Initial GSN-EDF-split plugin

2 files changed, 966 insertions, 0 deletions

diff --git a/litmus/Makefile b/litmus/Makefile
index 7338180f196f..8efcb699935b 100644
--- a/litmus/Makefile
+++ b/litmus/Makefile
@@ -17,6 +17,7 @@ obj-y     = sched_plugin.o litmus.o \
             bheap.o \
             ctrldev.o \
             sched_gsn_edf.o \
+        sched_gsn_edf_split.o \
             sched_psn_edf.o
 
 obj-$(CONFIG_PLUGIN_CEDF) += sched_cedf.o
diff --git a/litmus/sched_gsn_edf_split.c b/litmus/sched_gsn_edf_split.c
new file mode 100644
index 000000000000..1a4b06391d12
--- /dev/null
+++ b/litmus/sched_gsn_edf_split.c
@@ -0,0 +1,965 @@
+/*
+ * litmus/sched_gsn_edf.c
+ *
+ * Implementation of the GSN-EDF scheduling algorithm with job splitting, i.e.
+ * GSN-EDF-split.
+ *
+ * This plugin is a modified version of the prior GSN-EDF plugin in
+ * litmus/sched_gsn_edf_split.c
+ */
+
+#include <linux/spinlock.h>
+#include <linux/percpu.h>
+#include <linux/sched.h>
+#include <linux/slab.h>
+
+#include <litmus/litmus.h>
+#include <litmus/jobs.h>
+#include <litmus/sched_plugin.h>
+#include <litmus/edf_common.h>
+#include <litmus/sched_trace.h>
+#include <litmus/trace.h>
+
+#include <litmus/preempt.h>
+
+#include <litmus/bheap.h>
+
+#ifdef CONFIG_SCHED_CPU_AFFINITY
+#include <litmus/affinity.h>
+#endif
+
+#include <linux/module.h>
+
+/* cpu_entry_t - maintain the linked and scheduled state
+ */
+typedef struct  {
+        int                     cpu;
+        struct task_struct*     linked;         /* only RT tasks */
+        struct task_struct*     scheduled;      /* only RT tasks */
+        struct bheap_node*      hn;
+    struct hrtimer      split_timer;
+    int                 timer_armed;
+} cpu_entry_t;
+DEFINE_PER_CPU(cpu_entry_t, gsnedfsplit_cpu_entries);
+
+cpu_entry_t* gsnedfsplit_cpus[NR_CPUS];
+
+/* the cpus queue themselves according to priority in here */
+static struct bheap_node gsnedfsplit_heap_node[NR_CPUS];
+static struct bheap      gsnedfsplit_cpu_heap;
+
+static rt_domain_t gsnedfsplit;
+#define gsnedfsplit_lock (gsnedfsplit.ready_lock)
+
+
+/* Uncomment this if you want to see all scheduling decisions in the
+ * TRACE() log.
+#define WANT_ALL_SCHED_EVENTS
+ */
+
+static int cpu_lower_prio(struct bheap_node *_a, struct bheap_node *_b)
+{
+        cpu_entry_t *a, *b;
+        a = _a->value;
+        b = _b->value;
+        /* 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 gsnedfsplit lock.
+ */
+static void update_cpu_position(cpu_entry_t *entry)
+{
+        if (likely(bheap_node_in_heap(entry->hn)))
+                bheap_delete(cpu_lower_prio, &gsnedfsplit_cpu_heap, entry->hn);
+        bheap_insert(cpu_lower_prio, &gsnedfsplit_cpu_heap, entry->hn);
+}
+
+/* caller must hold gsnedfsplit lock */
+static cpu_entry_t* lowest_prio_cpu(void)
+{
+        struct bheap_node* hn;
+        hn = bheap_peek(cpu_lower_prio, &gsnedfsplit_cpu_heap);
+        return hn->value;
+}
+
+
+/* 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(gsnedfsplit_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 gsnedfsplit_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(gsnedfsplit_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.
+                 */
+                remove(&gsnedfsplit, 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 gsnedfsplit_lock.
+ */
+static noinline void requeue(struct task_struct* task)
+{
+        BUG_ON(!task);
+        /* sanity check before insertion */
+        BUG_ON(is_queued(task));
+
+        if (is_released(task, litmus_clock()))
+                __add_ready(&gsnedfsplit, task);
+        else {
+                /* it has got to wait */
+                add_release(&gsnedfsplit, task);
+        }
+}
+
+#ifdef CONFIG_SCHED_CPU_AFFINITY
+static cpu_entry_t* gsnedfsplit_get_nearest_available_cpu(cpu_entry_t *start)
+{
+        cpu_entry_t *affinity;
+
+        get_nearest_available_cpu(affinity, start, gsnedfsplit_cpu_entries,
+#ifdef CONFIG_RELEASE_MASTER
+                        gsnedfsplit.release_master
+#else
+                        NO_CPU
+#endif
+                        );
+
+        return(affinity);
+}
+#endif
+
+/* check for any necessary preemptions */
+static void check_for_preemptions(void)
+{
+        struct task_struct *task;
+        cpu_entry_t *last;
+
+        for (last = lowest_prio_cpu();
+             edf_preemption_needed(&gsnedfsplit, last->linked);
+             last = lowest_prio_cpu()) {
+                /* preemption necessary */
+                task = __take_ready(&gsnedfsplit);
+                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 =
+                                        gsnedfsplit_get_nearest_available_cpu(
+                                                &per_cpu(gsnedfsplit_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);
+        }
+}
+
+/* gsnedfsplit_job_arrival: task is either resumed or released */
+static noinline void gsnedfsplit_job_arrival(struct task_struct* task)
+{
+        BUG_ON(!task);
+
+        requeue(task);
+        check_for_preemptions();
+}
+
+static void gsnedfsplit_release_jobs(rt_domain_t* rt, struct bheap* tasks)
+{
+        unsigned long flags;
+
+        raw_spin_lock_irqsave(&gsnedfsplit_lock, flags);
+
+        __merge_ready(rt, tasks);
+        check_for_preemptions();
+
+        raw_spin_unlock_irqrestore(&gsnedfsplit_lock, flags);
+}
+
+/* caller holds gsnedfsplit_lock */
+static noinline void job_completion(struct task_struct *t, int forced)
+{
+        BUG_ON(!t);
+
+        sched_trace_task_completion(t, forced);
+
+        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))
+                gsnedfsplit_job_arrival(t);
+}
+
+/* gsnedfsplit_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 gsnedfsplit_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();
+                        TRACE("gsnedfsplit_scheduler_tick: "
+                              "%d is preemptable "
+                              " => FORCE_RESCHED\n", t->pid);
+                } else if (is_user_np(t)) {
+                        TRACE("gsnedfsplit_scheduler_tick: "
+                              "%d is non-preemptable, "
+                              "preemption delayed.\n", t->pid);
+                        request_exit_np(t);
+                }
+        }
+}
+
+/* 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* gsnedfsplit_schedule(struct task_struct * prev)
+{
+        cpu_entry_t* entry = &__get_cpu_var(gsnedfsplit_cpu_entries);
+        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(gsnedfsplit.release_master == entry->cpu)) {
+                sched_state_task_picked();
+                return NULL;
+        }
+#endif
+
+        raw_spin_lock(&gsnedfsplit_lock);
+
+        /* 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 gsnedfsplit_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);
+
+        /* 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(&gsnedfsplit), 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;
+                        TRACE_TASK(next, "scheduled_on = P%d\n", smp_processor_id());
+                }
+                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(&gsnedfsplit_lock);
+
+#ifdef WANT_ALL_SCHED_EVENTS
+        TRACE("gsnedfsplit_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 gsnedfsplit_finish_switch(struct task_struct *prev)
+{
+        cpu_entry_t*    entry = &__get_cpu_var(gsnedfsplit_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 gsnedfsplit_task_new(struct task_struct * t, int on_rq, int running)
+{
+        unsigned long           flags;
+        cpu_entry_t*            entry;
+
+        TRACE("gsn edf: task new %d\n", t->pid);
+
+        raw_spin_lock_irqsave(&gsnedfsplit_lock, flags);
+
+        /* setup job params */
+        release_at(t, litmus_clock());
+
+        if (running) {
+                entry = &per_cpu(gsnedfsplit_cpu_entries, task_cpu(t));
+                BUG_ON(entry->scheduled);
+
+#ifdef CONFIG_RELEASE_MASTER
+                if (entry->cpu != gsnedfsplit.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;
+
+        gsnedfsplit_job_arrival(t);
+        raw_spin_unlock_irqrestore(&gsnedfsplit_lock, flags);
+}
+
+static void gsnedfsplit_task_wake_up(struct task_struct *task)
+{
+        unsigned long flags;
+        lt_t now;
+
+        TRACE_TASK(task, "wake_up at %llu\n", litmus_clock());
+
+        raw_spin_lock_irqsave(&gsnedfsplit_lock, flags);
+        /* 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);
+                        }
+                }
+        }
+        gsnedfsplit_job_arrival(task);
+        raw_spin_unlock_irqrestore(&gsnedfsplit_lock, flags);
+}
+
+static void gsnedfsplit_task_block(struct task_struct *t)
+{
+        unsigned long flags;
+
+        TRACE_TASK(t, "block at %llu\n", litmus_clock());
+
+        /* unlink if necessary */
+        raw_spin_lock_irqsave(&gsnedfsplit_lock, flags);
+        unlink(t);
+        raw_spin_unlock_irqrestore(&gsnedfsplit_lock, flags);
+
+        BUG_ON(!is_realtime(t));
+}
+
+
+static void gsnedfsplit_task_exit(struct task_struct * t)
+{
+        unsigned long flags;
+
+        /* unlink if necessary */
+        raw_spin_lock_irqsave(&gsnedfsplit_lock, flags);
+        unlink(t);
+        if (tsk_rt(t)->scheduled_on != NO_CPU) {
+                gsnedfsplit_cpus[tsk_rt(t)->scheduled_on]->scheduled = NULL;
+                tsk_rt(t)->scheduled_on = NO_CPU;
+        }
+        raw_spin_unlock_irqrestore(&gsnedfsplit_lock, flags);
+
+        BUG_ON(!is_realtime(t));
+        TRACE_TASK(t, "RIP\n");
+}
+
+
+static long gsnedfsplit_admit_task(struct task_struct* tsk)
+{
+        return 0;
+}
+
+#ifdef CONFIG_LITMUS_LOCKING
+
+#include <litmus/fdso.h>
+
+/* called with IRQs off */
+static void set_priority_inheritance(struct task_struct* t, struct task_struct* prio_inh)
+{
+        int linked_on;
+        int check_preempt = 0;
+
+        raw_spin_lock(&gsnedfsplit_lock);
+
+        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. */
+                bheap_delete(cpu_lower_prio, &gsnedfsplit_cpu_heap,
+                            gsnedfsplit_cpus[linked_on]->hn);
+                bheap_insert(cpu_lower_prio, &gsnedfsplit_cpu_heap,
+                            gsnedfsplit_cpus[linked_on]->hn);
+        } else {
+                /* holder may be queued: first stop queue changes */
+                raw_spin_lock(&gsnedfsplit.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(&gsnedfsplit.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,
+                                         &gsnedfsplit.ready_queue);
+                        check_for_preemptions();
+                }
+        }
+
+        raw_spin_unlock(&gsnedfsplit_lock);
+}
+
+/* called with IRQs off */
+static void clear_priority_inheritance(struct task_struct* t)
+{
+        raw_spin_lock(&gsnedfsplit_lock);
+
+        /* A job only stops inheriting a priority when it releases a
+         * resource. Thus we can make the following assumption.*/
+        BUG_ON(tsk_rt(t)->scheduled_on == NO_CPU);
+
+        TRACE_TASK(t, "priority restored\n");
+        tsk_rt(t)->inh_task = NULL;
+
+        /* Check if rescheduling is necessary. We can't use heap_decrease()
+         * since the priority was effectively lowered. */
+        unlink(t);
+        gsnedfsplit_job_arrival(t);
+
+        raw_spin_unlock(&gsnedfsplit_lock);
+}
+
+
+/* ******************** FMLP support ********************** */
+
+/* struct for semaphore with priority inheritance */
+struct fmlp_semaphore {
+        struct litmus_lock litmus_lock;
+
+        /* current resource holder */
+        struct task_struct *owner;
+
+        /* highest-priority waiter */
+        struct task_struct *hp_waiter;
+
+        /* FIFO queue of waiting tasks */
+        wait_queue_head_t wait;
+};
+
+static inline struct fmlp_semaphore* fmlp_from_lock(struct litmus_lock* lock)
+{
+        return container_of(lock, struct fmlp_semaphore, litmus_lock);
+}
+
+/* caller is responsible for locking */
+static struct task_struct* find_hp_waiter(struct fmlp_semaphore *sem,
+                                   struct task_struct* skip)
+{
+        struct list_head        *pos;
+        struct task_struct      *queued, *found = NULL;
+
+        list_for_each(pos, &sem->wait.task_list) {
+                queued  = (struct task_struct*) list_entry(pos, wait_queue_t,
+                                                           task_list)->private;
+
+                /* Compare task prios, find high prio task. */
+                if (queued != skip && edf_higher_prio(queued, found))
+                        found = queued;
+        }
+        return found;
+}
+
+int gsnedfsplit_fmlp_lock(struct litmus_lock* l)
+{
+        struct task_struct* t = current;
+        struct fmlp_semaphore *sem = fmlp_from_lock(l);
+        wait_queue_t wait;
+        unsigned long flags;
+
+        if (!is_realtime(t))
+                return -EPERM;
+
+        spin_lock_irqsave(&sem->wait.lock, flags);
+
+        if (sem->owner) {
+                /* resource is not free => must suspend and wait */
+
+                init_waitqueue_entry(&wait, t);
+
+                /* FIXME: interruptible would be nice some day */
+                set_task_state(t, TASK_UNINTERRUPTIBLE);
+
+                __add_wait_queue_tail_exclusive(&sem->wait, &wait);
+
+                /* check if we need to activate priority inheritance */
+                if (edf_higher_prio(t, sem->hp_waiter)) {
+                        sem->hp_waiter = t;
+                        if (edf_higher_prio(t, sem->owner))
+                                set_priority_inheritance(sem->owner, sem->hp_waiter);
+                }
+
+                TS_LOCK_SUSPEND;
+
+                /* release lock before sleeping */
+                spin_unlock_irqrestore(&sem->wait.lock, flags);
+
+                /* We depend on the FIFO order.  Thus, we don't need to recheck
+                 * when we wake up; we are guaranteed to have the lock since
+                 * there is only one wake up per release.
+                 */
+
+                schedule();
+
+                TS_LOCK_RESUME;
+
+                /* Since we hold the lock, no other task will change
+                 * ->owner. We can thus check it without acquiring the spin
+                 * lock. */
+                BUG_ON(sem->owner != t);
+        } else {
+                /* it's ours now */
+                sem->owner = t;
+
+                spin_unlock_irqrestore(&sem->wait.lock, flags);
+        }
+
+        return 0;
+}
+
+int gsnedfsplit_fmlp_unlock(struct litmus_lock* l)
+{
+        struct task_struct *t = current, *next;
+        struct fmlp_semaphore *sem = fmlp_from_lock(l);
+        unsigned long flags;
+        int err = 0;
+
+        spin_lock_irqsave(&sem->wait.lock, flags);
+
+        if (sem->owner != t) {
+                err = -EINVAL;
+                goto out;
+        }
+
+        /* check if there are jobs waiting for this resource */
+        next = __waitqueue_remove_first(&sem->wait);
+        if (next) {
+                /* next becomes the resouce holder */
+                sem->owner = next;
+                TRACE_CUR("lock ownership passed to %s/%d\n", next->comm, next->pid);
+
+                /* determine new hp_waiter if necessary */
+                if (next == sem->hp_waiter) {
+                        TRACE_TASK(next, "was highest-prio waiter\n");
+                        /* next has the highest priority --- it doesn't need to
+                         * inherit.  However, we need to make sure that the
+                         * next-highest priority in the queue is reflected in
+                         * hp_waiter. */
+                        sem->hp_waiter = find_hp_waiter(sem, next);
+                        if (sem->hp_waiter)
+                                TRACE_TASK(sem->hp_waiter, "is new highest-prio waiter\n");
+                        else
+                                TRACE("no further waiters\n");
+                } else {
+                        /* Well, if next is not the highest-priority waiter,
+                         * then it ought to inherit the highest-priority
+                         * waiter's priority. */
+                        set_priority_inheritance(next, sem->hp_waiter);
+                }
+
+                /* wake up next */
+                wake_up_process(next);
+        } else
+                /* becomes available */
+                sem->owner = NULL;
+
+        /* we lose the benefit of priority inheritance (if any) */
+        if (tsk_rt(t)->inh_task)
+                clear_priority_inheritance(t);
+
+out:
+        spin_unlock_irqrestore(&sem->wait.lock, flags);
+
+        return err;
+}
+
+int gsnedfsplit_fmlp_close(struct litmus_lock* l)
+{
+        struct task_struct *t = current;
+        struct fmlp_semaphore *sem = fmlp_from_lock(l);
+        unsigned long flags;
+
+        int owner;
+
+        spin_lock_irqsave(&sem->wait.lock, flags);
+
+        owner = sem->owner == t;
+
+        spin_unlock_irqrestore(&sem->wait.lock, flags);
+
+        if (owner)
+                gsnedfsplit_fmlp_unlock(l);
+
+        return 0;
+}
+
+void gsnedfsplit_fmlp_free(struct litmus_lock* lock)
+{
+        kfree(fmlp_from_lock(lock));
+}
+
+static struct litmus_lock_ops gsnedfsplit_fmlp_lock_ops = {
+        .close  = gsnedfsplit_fmlp_close,
+        .lock   = gsnedfsplit_fmlp_lock,
+        .unlock = gsnedfsplit_fmlp_unlock,
+        .deallocate = gsnedfsplit_fmlp_free,
+};
+
+static struct litmus_lock* gsnedfsplit_new_fmlp(void)
+{
+        struct fmlp_semaphore* sem;
+
+        sem = kmalloc(sizeof(*sem), GFP_KERNEL);
+        if (!sem)
+                return NULL;
+
+        sem->owner   = NULL;
+        sem->hp_waiter = NULL;
+        init_waitqueue_head(&sem->wait);
+        sem->litmus_lock.ops = &gsnedfsplit_fmlp_lock_ops;
+
+        return &sem->litmus_lock;
+}
+
+/* **** lock constructor **** */
+
+
+static long gsnedfsplit_allocate_lock(struct litmus_lock **lock, int type,
+                                 void* __user unused)
+{
+        int err = -ENXIO;
+
+        /* GSN-EDF-split currently only supports the FMLP for global resources. */
+        switch (type) {
+
+        case FMLP_SEM:
+                /* Flexible Multiprocessor Locking Protocol */
+                *lock = gsnedfsplit_new_fmlp();
+                if (*lock)
+                        err = 0;
+                else
+                        err = -ENOMEM;
+                break;
+
+        };
+
+        return err;
+}
+
+#endif
+
+
+static long gsnedfsplit_activate_plugin(void)
+{
+        int cpu;
+        cpu_entry_t *entry;
+
+        bheap_init(&gsnedfsplit_cpu_heap);
+#ifdef CONFIG_RELEASE_MASTER
+        gsnedfsplit.release_master = atomic_read(&release_master_cpu);
+#endif
+
+        for_each_online_cpu(cpu) {
+                entry = &per_cpu(gsnedfsplit_cpu_entries, cpu);
+                bheap_node_init(&entry->hn, entry);
+                entry->linked    = NULL;
+                entry->scheduled = NULL;
+#ifdef CONFIG_RELEASE_MASTER
+                if (cpu != gsnedfsplit.release_master) {
+#endif
+                        TRACE("GSN-EDF-split: Initializing CPU #%d.\n", cpu);
+                        update_cpu_position(entry);
+#ifdef CONFIG_RELEASE_MASTER
+                } else {
+                        TRACE("GSN-EDF-split: CPU %d is release master.\n", cpu);
+                }
+#endif
+        }
+        return 0;
+}
+
+/*      Plugin object   */
+static struct sched_plugin gsn_edf_plugin __cacheline_aligned_in_smp = {
+        .plugin_name            = "GSN-EDF-split",
+        .finish_switch          = gsnedfsplit_finish_switch,
+        .tick                   = gsnedfsplit_tick,
+        .task_new               = gsnedfsplit_task_new,
+        .complete_job           = complete_job,
+        .task_exit              = gsnedfsplit_task_exit,
+        .schedule               = gsnedfsplit_schedule,
+        .task_wake_up           = gsnedfsplit_task_wake_up,
+        .task_block             = gsnedfsplit_task_block,
+        .admit_task             = gsnedfsplit_admit_task,
+        .activate_plugin        = gsnedfsplit_activate_plugin,
+#ifdef CONFIG_LITMUS_LOCKING
+        .allocate_lock          = gsnedfsplit_allocate_lock,
+#endif
+};
+
+
+static int __init init_gsn_edf(void)
+{
+        int cpu;
+        cpu_entry_t *entry;
+
+        bheap_init(&gsnedfsplit_cpu_heap);
+        /* initialize CPU state */
+        for (cpu = 0; cpu < NR_CPUS; cpu++)  {
+                entry = &per_cpu(gsnedfsplit_cpu_entries, cpu);
+                gsnedfsplit_cpus[cpu] = entry;
+                entry->cpu       = cpu;
+                entry->hn        = &gsnedfsplit_heap_node[cpu];
+                bheap_node_init(&entry->hn, entry);
+        }
+        edf_domain_init(&gsnedfsplit, NULL, gsnedfsplit_release_jobs);
+        return register_sched_plugin(&gsn_edf_plugin);
+}
+
+
+module_init(init_gsn_edf);