[RFC, PATCH 3/9] CPU Controller V2 - group timeslice management

[Srivatsa Vaddagiri <vatsa@in.ibm.com>]

This patch does the actual job of dividing up CPU time amonmg various
task-groups as per their quota.

High-level overview:

	Lets say that we have three task-groups/classes A, B and C
	whose quota is 50%, 30% and 20%. The quota specifies how
	much of execution time each group gets per some quantum.
	If 10 seconds is chosen as this quantum, then group A's tasks
	get 5 seconds worth of execution time -every- 10 seconds, 
	grp B gets 3 seconds and so on, as shown below:


		A (5 sec)	      B (3 sec)     C (2 sec)
	|------------------------|---------------|-----------|
	
	T0<---------------- 10 sec quantum ----------------->T1


	In this scheme, grp C's tasks could starve potentially,
	waiting for their turn. 

	This patch avoids the above problem by breaking the 10 sec
	quantum into several smaller quantums. A, B and C get their
	due share in each smaller quantum and hence each group
	progress more quickly than in the above scheme.

	
	   A    B   C   A    B   C   	         A    B   C   A    B   C
	|-----|---|--|-----|---|--| .../ / ...|-----|---|--|-----|---|--|

	|<---------->|<---------->|			   |<---------->|
	    1 sec	  1 sec				       1 sec
	       ^	    ^
	       |	    |
	     (CPU_CONTROL_SHORT_WINDOW)

	|<------------------------------------------------------------->|
			10 sec quantum (CPU_CONTROL_LONG_WINDOW)


	The latter scheme is implemented by maintaining two counters (tokens)
	for each task-group (on each CPU).

		o ticks -> which is exhausted over the short quantum
		o long_ticks -> which is exhausted over the longer quantum

	To begin with:

	 grp->ticks = grp->quota * CPU_CONTROL_SHORT_WINDOW * HZ	
	 grp->long_ticks = CPU_CONTROL_LONG_WINDOW/CPU_CONTROL_SHORT_WINDOW


	grp->ticks is decremented every timer ticks. When it hits zero, 
	grp->long_ticks is decremented. If resulting grp->long_ticks is
	non-zero, grp->ticks wraps back to the initial value and the group
	is put in an expired bucket. Else if resulting grp->long_ticks is
	zero, the group is put in a greedy bucket.


Signed-off-by : Srivatsa Vaddagiri <vatsa@in.ibm.com>

---

 linux-2.6.18-root/kernel/sched.c |  124 ++++++++++++++++++++++++++++++++++++---
 1 files changed, 115 insertions(+), 9 deletions(-)

diff -puN kernel/sched.c~handle_grp_timeslice kernel/sched.c
--- linux-2.6.18/kernel/sched.c~handle_grp_timeslice	2006-09-28 16:39:22.407755712 +0530
+++ linux-2.6.18-root/kernel/sched.c	2006-09-28 17:23:20.141759192 +0530
@@ -157,9 +157,6 @@
 	(JIFFIES_TO_NS(MAX_SLEEP_AVG * \
 		(MAX_BONUS / 2 + DELTA((p)) + 1) / MAX_BONUS - 1))
 
-#define TASK_PREEMPTS_CURR(p, rq) \
-	((p)->prio < (rq)->curr->prio)
-
 /*
  * task_timeslice() scales user-nice values [ -20 ... 0 ... 19 ]
  * to time slice values: [800ms ... 100ms ... 5ms]
@@ -217,7 +214,9 @@ struct task_grp_rq {
 	struct prio_array arrays[2];
 	struct prio_array *active, *expired, *rq_array;
 	unsigned long expired_timestamp;
-	int prio;			/* Priority of the task-group */
+	unsigned short ticks, long_ticks;
+  	int prio;			/* Priority of the task-group */
+	unsigned long last_update;
 	struct list_head list;
 	struct task_grp *tg;
 };
@@ -226,6 +225,7 @@ static DEFINE_PER_CPU(struct task_grp_rq
 
 /* task-group object - maintains information about each task-group */
 struct task_grp {
+	unsigned short ticks, long_ticks; /* bandwidth given to task-group */
 	struct task_grp_rq *rq[NR_CPUS]; /* runqueue pointer for every cpu */
 };
 
@@ -267,7 +267,9 @@ struct rq {
 	/* these arrays hold task-groups.
 	 * xxx: Avoid this for !CONFIG_CPUMETER?
 	 */
-	struct prio_array *active, *expired, arrays[2];
+	struct prio_array *active, *expired, *greedy_active, *greedy_expired;
+	struct prio_array arrays[4];
+	unsigned long last_update;
 	atomic_t nr_iowait;
 
 #ifdef CONFIG_SMP
@@ -710,6 +712,14 @@ sched_info_switch(struct task_struct *pr
 #define sched_info_switch(t, next)	do { } while (0)
 #endif /* CONFIG_SCHEDSTATS || CONFIG_TASK_DELAY_ACCT */
 
+#define CPU_CONTROL_SHORT_WINDOW        (HZ)
+#define CPU_CONTROL_LONG_WINDOW         (5 * HZ)
+#define NUM_LONG_TICKS  (unsigned short) (CPU_CONTROL_LONG_WINDOW /	\
+						CPU_CONTROL_SHORT_WINDOW)
+
+#define greedy_grp(grp) (!grp->long_ticks)
+#define greedy_task(p)	(greedy_grp(task_grp_rq(p)))
+
 /* Dequeue task-group object from the main per-cpu runqueue */
 static void dequeue_task_grp(struct task_grp_rq *tgrq)
 {
@@ -753,12 +763,29 @@ static inline void update_task_grp_prio(
 		dequeue_task_grp(tgrq);
 	tgrq->prio = new_prio;
 	if (new_prio != MAX_PRIO) {
-		if (!array)
+		if (!greedy_grp(tgrq))
 			array = rq->active;
+		else
+			array = rq->greedy_active;
 		enqueue_task_grp(tgrq, array, head);
 	}
 }
 
+#ifdef CONFIG_CPUMETER
+
+#define TASK_PREEMPTS_CURR(p, rq) \
+	(idle_cpu(task_cpu(p)) || \
+	 (greedy_task((rq)->curr) && !greedy_task(p)) || \
+	 ((task_grp_rq(p)->rq_array == task_grp_rq((rq)->curr)->rq_array) && \
+                                          (((p)->prio < (rq)->curr->prio))))
+
+#else
+
+#define TASK_PREEMPTS_CURR(p, rq) \
+		        ((p)->prio < (rq)->curr->prio)
+
+#endif
+
 /*
  * Adding/removing a task to/from a priority array:
  */
@@ -3111,6 +3138,29 @@ void account_steal_time(struct task_stru
 		cpustat->steal = cputime64_add(cpustat->steal, tmp);
 }
 
+#ifdef CONFIG_CPUMETER
+static inline void move_task_grp_list(struct prio_array *src,
+					struct prio_array *dst)
+{
+	unsigned int i;
+
+	if (!src->nr_active)
+		return;
+
+	for (i = 0; i < MAX_PRIO; i++) {
+		struct list_head *list = src->queue + i;
+
+		while (!list_empty(list)) {
+			struct task_grp_rq *tgrq;
+
+			tgrq = list_entry(list->next, struct task_grp_rq, list);
+			dequeue_task_grp(tgrq);
+			enqueue_task_grp(tgrq, dst, 0);
+		}
+	}
+}
+#endif
+
 /*
  * This function gets called by the timer code, with HZ frequency.
  * We call it with interrupts disabled.
@@ -3209,6 +3259,39 @@ void scheduler_tick(void)
 	}
 
 out_unlock:
+#ifdef CONFIG_CPUMETER
+	if (!--tgrq->ticks) {
+                struct prio_array *target;
+
+		if (tgrq->long_ticks) {
+			--tgrq->long_ticks;
+			if (tgrq->long_ticks)
+				target = rq->expired;
+			else
+				target = rq->greedy_active;
+		} else /* should be in rq->greedy_active */
+			target = rq->greedy_expired;
+
+		/* Move the task group to expired list */
+		dequeue_task_grp(tgrq);
+		tgrq->ticks = task_grp(p)->ticks;
+		enqueue_task_grp(tgrq, target, 0);
+		set_tsk_need_resched(p);
+	}
+
+	if (unlikely(jiffies - rq->last_update > CPU_CONTROL_LONG_WINDOW)) {
+		if (rq->active->nr_active || rq->expired->nr_active) {
+			move_task_grp_list(rq->greedy_active, rq->active);
+			move_task_grp_list(rq->greedy_expired, rq->active);
+		} else {
+			switch_array(rq->active, rq->greedy_active);
+			switch_array(rq->expired, rq->greedy_expired);
+		}
+		rq->last_update = jiffies;
+		set_tsk_need_resched(p);
+	}
+#endif
+
 	spin_unlock(&rq->lock);
 out:
 	rebalance_tick(cpu, rq, NOT_IDLE);
@@ -3478,12 +3561,27 @@ need_resched_nonpreemptible:
 #ifdef CONFIG_CPUMETER
 	array = rq->active;
 	if (unlikely(!array->nr_active)) {
-		switch_array(rq->active, rq->expired);
-		array = rq->active;
+		if (rq->expired->nr_active) {
+			switch_array(rq->active, rq->expired);
+			array = rq->active;
+		} else {
+			array = rq->greedy_active;
+			if (!array->nr_active) {
+				switch_array(rq->greedy_active,
+						rq->greedy_expired);
+				array = rq->greedy_active;
+			}
+		}
 	}
 	idx = sched_find_first_bit(array->bitmap);
 	queue = array->queue + idx;
 	next_grp = list_entry(queue->next, struct task_grp_rq, list);
+
+	if (unlikely(next_grp->last_update != rq->last_update)) {
+		next_grp->ticks = next_grp->tg->ticks;
+		next_grp->long_ticks = next_grp->tg->long_ticks;
+		next_grp->last_update = rq->last_update;
+	}
 #else
 	next_grp = init_task_grp.rq[cpu];
 #endif
@@ -6909,6 +7007,8 @@ static void task_grp_rq_init(struct task
 	tgrq->active->best_static_prio = MAX_PRIO;
 	tgrq->active->best_dyn_prio = MAX_PRIO;
 	tgrq->prio = MAX_PRIO;
+	tgrq->ticks = tg->ticks;
+	tgrq->long_ticks = tg->long_ticks;
 	tgrq->tg = tg;
 	INIT_LIST_HEAD(&tgrq->list);
 
@@ -6929,6 +7029,9 @@ void __init sched_init(void)
 {
 	int i, j;
 
+	init_task_grp.ticks = CPU_CONTROL_SHORT_WINDOW;   /* 100% bandwidth */
+	init_task_grp.long_ticks = NUM_LONG_TICKS;
+
 	for_each_possible_cpu(i) {
 		struct rq *rq;
 		struct task_grp_rq *tgrq;
@@ -6941,6 +7044,9 @@ void __init sched_init(void)
 		rq->nr_running = 0;
 		rq->active = rq->arrays;
 		rq->expired = rq->arrays + 1;
+		rq->greedy_active = rq->arrays + 2;
+		rq->greedy_expired = rq->arrays + 3;
+		rq->last_update = tgrq->last_update = jiffies;
 
 #ifdef CONFIG_SMP
 		rq->sd = NULL;
@@ -6953,7 +7059,7 @@ void __init sched_init(void)
 #endif
 		atomic_set(&rq->nr_iowait, 0);
 
-		for (j = 0; j < 2; j++) {
+		for (j = 0; j < 4; j++) {
 			struct prio_array *array;
 			int k;
 
_
-- 
Regards,
vatsa
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/