path: root/litmus/sched_cedf.c

/*
 * litmus/sched_cedf.c
 *
 * Implementation of the C-EDF scheduling algorithm.
 *
 * This implementation is based on G-EDF:
 * - CPUs are clustered around L2 or L3 caches.
 * - Clusters topology is automatically detected (this is arch dependent
 *   and is working only on x86 at the moment --- and only with modern
 *   cpus that exports cpuid4 information)
 * - The plugins _does not_ attempt to put tasks in the right cluster i.e.
 *   the programmer needs to be aware of the topology to place tasks
 *   in the desired cluster
 * - default clustering is around L2 cache (cache index = 2)
 *   supported clusters are: L1 (private cache: pedf), L2, L3, ALL (all
 *   online_cpus are placed in a single cluster).
 *
 *   For details on functions, take a look at sched_gsn_edf.c
 *
 * Currently, we do not support changes in the number of online cpus.
 * If the num_online_cpus() dynamically changes, the plugin is broken.
 *
 * This version uses the simple approach and serializes all scheduling
 * decisions by the use of a queue lock. This is probably not the
 * best way to do it, but it should suffice for now.
 */

#include <linux/spinlock.h>
#include <linux/percpu.h>
#include <linux/sched.h>
#include <linux/slab.h>
#include <linux/uaccess.h>
#include <linux/module.h>

#include <litmus/litmus.h>
#include <litmus/jobs.h>
#include <litmus/preempt.h>
#include <litmus/sched_plugin.h>
#include <litmus/edf_common.h>
#include <litmus/sched_trace.h>

#include <litmus/clustered.h>

#include <litmus/bheap.h>
#include <litmus/binheap.h>

#ifdef CONFIG_LITMUS_LOCKING
#include <litmus/kfmlp_lock.h>
#endif

#ifdef CONFIG_LITMUS_NESTED_LOCKING
#include <litmus/rsm_lock.h>
#include <litmus/ikglp_lock.h>
#endif

#ifdef CONFIG_SCHED_CPU_AFFINITY
#include <litmus/affinity.h>
#endif

/* to configure the cluster size */
#include <litmus/litmus_proc.h>

#ifdef CONFIG_SCHED_CPU_AFFINITY
#include <litmus/affinity.h>
#endif

#ifdef CONFIG_LITMUS_SOFTIRQD
#include <litmus/litmus_softirq.h>
#endif

#ifdef CONFIG_LITMUS_PAI_SOFTIRQD
#include <linux/interrupt.h>
#include <litmus/trace.h>
#endif

#ifdef CONFIG_LITMUS_NVIDIA
#include <litmus/nvidia_info.h>
#endif

#if defined(CONFIG_LITMUS_AFFINITY_LOCKING) && defined(CONFIG_LITMUS_NVIDIA)
#include <litmus/gpu_affinity.h>
#endif

/* Reference configuration variable. Determines which cache level is used to
 * group CPUs into clusters.  GLOBAL_CLUSTER, which is the default, means that
 * all CPUs form a single cluster (just like GSN-EDF).
 */
static enum cache_level cluster_config = GLOBAL_CLUSTER;

struct clusterdomain;

/* cpu_entry_t - maintain the linked and scheduled state
 *
 * A cpu also contains a pointer to the cedf_domain_t cluster
 * that owns it (struct clusterdomain*)
 */
typedef struct  {
	int 			cpu;
	struct clusterdomain*	cluster;	/* owning cluster */
	struct task_struct*	linked;		/* only RT tasks */
	struct task_struct*	scheduled;	/* only RT tasks */
	atomic_t		will_schedule;	/* prevent unneeded IPIs */
	struct binheap_node hn;
} cpu_entry_t;

/* one cpu_entry_t per CPU */
DEFINE_PER_CPU(cpu_entry_t, cedf_cpu_entries);

#define set_will_schedule() \
	(atomic_set(&__get_cpu_var(cedf_cpu_entries).will_schedule, 1))
#define clear_will_schedule() \
	(atomic_set(&__get_cpu_var(cedf_cpu_entries).will_schedule, 0))
#define test_will_schedule(cpu) \
	(atomic_read(&per_cpu(cedf_cpu_entries, cpu).will_schedule))

#ifdef CONFIG_LITMUS_PAI_SOFTIRQD
struct tasklet_head
{
	struct tasklet_struct *head;
	struct tasklet_struct **tail;
};
#endif

/*
 * In C-EDF there is a cedf domain _per_ cluster
 * The number of clusters is dynamically determined accordingly to the
 * total cpu number and the cluster size
 */
typedef struct clusterdomain {
	/* rt_domain for this cluster */
	rt_domain_t	domain;
	/* cpus in this cluster */
	cpu_entry_t*	*cpus;
	/* map of this cluster cpus */
	cpumask_var_t	cpu_map;
	/* the cpus queue themselves according to priority in here */
	struct binheap_handle cpu_heap;
	/* lock for this cluster */
#define cluster_lock domain.ready_lock

#ifdef CONFIG_LITMUS_PAI_SOFTIRQD
	struct tasklet_head pending_tasklets;
#endif

#ifdef CONFIG_LITMUS_DGL_SUPPORT
	raw_spinlock_t dgl_lock;
#endif
} cedf_domain_t;

/* a cedf_domain per cluster; allocation is done at init/activation time */
cedf_domain_t *cedf;

#define remote_cluster(cpu)	((cedf_domain_t *) per_cpu(cedf_cpu_entries, cpu).cluster)
#define task_cpu_cluster(task)	remote_cluster(get_partition(task))

/* total number of cluster */
static int num_clusters;
/* we do not support cluster of different sizes */
static unsigned int cluster_size;

static int clusters_allocated = 0;

#ifdef CONFIG_LITMUS_DGL_SUPPORT
static raw_spinlock_t* cedf_get_dgl_spinlock(struct task_struct *t)
{
	cedf_domain_t *cluster = task_cpu_cluster(t);
	return(&cluster->dgl_lock);
}
#endif


/* Uncomment WANT_ALL_SCHED_EVENTS if you want to see all scheduling
 * decisions in the TRACE() log; uncomment VERBOSE_INIT for verbose
 * information during the initialization of the plugin (e.g., topology)
#define WANT_ALL_SCHED_EVENTS
 */
#define VERBOSE_INIT

static int cpu_lower_prio(struct binheap_node *_a, struct binheap_node *_b)
{
	cpu_entry_t *a = binheap_entry(_a, cpu_entry_t, hn);
	cpu_entry_t *b = binheap_entry(_b, cpu_entry_t, hn);

	/* Note that a and b are inverted: we want the lowest-priority CPU at
	 * the top of the heap.
	 */
	return edf_higher_prio(b->linked, a->linked);
}

/* update_cpu_position - Move the cpu entry to the correct place to maintain
 *                       order in the cpu queue. Caller must hold cedf lock.
 */
static void update_cpu_position(cpu_entry_t *entry)
{
	cedf_domain_t *cluster = entry->cluster;

	if (likely(binheap_is_in_heap(&entry->hn))) {
		binheap_delete(&entry->hn, &cluster->cpu_heap);
	}

	binheap_add(&entry->hn, &cluster->cpu_heap, cpu_entry_t, hn);
}

/* caller must hold cedf lock */
static cpu_entry_t* lowest_prio_cpu(cedf_domain_t *cluster)
{
	return binheap_top_entry(&cluster->cpu_heap, cpu_entry_t, hn);
}


/* link_task_to_cpu - Update the link of a CPU.
 *                    Handles the case where the to-be-linked task is already
 *                    scheduled on a different CPU.
 */
static noinline void link_task_to_cpu(struct task_struct* linked,
				      cpu_entry_t *entry)
{
	cpu_entry_t *sched;
	struct task_struct* tmp;
	int on_cpu;

	BUG_ON(linked && !is_realtime(linked));

	/* Currently linked task is set to be unlinked. */
	if (entry->linked) {
		entry->linked->rt_param.linked_on = NO_CPU;
	}

	/* Link new task to CPU. */
	if (linked) {
		set_rt_flags(linked, RT_F_RUNNING);
		/* handle task is already scheduled somewhere! */
		on_cpu = linked->rt_param.scheduled_on;
		if (on_cpu != NO_CPU) {
			sched = &per_cpu(cedf_cpu_entries, on_cpu);
			/* this should only happen if not linked already */
			BUG_ON(sched->linked == linked);

			/* If we are already scheduled on the CPU to which we
			 * wanted to link, we don't need to do the swap --
			 * we just link ourselves to the CPU and depend on
			 * the caller to get things right.
			 */
			if (entry != sched) {
				TRACE_TASK(linked,
					   "already scheduled on %d, updating link.\n",
					   sched->cpu);
				tmp = sched->linked;
				linked->rt_param.linked_on = sched->cpu;
				sched->linked = linked;
				update_cpu_position(sched);
				linked = tmp;
			}
		}
		if (linked) /* might be NULL due to swap */
			linked->rt_param.linked_on = entry->cpu;
	}
	entry->linked = linked;
#ifdef WANT_ALL_SCHED_EVENTS
	if (linked)
		TRACE_TASK(linked, "linked to %d.\n", entry->cpu);
	else
		TRACE("NULL linked to %d.\n", entry->cpu);
#endif
	update_cpu_position(entry);
}

/* unlink - Make sure a task is not linked any longer to an entry
 *          where it was linked before. Must hold cluster_lock.
 */
static noinline void unlink(struct task_struct* t)
{
    	cpu_entry_t *entry;

	if (t->rt_param.linked_on != NO_CPU) {
		/* unlink */
		entry = &per_cpu(cedf_cpu_entries, t->rt_param.linked_on);
		t->rt_param.linked_on = NO_CPU;
		link_task_to_cpu(NULL, entry);
	} else if (is_queued(t)) {
		/* This is an interesting situation: t is scheduled,
		 * but was just recently unlinked.  It cannot be
		 * linked anywhere else (because then it would have
		 * been relinked to this CPU), thus it must be in some
		 * queue. We must remove it from the list in this
		 * case.
		 *
		 * in C-EDF case is should be somewhere in the queue for
		 * its domain, therefore and we can get the domain using
		 * task_cpu_cluster
		 */
		remove(&(task_cpu_cluster(t))->domain, t);
	}
}


/* preempt - force a CPU to reschedule
 */
static void preempt(cpu_entry_t *entry)
{
	preempt_if_preemptable(entry->scheduled, entry->cpu);
}

/* requeue - Put an unlinked task into gsn-edf domain.
 *           Caller must hold cluster_lock.
 */
static noinline void requeue(struct task_struct* task)
{
	cedf_domain_t *cluster = task_cpu_cluster(task);
	BUG_ON(!task);
	/* sanity check before insertion */
	BUG_ON(is_queued(task));

	if (is_released(task, litmus_clock()))
		__add_ready(&cluster->domain, task);
	else {
		/* it has got to wait */
		add_release(&cluster->domain, task);
	}
}

#ifdef CONFIG_SCHED_CPU_AFFINITY
static cpu_entry_t* cedf_get_nearest_available_cpu(
				cedf_domain_t *cluster, cpu_entry_t *start)
{
	cpu_entry_t *affinity;

	get_nearest_available_cpu(affinity, start, cedf_cpu_entries,
#ifdef CONFIG_RELEASE_MASTER
		cluster->domain.release_master
#else
		NO_CPU
#endif
		);

	/* make sure CPU is in our cluster */
	if (affinity && cpu_isset(affinity->cpu, *cluster->cpu_map))
		return(affinity);
	else
		return(NULL);
}
#endif


/* check for any necessary preemptions */
static void check_for_preemptions(cedf_domain_t *cluster)
{
	struct task_struct *task;
	cpu_entry_t *last;

	for(last = lowest_prio_cpu(cluster);
	    edf_preemption_needed(&cluster->domain, last->linked);
	    last = lowest_prio_cpu(cluster)) {
		/* preemption necessary */
		task = __take_ready(&cluster->domain);
		TRACE("check_for_preemptions: attempting to link task %d to %d\n",
		      task->pid, last->cpu);
#ifdef CONFIG_SCHED_CPU_AFFINITY
		{
			cpu_entry_t *affinity =
					cedf_get_nearest_available_cpu(cluster,
						&per_cpu(cedf_cpu_entries, task_cpu(task)));
			if(affinity)
				last = affinity;
			else if(last->linked)
				requeue(last->linked);
		}
#else
		if (last->linked)
			requeue(last->linked);
#endif
		link_task_to_cpu(task, last);
		preempt(last);
	}
}

/* cedf_job_arrival: task is either resumed or released */
static noinline void cedf_job_arrival(struct task_struct* task)
{
	cedf_domain_t *cluster = task_cpu_cluster(task);
	BUG_ON(!task);

	requeue(task);
	check_for_preemptions(cluster);
}

static void cedf_release_jobs(rt_domain_t* rt, struct bheap* tasks)
{
	cedf_domain_t* cluster = container_of(rt, cedf_domain_t, domain);
	unsigned long flags;

	raw_spin_lock_irqsave(&cluster->cluster_lock, flags);

	__merge_ready(&cluster->domain, tasks);
	check_for_preemptions(cluster);

	raw_spin_unlock_irqrestore(&cluster->cluster_lock, flags);
}

/* caller holds cluster_lock */
static noinline void job_completion(struct task_struct *t, int forced)
{
	BUG_ON(!t);

	sched_trace_task_completion(t, forced);

#ifdef CONFIG_LITMUS_NVIDIA
	atomic_set(&tsk_rt(t)->nv_int_count, 0);
#endif

	TRACE_TASK(t, "job_completion().\n");

	/* set flags */
	set_rt_flags(t, RT_F_SLEEP);
	/* prepare for next period */
	prepare_for_next_period(t);
	if (is_released(t, litmus_clock()))
		sched_trace_task_release(t);
	/* unlink */
	unlink(t);
	/* requeue
	 * But don't requeue a blocking task. */
	if (is_running(t))
		cedf_job_arrival(t);
}

/* cedf_tick - this function is called for every local timer
 *                         interrupt.
 *
 *                   checks whether the current task has expired and checks
 *                   whether we need to preempt it if it has not expired
 */
static void cedf_tick(struct task_struct* t)
{
	if (is_realtime(t) && budget_enforced(t) && budget_exhausted(t)) {
		if (!is_np(t)) {
			/* np tasks will be preempted when they become
			 * preemptable again
			 */
			litmus_reschedule_local();
			set_will_schedule();
			TRACE("cedf_scheduler_tick: "
			      "%d is preemptable "
			      " => FORCE_RESCHED\n", t->pid);
		} else if (is_user_np(t)) {
			TRACE("cedf_scheduler_tick: "
			      "%d is non-preemptable, "
			      "preemption delayed.\n", t->pid);
			request_exit_np(t);
		}
	}
}













#ifdef CONFIG_LITMUS_PAI_SOFTIRQD


static void __do_lit_tasklet(struct tasklet_struct* tasklet, unsigned long flushed)
{
	if (!atomic_read(&tasklet->count)) {
		if(tasklet->owner) {
			sched_trace_tasklet_begin(tasklet->owner);
		}

		if (!test_and_clear_bit(TASKLET_STATE_SCHED, &tasklet->state))
		{
			BUG();
		}
		TRACE("%s: Invoking tasklet with owner pid = %d (flushed = %d).\n",
			  __FUNCTION__,
			  (tasklet->owner) ? tasklet->owner->pid : -1,
			  (tasklet->owner) ? 0 : 1);
		tasklet->func(tasklet->data);
		tasklet_unlock(tasklet);

		if(tasklet->owner) {
			sched_trace_tasklet_end(tasklet->owner, flushed);
		}
	}
	else {
		BUG();
	}
}


static void do_lit_tasklets(cedf_domain_t* cluster, struct task_struct* sched_task)
{
	int work_to_do = 1;
	struct tasklet_struct *tasklet = NULL;
	unsigned long flags;

	while(work_to_do) {

		TS_NV_SCHED_BOTISR_START;

		raw_spin_lock_irqsave(&cluster->cluster_lock, flags);

		if(cluster->pending_tasklets.head != NULL) {
			// remove tasklet at head.
			struct tasklet_struct *prev = NULL;
			tasklet = cluster->pending_tasklets.head;

			// find a tasklet with prio to execute; skip ones where
			// sched_task has a higher priority.
			// We use the '!edf' test instead of swaping function arguments since
			// both sched_task and owner could be NULL.  In this case, we want to
			// still execute the tasklet.
			while(tasklet && !edf_higher_prio(tasklet->owner, sched_task)) {
				prev = tasklet;
				tasklet = tasklet->next;
			}

			if(tasklet) {  // found something to execuite
				// remove the tasklet from the queue
				if(prev) {
					prev->next = tasklet->next;
					if(prev->next == NULL) {
						TRACE("%s: Tasklet for %d is the last element in tasklet queue.\n", __FUNCTION__, tasklet->owner->pid);
						cluster->pending_tasklets.tail = &(prev);
					}
				}
				else {
					cluster->pending_tasklets.head = tasklet->next;
					if(tasklet->next == NULL) {
						TRACE("%s: Tasklet for %d is the last element in tasklet queue.\n", __FUNCTION__, tasklet->owner->pid);
						cluster->pending_tasklets.tail = &(cluster->pending_tasklets.head);
					}
				}
			}
			else {
				TRACE("%s: No tasklets with eligible priority.\n", __FUNCTION__);
			}
		}
		else {
			TRACE("%s: Tasklet queue is empty.\n", __FUNCTION__);
		}

		raw_spin_unlock_irqrestore(&cluster->cluster_lock, flags);

		if(tasklet) {
			__do_lit_tasklet(tasklet, 0ul);
			tasklet = NULL;
		}
		else {
			work_to_do = 0;
		}

		TS_NV_SCHED_BOTISR_END;
	}
}

static void __add_pai_tasklet(struct tasklet_struct* tasklet, cedf_domain_t* cluster)
{
	struct tasklet_struct* step;

	tasklet->next = NULL;  // make sure there are no old values floating around

	step = cluster->pending_tasklets.head;
	if(step == NULL) {
		TRACE("%s: tasklet queue empty.  inserting tasklet for %d at head.\n", __FUNCTION__, tasklet->owner->pid);
		// insert at tail.
		*(cluster->pending_tasklets.tail) = tasklet;
		cluster->pending_tasklets.tail = &(tasklet->next);
	}
	else if((*(cluster->pending_tasklets.tail) != NULL) &&
			edf_higher_prio((*(cluster->pending_tasklets.tail))->owner, tasklet->owner)) {
		// insert at tail.
		TRACE("%s: tasklet belongs at end.  inserting tasklet for %d at tail.\n", __FUNCTION__, tasklet->owner->pid);

		*(cluster->pending_tasklets.tail) = tasklet;
		cluster->pending_tasklets.tail = &(tasklet->next);
	}
	else {

		// insert the tasklet somewhere in the middle.

        TRACE("%s: tasklet belongs somewhere in the middle.\n", __FUNCTION__);

		while(step->next && edf_higher_prio(step->next->owner, tasklet->owner)) {
			step = step->next;
		}

		// insert tasklet right before step->next.

		TRACE("%s: inserting tasklet for %d between %d and %d.\n", __FUNCTION__,
			  tasklet->owner->pid,
			  (step->owner) ?
			  step->owner->pid :
			  -1,
			  (step->next) ?
			  ((step->next->owner) ?
			   step->next->owner->pid :
			   -1) :
			  -1);

		tasklet->next = step->next;
		step->next = tasklet;

		// patch up the head if needed.
		if(cluster->pending_tasklets.head == step)
		{
			TRACE("%s: %d is the new tasklet queue head.\n", __FUNCTION__, tasklet->owner->pid);
			cluster->pending_tasklets.head = tasklet;
		}
	}
}

static void cedf_run_tasklets(struct task_struct* sched_task)
{
	cedf_domain_t* cluster;

	preempt_disable();

	cluster = (is_realtime(sched_task)) ?
		task_cpu_cluster(sched_task) :
		remote_cluster(smp_processor_id());

	if(cluster && cluster->pending_tasklets.head != NULL) {
		TRACE("%s: There are tasklets to process.\n", __FUNCTION__);
		do_lit_tasklets(cluster, sched_task);
	}

	preempt_enable_no_resched();
}



static int cedf_enqueue_pai_tasklet(struct tasklet_struct* tasklet)
{
#if 0
	cedf_domain_t *cluster = NULL;
	cpu_entry_t *targetCPU = NULL;
	int thisCPU;
	int runLocal = 0;
	int runNow = 0;
	unsigned long flags;

    if(unlikely((tasklet->owner == NULL) || !is_realtime(tasklet->owner)))
    {
        TRACE("%s: No owner associated with this tasklet!\n", __FUNCTION__);
		return 0;
    }

	cluster = task_cpu_cluster(tasklet->owner);

	raw_spin_lock_irqsave(&cluster->cluster_lock, flags);

	thisCPU = smp_processor_id();

#ifdef CONFIG_SCHED_CPU_AFFINITY
	{
		cpu_entry_t* affinity = NULL;

		// use this CPU if it is in our cluster and isn't running any RT work.
		if(cpu_isset(thisCPU, *cluster->cpu_map) && (__get_cpu_var(cedf_cpu_entries).linked == NULL)) {
			affinity = &(__get_cpu_var(cedf_cpu_entries));
		}
		else {
			// this CPU is busy or shouldn't run tasklet in this cluster.
			// look for available near by CPUs.
			// NOTE: Affinity towards owner and not this CPU.  Is this right?
			affinity =
				cedf_get_nearest_available_cpu(cluster,
								&per_cpu(cedf_cpu_entries, task_cpu(tasklet->owner)));
		}

		targetCPU = affinity;
	}
#endif

	if (targetCPU == NULL) {
		targetCPU = lowest_prio_cpu(cluster);
	}

	if (edf_higher_prio(tasklet->owner, targetCPU->linked)) {
		if (thisCPU == targetCPU->cpu) {
			TRACE("%s: Run tasklet locally (and now).\n", __FUNCTION__);
			runLocal = 1;
			runNow = 1;
		}
		else {
			TRACE("%s: Run tasklet remotely (and now).\n", __FUNCTION__);
			runLocal = 0;
			runNow = 1;
		}
	}
	else {
		runLocal = 0;
		runNow = 0;
	}

	if(!runLocal) {
		// enqueue the tasklet
		__add_pai_tasklet(tasklet, cluster);
	}

	raw_spin_unlock_irqrestore(&cluster->cluster_lock, flags);


	if (runLocal /*&& runNow */) {  // runNow == 1 is implied
		TRACE("%s: Running tasklet on CPU where it was received.\n", __FUNCTION__);
		__do_lit_tasklet(tasklet, 0ul);
	}
	else if (runNow /*&& !runLocal */) {  // runLocal == 0 is implied
		TRACE("%s: Triggering CPU %d to run tasklet.\n", __FUNCTION__, targetCPU->cpu);
		preempt(targetCPU);  // need to be protected by cluster_lock?
	}
	else {
		TRACE("%s: Scheduling of tasklet was deferred.\n", __FUNCTION__);
	}
#else
	TRACE("%s: Running tasklet on CPU where it was received.\n", __FUNCTION__);
	__do_lit_tasklet(tasklet, 0ul);
#endif
	return(1); // success
}

static void cedf_change_prio_pai_tasklet(struct task_struct *old_prio,
										 struct task_struct *new_prio)
{
	struct tasklet_struct* step;
	unsigned long flags;
	cedf_domain_t *cluster;
	struct task_struct *probe;

	// identify the cluster by the assignment of these tasks.  one should
	// be non-NULL.
	probe = (old_prio) ? old_prio : new_prio;

	if(probe) {
		cluster = task_cpu_cluster(probe);

		if(cluster->pending_tasklets.head != NULL) {
			raw_spin_lock_irqsave(&cluster->cluster_lock, flags);
			for(step = cluster->pending_tasklets.head; step != NULL; step = step->next) {
				if(step->owner == old_prio) {
					TRACE("%s: Found tasklet to change: %d\n", __FUNCTION__, step->owner->pid);
					step->owner = new_prio;
				}
			}
			raw_spin_unlock_irqrestore(&cluster->cluster_lock, flags);
		}
	}
	else {
		TRACE("%s: Both priorities were NULL\n");
	}
}

#endif  // PAI

/* Getting schedule() right is a bit tricky. schedule() may not make any
 * assumptions on the state of the current task since it may be called for a
 * number of reasons. The reasons include a scheduler_tick() determined that it
 * was necessary, because sys_exit_np() was called, because some Linux
 * subsystem determined so, or even (in the worst case) because there is a bug
 * hidden somewhere. Thus, we must take extreme care to determine what the
 * current state is.
 *
 * The CPU could currently be scheduling a task (or not), be linked (or not).
 *
 * The following assertions for the scheduled task could hold:
 *
 *      - !is_running(scheduled)        // the job blocks
 *	- scheduled->timeslice == 0	// the job completed (forcefully)
 *	- get_rt_flag() == RT_F_SLEEP	// the job completed (by syscall)
 * 	- linked != scheduled		// we need to reschedule (for any reason)
 * 	- is_np(scheduled)		// rescheduling must be delayed,
 *					   sys_exit_np must be requested
 *
 * Any of these can occur together.
 */
static struct task_struct* cedf_schedule(struct task_struct * prev)
{
	cpu_entry_t* entry = &__get_cpu_var(cedf_cpu_entries);
	cedf_domain_t *cluster = entry->cluster;
	int out_of_time, sleep, preempt, np, exists, blocks;
	struct task_struct* next = NULL;

#ifdef CONFIG_RELEASE_MASTER
	/* Bail out early if we are the release master.
	 * The release master never schedules any real-time tasks.
	 */
	if (unlikely(cluster->domain.release_master == entry->cpu)) {
		sched_state_task_picked();
		return NULL;
	}
#endif

	raw_spin_lock(&cluster->cluster_lock);
	clear_will_schedule();

	/* sanity checking */
	BUG_ON(entry->scheduled && entry->scheduled != prev);
	BUG_ON(entry->scheduled && !is_realtime(prev));
	BUG_ON(is_realtime(prev) && !entry->scheduled);

	/* (0) Determine state */
	exists      = entry->scheduled != NULL;
	blocks      = exists && !is_running(entry->scheduled);
	out_of_time = exists &&
				  budget_enforced(entry->scheduled) &&
				  budget_exhausted(entry->scheduled);
	np 	    = exists && is_np(entry->scheduled);
	sleep	    = exists && get_rt_flags(entry->scheduled) == RT_F_SLEEP;
	preempt     = entry->scheduled != entry->linked;

#ifdef WANT_ALL_SCHED_EVENTS
	TRACE_TASK(prev, "invoked cedf_schedule.\n");
#endif

	if (exists)
		TRACE_TASK(prev,
			   "blocks:%d out_of_time:%d np:%d sleep:%d preempt:%d "
			   "state:%d sig:%d\n",
			   blocks, out_of_time, np, sleep, preempt,
			   prev->state, signal_pending(prev));
	if (entry->linked && preempt)
		TRACE_TASK(prev, "will be preempted by %s/%d\n",
			   entry->linked->comm, entry->linked->pid);


	/* If a task blocks we have no choice but to reschedule.
	 */
	if (blocks)
		unlink(entry->scheduled);

#if defined(CONFIG_LITMUS_NVIDIA) && defined(CONFIG_LITMUS_AFFINITY_LOCKING)
	if(exists && is_realtime(entry->scheduled) && tsk_rt(entry->scheduled)->held_gpus) {
		if(!blocks || tsk_rt(entry->scheduled)->suspend_gpu_tracker_on_block) {
			// don't track preemptions or locking protocol suspensions.
			TRACE_TASK(entry->scheduled, "stopping GPU tracker.\n");
			stop_gpu_tracker(entry->scheduled);
		}
		else if(blocks && !tsk_rt(entry->scheduled)->suspend_gpu_tracker_on_block) {
			TRACE_TASK(entry->scheduled, "GPU tracker remains on during suspension.\n");
		}
	}
#endif

	/* Request a sys_exit_np() call if we would like to preempt but cannot.
	 * We need to make sure to update the link structure anyway in case
	 * that we are still linked. Multiple calls to request_exit_np() don't
	 * hurt.
	 */
	if (np && (out_of_time || preempt || sleep)) {
		unlink(entry->scheduled);
		request_exit_np(entry->scheduled);
	}

	/* Any task that is preemptable and either exhausts its execution
	 * budget or wants to sleep completes. We may have to reschedule after
	 * this. Don't do a job completion if we block (can't have timers running
	 * for blocked jobs). Preemption go first for the same reason.
	 */
	if (!np && (out_of_time || sleep) && !blocks && !preempt)
		job_completion(entry->scheduled, !sleep);

	/* Link pending task if we became unlinked.
	 */
	if (!entry->linked)
		link_task_to_cpu(__take_ready(&cluster->domain), entry);

	/* The final scheduling decision. Do we need to switch for some reason?
	 * If linked is different from scheduled, then select linked as next.
	 */
	if ((!np || blocks) &&
	    entry->linked != entry->scheduled) {
		/* Schedule a linked job? */
		if (entry->linked) {
			entry->linked->rt_param.scheduled_on = entry->cpu;
			next = entry->linked;
		}
		if (entry->scheduled) {
			/* not gonna be scheduled soon */
			entry->scheduled->rt_param.scheduled_on = NO_CPU;
			TRACE_TASK(entry->scheduled, "scheduled_on = NO_CPU\n");
		}
	} else
		/* Only override Linux scheduler if we have a real-time task
		 * scheduled that needs to continue.
		 */
		if (exists)
			next = prev;

	sched_state_task_picked();
	raw_spin_unlock(&cluster->cluster_lock);

#ifdef WANT_ALL_SCHED_EVENTS
	TRACE("cluster_lock released, next=0x%p\n", next);

	if (next)
		TRACE_TASK(next, "scheduled at %llu\n", litmus_clock());
	else if (exists && !next)
		TRACE("becomes idle at %llu.\n", litmus_clock());
#endif

	return next;
}


/* _finish_switch - we just finished the switch away from prev
 */
static void cedf_finish_switch(struct task_struct *prev)
{
	cpu_entry_t* 	entry = &__get_cpu_var(cedf_cpu_entries);

	entry->scheduled = is_realtime(current) ? current : NULL;
#ifdef WANT_ALL_SCHED_EVENTS
	TRACE_TASK(prev, "switched away from\n");
#endif
}


/*	Prepare a task for running in RT mode
 */
static void cedf_task_new(struct task_struct * t, int on_rq, int running)
{
	unsigned long 		flags;
	cpu_entry_t* 		entry;
	cedf_domain_t*		cluster;

	TRACE("c-edf: task new %d\n", t->pid);

	/* the cluster doesn't change even if t is running */
	cluster = task_cpu_cluster(t);

	raw_spin_lock_irqsave(&cluster->cluster_lock, flags);

	/* setup job params */
	release_at(t, litmus_clock());

	if (running) {
		entry = &per_cpu(cedf_cpu_entries, task_cpu(t));
		BUG_ON(entry->scheduled);

#ifdef CONFIG_RELEASE_MASTER
		if (entry->cpu != cluster->domain.release_master) {
#endif
			entry->scheduled = t;
			tsk_rt(t)->scheduled_on = task_cpu(t);
#ifdef CONFIG_RELEASE_MASTER
		} else {
			/* do not schedule on release master */
			preempt(entry); /* force resched */
			tsk_rt(t)->scheduled_on = NO_CPU;
		}
#endif
	} else {
		t->rt_param.scheduled_on = NO_CPU;
	}
	t->rt_param.linked_on          = NO_CPU;

	cedf_job_arrival(t);
	raw_spin_unlock_irqrestore(&(cluster->cluster_lock), flags);
}

static void cedf_task_wake_up(struct task_struct *task)
{
	unsigned long flags;
	//lt_t now;
	cedf_domain_t *cluster;

	TRACE_TASK(task, "wake_up at %llu\n", litmus_clock());

	cluster = task_cpu_cluster(task);

	raw_spin_lock_irqsave(&cluster->cluster_lock, flags);

#if 0 // sproadic task model
	/* We need to take suspensions because of semaphores into
	 * account! If a job resumes after being suspended due to acquiring
	 * a semaphore, it should never be treated as a new job release.
	 */
	if (get_rt_flags(task) == RT_F_EXIT_SEM) {
		set_rt_flags(task, RT_F_RUNNING);
	} else {
		now = litmus_clock();
		if (is_tardy(task, now)) {
			/* new sporadic release */
			release_at(task, now);
			sched_trace_task_release(task);
		}
		else {
			if (task->rt.time_slice) {
				/* came back in time before deadline
				*/
				set_rt_flags(task, RT_F_RUNNING);
			}
		}
	}
#else
	set_rt_flags(task, RT_F_RUNNING);  // periodic model
#endif

	if(tsk_rt(task)->linked_on == NO_CPU)
		cedf_job_arrival(task);

	raw_spin_unlock_irqrestore(&cluster->cluster_lock, flags);
}

static void cedf_task_block(struct task_struct *t)
{
	unsigned long flags;
	cedf_domain_t *cluster;

	TRACE_TASK(t, "block at %llu\n", litmus_clock());

	cluster = task_cpu_cluster(t);

	/* unlink if necessary */
	raw_spin_lock_irqsave(&cluster->cluster_lock, flags);
	unlink(t);
	raw_spin_unlock_irqrestore(&cluster->cluster_lock, flags);

	BUG_ON(!is_realtime(t));
}


static void cedf_task_exit(struct task_struct * t)
{
	unsigned long flags;
	cedf_domain_t *cluster = task_cpu_cluster(t);

#ifdef CONFIG_LITMUS_PAI_SOFTIRQD
	cedf_change_prio_pai_tasklet(t, NULL);
#endif

	/* unlink if necessary */
	raw_spin_lock_irqsave(&cluster->cluster_lock, flags);
	unlink(t);
	if (tsk_rt(t)->scheduled_on != NO_CPU) {
		cpu_entry_t *cpu;
		cpu = &per_cpu(cedf_cpu_entries, tsk_rt(t)->scheduled_on);
		cpu->scheduled = NULL;
		tsk_rt(t)->scheduled_on = NO_CPU;
	}
	raw_spin_unlock_irqrestore(&cluster->cluster_lock, flags);

	BUG_ON(!is_realtime(t));
        TRACE_TASK(t, "RIP\n");
}

static long cedf_admit_task(struct task_struct* tsk)
{
#ifdef CONFIG_LITMUS_NESTED_LOCKING
	INIT_BINHEAP_HANDLE(&tsk_rt(tsk)->hp_blocked_tasks,
						edf_max_heap_base_priority_order);
#endif

	return task_cpu(tsk) == tsk->rt_param.task_params.cpu ? 0 : -EINVAL;
}



#ifdef CONFIG_LITMUS_LOCKING

#include <litmus/fdso.h>



/* called with IRQs off */
static void __increase_priority_inheritance(struct task_struct* t,
										    struct task_struct* prio_inh)
{
	int linked_on;
	int check_preempt = 0;

	cedf_domain_t* cluster = task_cpu_cluster(t);

#ifdef CONFIG_LITMUS_NESTED_LOCKING
	/* this sanity check allows for weaker locking in protocols */
	/* TODO (klitirqd): Skip this check if 't' is a proxy thread (???) */
	if(__edf_higher_prio(prio_inh, BASE, t, EFFECTIVE)) {
#endif
		TRACE_TASK(t, "inherits priority from %s/%d\n",
				   prio_inh->comm, prio_inh->pid);
		tsk_rt(t)->inh_task = prio_inh;

		linked_on  = tsk_rt(t)->linked_on;

		/* If it is scheduled, then we need to reorder the CPU heap. */
		if (linked_on != NO_CPU) {
			TRACE_TASK(t, "%s: linked  on %d\n",
					   __FUNCTION__, linked_on);
			/* Holder is scheduled; need to re-order CPUs.
			 * We can't use heap_decrease() here since
			 * the cpu_heap is ordered in reverse direction, so
			 * it is actually an increase. */
			binheap_delete(&per_cpu(cedf_cpu_entries, linked_on).hn,
						   &cluster->cpu_heap);
			binheap_add(&per_cpu(cedf_cpu_entries, linked_on).hn,
						&cluster->cpu_heap, cpu_entry_t, hn);

		} else {
			/* holder may be queued: first stop queue changes */
			raw_spin_lock(&cluster->domain.release_lock);
			if (is_queued(t)) {
				TRACE_TASK(t, "%s: is queued\n",
						   __FUNCTION__);
				/* We need to update the position of holder in some
				 * heap. Note that this could be a release heap if we
				 * budget enforcement is used and this job overran. */
				check_preempt =
					!bheap_decrease(edf_ready_order, tsk_rt(t)->heap_node);
			} else {
				/* Nothing to do: if it is not queued and not linked
				 * then it is either sleeping or currently being moved
				 * by other code (e.g., a timer interrupt handler) that
				 * will use the correct priority when enqueuing the
				 * task. */
				TRACE_TASK(t, "%s: is NOT queued => Done.\n",
						   __FUNCTION__);
			}
			raw_spin_unlock(&cluster->domain.release_lock);

			/* If holder was enqueued in a release heap, then the following
			 * preemption check is pointless, but we can't easily detect
			 * that case. If you want to fix this, then consider that
			 * simply adding a state flag requires O(n) time to update when
			 * releasing n tasks, which conflicts with the goal to have
			 * O(log n) merges. */
			if (check_preempt) {
				/* heap_decrease() hit the top level of the heap: make
				 * sure preemption checks get the right task, not the
				 * potentially stale cache. */
				bheap_uncache_min(edf_ready_order,
								  &cluster->domain.ready_queue);
				check_for_preemptions(cluster);
			}
		}
#ifdef CONFIG_LITMUS_NESTED_LOCKING
	}
	else {
		TRACE_TASK(t, "Spurious invalid priority increase. "
				   "Inheritance request: %s/%d [eff_prio = %s/%d] to inherit from %s/%d\n"
				   "Occurance is likely okay: probably due to (hopefully safe) concurrent priority updates.\n",
				   t->comm, t->pid,
				   effective_priority(t)->comm, effective_priority(t)->pid,
				   (prio_inh) ? prio_inh->comm : "nil",
				   (prio_inh) ? prio_inh->pid : -1);
		WARN_ON(!prio_inh);
	}
#endif
}

/* called with IRQs off */
static void increase_priority_inheritance(struct task_struct* t, struct task_struct* prio_inh)
{
	cedf_domain_t* cluster = task_cpu_cluster(t);

	raw_spin_lock(&cluster->cluster_lock);

	__increase_priority_inheritance(t, prio_inh);

#ifdef CONFIG_LITMUS_SOFTIRQD
	if(tsk_rt(t)->cur_klitirqd != NULL)
	{
		TRACE_TASK(t, "%s/%d inherits a new priority!\n",
				   tsk_rt(t)->cur_klitirqd->comm, tsk_rt(t)->cur_klitirqd->pid);

		__increase_priority_inheritance(tsk_rt(t)->cur_klitirqd, prio_inh);
	}
#endif

	raw_spin_unlock(&cluster->cluster_lock);

#if defined(CONFIG_LITMUS_PAI_SOFTIRQD) && defined(CONFIG_LITMUS_NVIDIA)
	if(tsk_rt(t)->held_gpus) {
		int i;
		for(i = find_first_bit(&tsk_rt(t)->held_gpus, sizeof(tsk_rt(t)->held_gpus));
			i < NV_DEVICE_NUM;
			i = find_next_bit(&tsk_rt(t)->held_gpus, sizeof(tsk_rt(t)->held_gpus), i+1)) {
			pai_check_priority_increase(t, i);
		}
	}
#endif
}

/* called with IRQs off */
static void __decrease_priority_inheritance(struct task_struct* t,
											struct task_struct* prio_inh)
{
#ifdef CONFIG_LITMUS_NESTED_LOCKING
	if(__edf_higher_prio(t, EFFECTIVE, prio_inh, BASE)) {
#endif
		/* A job only stops inheriting a priority when it releases a
		 * resource. Thus we can make the following assumption.*/
		if(prio_inh)
			TRACE_TASK(t, "EFFECTIVE priority decreased to %s/%d\n",
					   prio_inh->comm, prio_inh->pid);
		else
			TRACE_TASK(t, "base priority restored.\n");

		tsk_rt(t)->inh_task = prio_inh;

		if(tsk_rt(t)->scheduled_on != NO_CPU) {
			TRACE_TASK(t, "is scheduled.\n");

			/* Check if rescheduling is necessary. We can't use heap_decrease()
			 * since the priority was effectively lowered. */
			unlink(t);
			cedf_job_arrival(t);
		}
		else {
			cedf_domain_t* cluster = task_cpu_cluster(t);
			/* task is queued */
			raw_spin_lock(&cluster->domain.release_lock);
			if (is_queued(t)) {
				TRACE_TASK(t, "is queued.\n");

				/* decrease in priority, so we have to re-add to binomial heap */
				unlink(t);
				cedf_job_arrival(t);
			}
			else {
				TRACE_TASK(t, "is not in scheduler. Probably on wait queue somewhere.\n");
			}
			raw_spin_unlock(&cluster->domain.release_lock);
		}
#ifdef CONFIG_LITMUS_NESTED_LOCKING
	}
	else {
		TRACE_TASK(t, "Spurious invalid priority decrease. "
				   "Inheritance request: %s/%d [eff_prio = %s/%d] to inherit from %s/%d\n"
				   "Occurance is likely okay: probably due to (hopefully safe) concurrent priority updates.\n",
				   t->comm, t->pid,
				   effective_priority(t)->comm, effective_priority(t)->pid,
				   (prio_inh) ? prio_inh->comm : "nil",
				   (prio_inh) ? prio_inh->pid : -1);
	}
#endif
}

static void decrease_priority_inheritance(struct task_struct* t,
										struct task_struct* prio_inh)
{
	cedf_domain_t* cluster = task_cpu_cluster(t);

	raw_spin_lock(&cluster->cluster_lock);
	__decrease_priority_inheritance(t, prio_inh);

#ifdef CONFIG_LITMUS_SOFTIRQD
	if(tsk_rt(t)->cur_klitirqd != NULL)
	{
		TRACE_TASK(t, "%s/%d decreases in priority!\n",
				   tsk_rt(t)->cur_klitirqd->comm, tsk_rt(t)->cur_klitirqd->pid);

		__decrease_priority_inheritance(tsk_rt(t)->cur_klitirqd, prio_inh);
	}
#endif

	raw_spin_unlock(&cluster->cluster_lock);

#if defined(CONFIG_LITMUS_PAI_SOFTIRQD) && defined(CONFIG_LITMUS_NVIDIA)
	if(tsk_rt(t)->held_gpus) {
		int i;
		for(i = find_first_bit(&tsk_rt(t)->held_gpus, sizeof(tsk_rt(t)->held_gpus));
			i < NV_DEVICE_NUM;
			i = find_next_bit(&tsk_rt(t)->held_gpus, sizeof(tsk_rt(t)->held_gpus), i+1)) {
			pai_check_priority_decrease(t, i);
		}
	}
#endif
}





#ifdef CONFIG_LITMUS_SOFTIRQD
/* called with IRQs off */
static void increase_priority_inheritance_klitirqd(struct task_struct* klitirqd,
											  struct task_struct* old_owner,
											  struct task_struct* new_owner)
{
	cedf_domain_t* cluster = task_cpu_cluster(klitirqd);

	BUG_ON(!(tsk_rt(klitirqd)->is_proxy_thread));

	raw_spin_lock(&cluster->cluster_lock);

	if(old_owner != new_owner)
	{
		if(old_owner)
		{
			// unreachable?
			tsk_rt(old_owner)->cur_klitirqd = NULL;
		}

		TRACE_TASK(klitirqd, "giving ownership to %s/%d.\n",
				   new_owner->comm, new_owner->pid);

		tsk_rt(new_owner)->cur_klitirqd = klitirqd;
	}

	__decrease_priority_inheritance(klitirqd, NULL);  // kludge to clear out cur prio.

	__increase_priority_inheritance(klitirqd,
			(tsk_rt(new_owner)->inh_task == NULL) ?
				new_owner :
				tsk_rt(new_owner)->inh_task);

	raw_spin_unlock(&cluster->cluster_lock);
}


/* called with IRQs off */
static void decrease_priority_inheritance_klitirqd(struct task_struct* klitirqd,
												   struct task_struct* old_owner,
												   struct task_struct* new_owner)
{
	cedf_domain_t* cluster = task_cpu_cluster(klitirqd);

	BUG_ON(!(tsk_rt(klitirqd)->is_proxy_thread));

	raw_spin_lock(&cluster->cluster_lock);

    TRACE_TASK(klitirqd, "priority restored\n");

	__decrease_priority_inheritance(klitirqd, new_owner);

	tsk_rt(old_owner)->cur_klitirqd = NULL;

	raw_spin_unlock(&cluster->cluster_lock);
}
#endif // CONFIG_LITMUS_SOFTIRQD







#ifdef CONFIG_LITMUS_NESTED_LOCKING

/* called with IRQs off */
/* preconditions:
 (1) The 'hp_blocked_tasks_lock' of task 't' is held.
 (2) The lock 'to_unlock' is held.
 */
static void nested_increase_priority_inheritance(struct task_struct* t,
												 struct task_struct* prio_inh,
												 raw_spinlock_t *to_unlock,
												 unsigned long irqflags)
{
	struct litmus_lock *blocked_lock = tsk_rt(t)->blocked_lock;

	if(tsk_rt(t)->inh_task != prio_inh) { 		// shield redundent calls.
		increase_priority_inheritance(t, prio_inh);  // increase our prio.
	}

	raw_spin_unlock(&tsk_rt(t)->hp_blocked_tasks_lock);  // unlock the t's heap.


	if(blocked_lock) {
		if(blocked_lock->ops->propagate_increase_inheritance) {
			TRACE_TASK(t, "Inheritor is blocked (...perhaps).  Checking lock %d.\n",
					   blocked_lock->ident);

			// beware: recursion
			blocked_lock->ops->propagate_increase_inheritance(blocked_lock,
															  t, to_unlock,
															  irqflags);
		}
		else {
			TRACE_TASK(t, "Inheritor is blocked on lock (%d) that does not support nesting!\n",
					   blocked_lock->ident);
			unlock_fine_irqrestore(to_unlock, irqflags);
		}
	}
	else {
		TRACE_TASK(t, "is not blocked.  No propagation.\n");
		unlock_fine_irqrestore(to_unlock, irqflags);
	}
}

/* called with IRQs off */
/* preconditions:
 (1) The 'hp_blocked_tasks_lock' of task 't' is held.
 (2) The lock 'to_unlock' is held.
 */
static void nested_decrease_priority_inheritance(struct task_struct* t,
												 struct task_struct* prio_inh,
												 raw_spinlock_t *to_unlock,
												 unsigned long irqflags)
{
	struct litmus_lock *blocked_lock = tsk_rt(t)->blocked_lock;
	decrease_priority_inheritance(t, prio_inh);

	raw_spin_unlock(&tsk_rt(t)->hp_blocked_tasks_lock);  // unlock the t's heap.

	if(blocked_lock) {
		if(blocked_lock->ops->propagate_decrease_inheritance) {
			TRACE_TASK(t, "Inheritor is blocked (...perhaps).  Checking lock %d.\n",
					   blocked_lock->ident);

			// beware: recursion
			blocked_lock->ops->propagate_decrease_inheritance(blocked_lock, t,
															  to_unlock,
															  irqflags);
		}
		else {
			TRACE_TASK(t, "Inheritor is blocked on lock (%p) that does not support nesting!\n",
					   blocked_lock);
			unlock_fine_irqrestore(to_unlock, irqflags);
		}
	}
	else {
		TRACE_TASK(t, "is not blocked.  No propagation.\n");
		unlock_fine_irqrestore(to_unlock, irqflags);
	}
}


/* ******************** RSM MUTEX ********************** */

static struct litmus_lock_ops cedf_rsm_mutex_lock_ops = {
	.lock   = rsm_mutex_lock,
	.unlock = rsm_mutex_unlock,
	.close  = rsm_mutex_close,
	.deallocate = rsm_mutex_free,

	.propagate_increase_inheritance = rsm_mutex_propagate_increase_inheritance,
	.propagate_decrease_inheritance = rsm_mutex_propagate_decrease_inheritance,

#ifdef CONFIG_LITMUS_DGL_SUPPORT
	.dgl_lock = rsm_mutex_dgl_lock,
	.is_owner = rsm_mutex_is_owner,
	.enable_priority = rsm_mutex_enable_priority,
#endif
};

static struct litmus_lock* cedf_new_rsm_mutex(void)
{
	return rsm_mutex_new(&cedf_rsm_mutex_lock_ops);
}

/* ******************** IKGLP ********************** */

static struct litmus_lock_ops cedf_ikglp_lock_ops = {
	.lock   = ikglp_lock,
	.unlock = ikglp_unlock,
	.close  = ikglp_close,
	.deallocate = ikglp_free,

	// ikglp can only be an outer-most lock.
	.propagate_increase_inheritance = NULL,
	.propagate_decrease_inheritance = NULL,
};

static struct litmus_lock* cedf_new_ikglp(void* __user arg)
{
	// assumes clusters of uniform size.
	return ikglp_new(cluster_size/num_clusters, &cedf_ikglp_lock_ops, arg);
}

#endif  /* CONFIG_LITMUS_NESTED_LOCKING */




/* ******************** KFMLP support ********************** */

static struct litmus_lock_ops cedf_kfmlp_lock_ops = {
	.lock   = kfmlp_lock,
	.unlock = kfmlp_unlock,
	.close  = kfmlp_close,
	.deallocate = kfmlp_free,

	// kfmlp can only be an outer-most lock.
	.propagate_increase_inheritance = NULL,
	.propagate_decrease_inheritance = NULL,
};


static struct litmus_lock* cedf_new_kfmlp(void* __user arg)
{
	return kfmlp_new(&cedf_kfmlp_lock_ops, arg);
}


/* **** lock constructor **** */

static long cedf_allocate_lock(struct litmus_lock **lock, int type,
								 void* __user args)
{
	int err;

	switch (type) {
#ifdef CONFIG_LITMUS_NESTED_LOCKING
		case RSM_MUTEX:
			*lock = cedf_new_rsm_mutex();
			break;

		case IKGLP_SEM:
			*lock = cedf_new_ikglp(args);
			break;
#endif
		case KFMLP_SEM:
			*lock = cedf_new_kfmlp(args);
			break;

		default:
			err = -ENXIO;
			goto UNSUPPORTED_LOCK;
	};

	if (*lock)
		err = 0;
	else
		err = -ENOMEM;

UNSUPPORTED_LOCK:
	return err;
}

#endif  // CONFIG_LITMUS_LOCKING


#ifdef CONFIG_LITMUS_AFFINITY_LOCKING
static struct affinity_observer_ops cedf_kfmlp_affinity_ops = {
	.close = kfmlp_aff_obs_close,
	.deallocate = kfmlp_aff_obs_free,
};

#ifdef CONFIG_LITMUS_NESTED_LOCKING
static struct affinity_observer_ops cedf_ikglp_affinity_ops = {
	.close = ikglp_aff_obs_close,
	.deallocate = ikglp_aff_obs_free,
};
#endif

static long cedf_allocate_affinity_observer(struct affinity_observer **aff_obs,
											int type,
											void* __user args)
{
	int err;

	switch (type) {

		case KFMLP_SIMPLE_GPU_AFF_OBS:
			*aff_obs = kfmlp_simple_gpu_aff_obs_new(&cedf_kfmlp_affinity_ops, args);
			break;

		case KFMLP_GPU_AFF_OBS:
			*aff_obs = kfmlp_gpu_aff_obs_new(&cedf_kfmlp_affinity_ops, args);
			break;

#ifdef CONFIG_LITMUS_NESTED_LOCKING
		case IKGLP_SIMPLE_GPU_AFF_OBS:
			*aff_obs = ikglp_simple_gpu_aff_obs_new(&cedf_ikglp_affinity_ops, args);
			break;

		case IKGLP_GPU_AFF_OBS:
			*aff_obs = ikglp_gpu_aff_obs_new(&cedf_ikglp_affinity_ops, args);
			break;
#endif
		default:
			err = -ENXIO;
			goto UNSUPPORTED_AFF_OBS;
	};

	if (*aff_obs)
		err = 0;
	else
		err = -ENOMEM;

UNSUPPORTED_AFF_OBS:
	return err;
}
#endif




#ifdef VERBOSE_INIT
static void print_cluster_topology(cpumask_var_t mask, int cpu)
{
	int chk;
	char buf[255];

	chk = cpulist_scnprintf(buf, 254, mask);
	buf[chk] = '\0';
	printk(KERN_INFO "CPU = %d, shared cpu(s) = %s\n", cpu, buf);

}
#endif

static void cleanup_cedf(void)
{
	int i;

#ifdef CONFIG_LITMUS_NVIDIA
	shutdown_nvidia_info();
#endif

	if (clusters_allocated) {
		for (i = 0; i < num_clusters; i++) {
			kfree(cedf[i].cpus);
			free_cpumask_var(cedf[i].cpu_map);
		}

		kfree(cedf);
	}
}

static long cedf_activate_plugin(void)
{
	int i, j, cpu, ccpu, cpu_count;
	cpu_entry_t *entry;

	cpumask_var_t mask;
	int chk = 0;

	/* de-allocate old clusters, if any */
	cleanup_cedf();

	printk(KERN_INFO "C-EDF: Activate Plugin, cluster configuration = %d\n",
			cluster_config);

	/* need to get cluster_size first */
	if(!zalloc_cpumask_var(&mask, GFP_ATOMIC))
		return -ENOMEM;

	if (unlikely(cluster_config == GLOBAL_CLUSTER)) {
		cluster_size = num_online_cpus();
	} else {
		chk = get_shared_cpu_map(mask, 0, cluster_config);
		if (chk) {
			/* if chk != 0 then it is the max allowed index */
			printk(KERN_INFO "C-EDF: Cluster configuration = %d "
			       "is not supported on this hardware.\n",
			       cluster_config);
			/* User should notice that the configuration failed, so
			 * let's bail out. */
			return -EINVAL;
		}

		cluster_size = cpumask_weight(mask);
	}

	if ((num_online_cpus() % cluster_size) != 0) {
		/* this can't be right, some cpus are left out */
		printk(KERN_ERR "C-EDF: Trying to group %d cpus in %d!\n",
				num_online_cpus(), cluster_size);
		return -1;
	}

	num_clusters = num_online_cpus() / cluster_size;
	printk(KERN_INFO "C-EDF: %d cluster(s) of size = %d\n",
			num_clusters, cluster_size);

	/* initialize clusters */
	cedf = kmalloc(num_clusters * sizeof(cedf_domain_t), GFP_ATOMIC);
	for (i = 0; i < num_clusters; i++) {

		cedf[i].cpus = kmalloc(cluster_size * sizeof(cpu_entry_t),
				GFP_ATOMIC);
		INIT_BINHEAP_HANDLE(&(cedf[i].cpu_heap), cpu_lower_prio);
		edf_domain_init(&(cedf[i].domain), NULL, cedf_release_jobs);


#ifdef CONFIG_LITMUS_PAI_SOFTIRQD
		cedf[i].pending_tasklets.head = NULL;
		cedf[i].pending_tasklets.tail = &(cedf[i].pending_tasklets.head);
#endif


		if(!zalloc_cpumask_var(&cedf[i].cpu_map, GFP_ATOMIC))
			return -ENOMEM;
#ifdef CONFIG_RELEASE_MASTER
		cedf[i].domain.release_master = atomic_read(&release_master_cpu);
#endif
	}

	/* cycle through cluster and add cpus to them */
	for (i = 0; i < num_clusters; i++) {

#ifdef CONFIG_LITMUS_DGL_SUPPORT
		raw_spin_lock_init(&cedf[i].dgl_lock);
#endif

		for_each_online_cpu(cpu) {
			/* check if the cpu is already in a cluster */
			for (j = 0; j < num_clusters; j++)
				if (cpumask_test_cpu(cpu, cedf[j].cpu_map))
					break;
			/* if it is in a cluster go to next cpu */
			if (j < num_clusters &&
					cpumask_test_cpu(cpu, cedf[j].cpu_map))
				continue;

			/* this cpu isn't in any cluster */
			/* get the shared cpus */
			if (unlikely(cluster_config == GLOBAL_CLUSTER))
				cpumask_copy(mask, cpu_online_mask);
			else
				get_shared_cpu_map(mask, cpu, cluster_config);

			cpumask_copy(cedf[i].cpu_map, mask);
#ifdef VERBOSE_INIT
			print_cluster_topology(mask, cpu);
#endif
			/* add cpus to current cluster and init cpu_entry_t */
			cpu_count = 0;
			for_each_cpu(ccpu, cedf[i].cpu_map) {

				entry = &per_cpu(cedf_cpu_entries, ccpu);
				cedf[i].cpus[cpu_count] = entry;
				atomic_set(&entry->will_schedule, 0);
				entry->cpu = ccpu;
				entry->cluster = &cedf[i];

				INIT_BINHEAP_NODE(&entry->hn);

				cpu_count++;

				entry->linked = NULL;
				entry->scheduled = NULL;
#ifdef CONFIG_RELEASE_MASTER
				/* only add CPUs that should schedule jobs */
				if (entry->cpu != entry->cluster->domain.release_master)
#endif
					update_cpu_position(entry);
			}
			/* done with this cluster */
			break;
		}
	}

#ifdef CONFIG_LITMUS_SOFTIRQD
	{
		/* distribute the daemons evenly across the clusters. */
		int* affinity = kmalloc(NR_LITMUS_SOFTIRQD * sizeof(int), GFP_ATOMIC);
		int num_daemons_per_cluster = NR_LITMUS_SOFTIRQD / num_clusters;
		int left_over = NR_LITMUS_SOFTIRQD % num_clusters;

		int daemon = 0;
		for(i = 0; i < num_clusters; ++i)
		{
			int num_on_this_cluster = num_daemons_per_cluster;
			if(left_over)
			{
				++num_on_this_cluster;
				--left_over;
			}

			for(j = 0; j < num_on_this_cluster; ++j)
			{
				// first CPU of this cluster
				affinity[daemon++] = i*cluster_size;
			}
		}

		spawn_klitirqd(affinity);

		kfree(affinity);
	}
#endif

#ifdef CONFIG_LITMUS_NVIDIA
	init_nvidia_info();
#endif

	free_cpumask_var(mask);
	clusters_allocated = 1;
	return 0;
}

/*	Plugin object	*/
static struct sched_plugin cedf_plugin __cacheline_aligned_in_smp = {
	.plugin_name		= "C-EDF",
	.finish_switch		= cedf_finish_switch,
	.tick			= cedf_tick,
	.task_new		= cedf_task_new,
	.complete_job		= complete_job,
	.task_exit		= cedf_task_exit,
	.schedule		= cedf_schedule,
	.task_wake_up		= cedf_task_wake_up,
	.task_block		= cedf_task_block,
	.admit_task		= cedf_admit_task,
	.activate_plugin	= cedf_activate_plugin,
	.compare		= edf_higher_prio,
#ifdef CONFIG_LITMUS_LOCKING
	.allocate_lock		= cedf_allocate_lock,
	.increase_prio		= increase_priority_inheritance,
	.decrease_prio		= decrease_priority_inheritance,
#endif
#ifdef CONFIG_LITMUS_NESTED_LOCKING
	.nested_increase_prio		= nested_increase_priority_inheritance,
	.nested_decrease_prio		= nested_decrease_priority_inheritance,
	.__compare					= __edf_higher_prio,
#endif
#ifdef CONFIG_LITMUS_DGL_SUPPORT
	.get_dgl_spinlock = cedf_get_dgl_spinlock,
#endif
#ifdef CONFIG_LITMUS_AFFINITY_LOCKING
	.allocate_aff_obs = cedf_allocate_affinity_observer,
#endif
#ifdef CONFIG_LITMUS_SOFTIRQD
	.increase_prio_klitirqd = increase_priority_inheritance_klitirqd,
	.decrease_prio_klitirqd = decrease_priority_inheritance_klitirqd,
#endif
#ifdef CONFIG_LITMUS_PAI_SOFTIRQD
	.enqueue_pai_tasklet = cedf_enqueue_pai_tasklet,
	.change_prio_pai_tasklet = cedf_change_prio_pai_tasklet,
	.run_tasklets = cedf_run_tasklets,
#endif
};

static struct proc_dir_entry *cluster_file = NULL, *cedf_dir = NULL;

static int __init init_cedf(void)
{
	int err, fs;

	err = register_sched_plugin(&cedf_plugin);
	if (!err) {
		fs = make_plugin_proc_dir(&cedf_plugin, &cedf_dir);
		if (!fs)
			cluster_file = create_cluster_file(cedf_dir, &cluster_config);
		else
			printk(KERN_ERR "Could not allocate C-EDF procfs dir.\n");
	}
	return err;
}

static void clean_cedf(void)
{
	cleanup_cedf();
	if (cluster_file)
		remove_proc_entry("cluster", cedf_dir);
	if (cedf_dir)
		remove_plugin_proc_dir(&cedf_plugin);
}

module_init(init_cedf);
module_exit(clean_cedf);