task_has_dl_policy(p) ||

				!p->dl.dl_throttled))

		p->sched_class->fixup_hmp_sched_stats(rq, p,

				p->ravg.demand,

				new);

	p->ravg.pred_demand = new;

}

/*

 * Account cpu activity in its busy time counters (rq->curr/prev_runnable_sum)

 */

	spin_unlock_irqrestore(&freq_max_load_lock, flags);

	return ret;

}

static inline u32 predict_and_update_buckets(struct rq *rq,

			struct task_struct *p, u32 runtime) {

	int bidx;

	u32 pred_demand;

	bidx = busy_to_bucket(runtime);

	pred_demand = get_pred_busy(rq, p, bidx, runtime);

	bucket_increase(p->ravg.busy_buckets, bidx);

	return pred_demand;

}

#define assign_ravg_pred_demand(x) (p->ravg.pred_demand = x)

#else	/* CONFIG_SCHED_FREQ_INPUT */

static inline void

update_task_pred_demand(struct rq *rq, struct task_struct *p, int event)

{

}

static inline void update_cpu_busy_time(struct task_struct *p, struct rq *rq,

	     int event, u64 wallclock, u64 irqtime)

{

}

static inline u32 predict_and_update_buckets(struct rq *rq,

			struct task_struct *p, u32 runtime)

{

	return 0;

}

#define assign_ravg_pred_demand(x)

#endif	/* CONFIG_SCHED_FREQ_INPUT */

static int account_busy_for_task_demand(struct task_struct *p, int event)

{

	u32 *hist = &p->ravg.sum_history[0];

	int ridx, widx;

	u32 max = 0, avg, demand;

	u32 max = 0, avg, demand, pred_demand;

	u64 sum = 0;

	/* Ignore windows where task had no activity */

		else

			demand = max(avg, runtime);

	}

	pred_demand = predict_and_update_buckets(rq, p, runtime);