1 files changed, 678 insertions, 0 deletions

diff --git a/litmus/sched_mc_edf.c b/litmus/sched_mc_edf.c
new file mode 100644
index 000000000000..bdd518523cff
--- /dev/null
+++ b/litmus/sched_mc_edf.c
@@ -0,0 +1,678 @@
+/*
+ * kernel/sched_mc_edf.c
+ *
+ * Implementation of the MC-EDF scheduler plugin.
+ * Based on kern/sched_part_edf.c and kern/sched_gsn_edf.c.
+ *
+ * Suspensions and non-preemptable sections are supported.
+ * Priority inheritance is not supported.
+ */
+
+#include <linux/percpu.h>
+#include <linux/sched.h>
+#include <linux/list.h>
+#include <linux/spinlock.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/trace.h>
+#include <litmus/budget.h>
+
+typedef struct {
+        int                     curr_cri; // mc-bipasa
+        rt_domain_t             high_domain; // mc-bipasa
+        rt_domain_t             domain;
+        int                     cpu;
+        struct task_struct*     scheduled; /* only RT tasks */
+/*
+ * scheduling lock slock
+ * protects the domain and serializes scheduling decisions
+ */
+#define slock domain.ready_lock
+
+} mcedf_domain_t;
+
+DEFINE_PER_CPU(mcedf_domain_t, mcedf_domains);
+
+#define local_edf               (&__get_cpu_var(mcedf_domains).domain)
+#define local_mcedf             (&__get_cpu_var(mcedf_domains))
+#define remote_edf(cpu)         (&per_cpu(mcedf_domains, cpu).domain)
+#define remote_mcedf(cpu)       (&per_cpu(mcedf_domains, cpu))
+#define task_edf(task)          remote_edf(get_partition(task))
+#define task_mcedf(task)        remote_mcedf(get_partition(task))//mc-bipasa always 0
+
+
+static void mcedf_domain_init(mcedf_domain_t* mcedf,
+                               check_resched_needed_t check,
+                               release_jobs_t release,
+                               int cpu)
+{ 
+        // high >= 1, low = 0
+        mcedf->curr_cri                 = 0; 
+        edf_mc_domain_init(&mcedf->high_domain, check, release, 1); // mc-bipasa
+        edf_mc_domain_init(&mcedf->domain, check, release, 0); // mc-bipasa
+        mcedf->cpu                      = cpu;
+        mcedf->scheduled                = NULL;
+}
+
+static void requeue(struct task_struct* t, rt_domain_t *edf)
+{
+        if (t->state != TASK_RUNNING)
+                TRACE_TASK(t, "requeue: !TASK_RUNNING\n");
+
+        set_rt_flags(t, RT_F_RUNNING);
+        if (is_released(t, litmus_clock()))
+                __add_ready(edf, t);
+        else
+                add_release(edf, t); /* it has got to wait */
+}
+
+/* we assume the lock is being held */
+static void preempt(mcedf_domain_t *mcedf)
+{
+        preempt_if_preemptable(mcedf->scheduled, mcedf->cpu);
+}
+
+#ifdef CONFIG_LITMUS_LOCKING
+
+static void boost_priority(struct task_struct* t)
+{
+        unsigned long           flags;
+        mcedf_domain_t*         mcedf = task_mcedf(t);
+        lt_t                    now;
+
+        raw_spin_lock_irqsave(&mcedf->slock, flags);
+        now = litmus_clock();
+
+        TRACE_TASK(t, "priority boosted at %llu\n", now);
+
+        tsk_rt(t)->priority_boosted = 1;
+        tsk_rt(t)->boost_start_time = now;
+
+        if (mcedf->scheduled != t) {
+                /* holder may be queued: first stop queue changes */
+                raw_spin_lock(&mcedf->domain.release_lock);
+                if (is_queued(t) &&
+                    /* If it is queued, then we need to re-order. */
+                    bheap_decrease(edf_ready_order, tsk_rt(t)->heap_node) &&
+                    /* If we bubbled to the top, then we need to check for preemptions. */
+                    edf_preemption_needed(&mcedf->domain, mcedf->scheduled))
+                                preempt(mcedf);
+                raw_spin_unlock(&mcedf->domain.release_lock);
+        } /* else: nothing to do since the job is not queued while scheduled */
+
+        raw_spin_unlock_irqrestore(&mcedf->slock, flags);
+}
+
+static void unboost_priority(struct task_struct* t)
+{
+        unsigned long           flags;
+        mcedf_domain_t*         mcedf = task_mcedf(t);
+        lt_t                    now;
+
+        raw_spin_lock_irqsave(&mcedf->slock, flags);
+        now = litmus_clock();
+
+        /* assumption: this only happens when the job is scheduled */
+        BUG_ON(mcedf->scheduled != t);
+
+        TRACE_TASK(t, "priority restored at %llu\n", now);
+
+        /* priority boosted jobs must be scheduled */
+        BUG_ON(mcedf->scheduled != t);
+
+        tsk_rt(t)->priority_boosted = 0;
+        tsk_rt(t)->boost_start_time = 0;
+
+        /* check if this changes anything */
+        if (edf_preemption_needed(&mcedf->domain, mcedf->scheduled))
+                preempt(mcedf);
+
+        raw_spin_unlock_irqrestore(&mcedf->slock, flags);
+}
+
+#endif
+
+/* This check is trivial in partioned systems as we only have to consider
+ * the CPU of the partition.
+ */
+static int mcedf_check_resched(rt_domain_t *edf)
+{
+        mcedf_domain_t *mcedf = container_of(edf, mcedf_domain_t, domain);
+
+        /* because this is a callback from rt_domain_t we already hold
+         * the necessary lock for the ready queue
+         */
+        if (edf_preemption_needed(edf, mcedf->scheduled)) {
+                preempt(mcedf);
+                return 1;
+        } else
+                return 0;
+}
+
+static void job_completion(struct task_struct* t, int forced)
+{
+        sched_trace_task_completion(t,forced);
+        TRACE_TASK(t, "job_completion().\n");
+
+        set_rt_flags(t, RT_F_SLEEP);
+
+        prepare_for_next_period(t); // check
+        /*if (forced) {
+          server_release(t);
+        } else {
+          task_release(t);
+        }*/
+}
+
+static void mcedf_tick(struct task_struct *t)
+{
+        mcedf_domain_t *mcedf = local_mcedf;
+
+        /* Check for inconsistency. We don't need the lock for this since
+         * ->scheduled is only changed in schedule, which obviously is not
+         *  executing in parallel on this CPU
+         */
+        BUG_ON(is_realtime(t) && t != mcedf->scheduled);
+
+        if (is_realtime(t) && budget_enforced(t) && budget_exhausted(t)) {
+                if (!is_np(t)) {
+                        litmus_reschedule_local();
+                        TRACE("mcedf_scheduler_tick: "
+                              "%d is preemptable "
+                              " => FORCE_RESCHED\n", t->pid);
+                } else if (is_user_np(t)) {
+                        TRACE("mcedf_scheduler_tick: "
+                              "%d is non-preemptable, "
+                              "preemption delayed.\n", t->pid);
+                        request_exit_np(t);
+                }
+        }
+}
+
+static struct task_struct* mcedf_schedule(struct task_struct * prev)
+{
+        mcedf_domain_t*         mcedf = local_mcedf; 
+        rt_domain_t*            edf  = &mcedf->domain; 
+        struct task_struct*     next;
+
+        int                     out_of_time, sleep, preempt,
+                                np, exists, blocks, resched;
+
+        raw_spin_lock(&mcedf->slock);
+
+        /* sanity checking
+         * differently from gedf, when a task exits (dead)
+         * mcedf->schedule may be null and prev _is_ realtime
+         */
+        BUG_ON(mcedf->scheduled && mcedf->scheduled != prev);
+        BUG_ON(mcedf->scheduled && !is_realtime(prev));
+
+        /* (0) Determine state */
+        exists      = mcedf->scheduled != NULL;
+        blocks      = exists && !is_running(mcedf->scheduled);
+        out_of_time = exists &&
+                                  budget_enforced(mcedf->scheduled) &&
+                                  budget_exhausted(mcedf->scheduled);
+        np          = exists && is_np(mcedf->scheduled);
+        sleep       = exists && get_rt_flags(mcedf->scheduled) == RT_F_SLEEP;
+        preempt     = edf_preemption_needed(edf, prev);
+
+        /* If we need to preempt do so.
+         * The following checks set resched to 1 in case of special
+         * circumstances.
+         */
+        resched = preempt;
+
+        /* If a task blocks we have no choice but to reschedule.
+         */
+        if (blocks)
+                resched = 1;
+
+        /* Request a sys_exit_np() call if we would like to preempt but cannot.
+         * Multiple calls to request_exit_np() don't hurt.
+         */
+        if (np && (out_of_time || preempt || sleep))
+                request_exit_np(mcedf->scheduled);
+
+        // prev code
+        /* Any task that is preemptable and either exhausts its execution
+         * budget or wants to sleep completes. We may have to reschedule after
+         * this.
+         */
+        /*if (!np && (out_of_time || sleep) && !blocks) {
+                job_completion(mcedf->scheduled, !sleep);
+                resched = 1;
+        }*/
+        
+        /* Any task that is premptable and either exhausts its execution
+         * budget or wants to sleep. We have to reschedule after this. 
+         */ 
+        
+        if (!np && (out_of_time || sleep) && !blocks){
+                /* This causes a switch to high criticality*/
+                if (out_of_time){
+                        TRACE_TASK(prev, "Ran out of budget at %llu. Switching to high criticality\n", litmus_clock()); 
+                        mcedf->curr_cri = 1; 
+                        TRACE("After this point no low criticality jobs must be scheduled\n"); 
+                }
+                //if (sleep){   
+                        job_completion(mcedf->scheduled, !sleep); 
+                //}
+                resched = 1; 
+        }
+
+        /* The final scheduling decision. Do we need to switch for some reason?
+         * Switch if we are in RT mode and have no task or if we need to
+         * resched.
+         */
+        next = NULL;
+        if ((!np || blocks) && (resched || !exists)) {
+                /* When preempting a task that does not block, then
+                 * re-insert it into either the ready queue or the
+                 * release queue (if it completed). requeue() picks
+                 * the appropriate queue.
+                 */
+                if (mcedf->scheduled && !blocks)
+                        requeue(mcedf->scheduled, edf);
+                next = __take_ready(edf);
+        } else
+                /* Only override Linux scheduler if we have a real-time task
+                 * scheduled that needs to continue.
+                 */
+                if (exists)
+                        next = prev;
+
+        if (next) {
+                TRACE_TASK(next, "scheduled at %llu\n", litmus_clock());
+                set_rt_flags(next, RT_F_RUNNING);
+        } else {
+                TRACE("becoming idle at %llu\n", litmus_clock());
+        }
+
+        mcedf->scheduled = next;
+        sched_state_task_picked();
+        raw_spin_unlock(&mcedf->slock);
+
+        return next;
+}
+
+
+/*      Prepare a task for running in RT mode
+ */
+static void mcedf_task_new(struct task_struct * t, int on_rq, int running)
+{
+        rt_domain_t*            edf  = task_edf(t);
+        mcedf_domain_t*         mcedf = task_mcedf(t);
+        unsigned long           flags;
+
+        TRACE_TASK(t, "mc edf: task new, cpu = %d\n",
+                   t->rt_param.task_params.cpu);
+        TRACE_TASK(t, "mc edf: task new, criticality = %d\n", 
+                   t->rt_param.task_params.cri); // mc-bipasa
+
+        /* setup job parameters */
+        release_at(t, litmus_clock());
+
+        /* The task should be running in the queue, otherwise signal
+         * code will try to wake it up with fatal consequences.
+         */
+        raw_spin_lock_irqsave(&mcedf->slock, flags);
+        if (running) {
+                /* there shouldn't be anything else running at the time */
+                BUG_ON(mcedf->scheduled);
+                mcedf->scheduled = t;
+        } else {
+                requeue(t, edf);
+                /* maybe we have to reschedule */
+                preempt(mcedf);
+        }
+        raw_spin_unlock_irqrestore(&mcedf->slock, flags);
+}
+
+static void mcedf_task_wake_up(struct task_struct *task)
+{
+        unsigned long           flags;
+        mcedf_domain_t*         mcedf = task_mcedf(task);
+        rt_domain_t*            edf  = task_edf(task);
+        lt_t                    now;
+
+        TRACE_TASK(task, "wake_up at %llu\n", litmus_clock());
+        raw_spin_lock_irqsave(&mcedf->slock, flags);
+        BUG_ON(is_queued(task));
+        now = litmus_clock();
+        if (is_tardy(task, now)
+#ifdef CONFIG_LITMUS_LOCKING
+        /* 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.
+         */
+            && !is_priority_boosted(task)
+#endif
+                ) {
+                /* new sporadic release */
+                release_at(task, now);
+                sched_trace_task_release(task);
+        }
+
+        /* Only add to ready queue if it is not the currently-scheduled
+         * task. This could be the case if a task was woken up concurrently
+         * on a remote CPU before the executing CPU got around to actually
+         * de-scheduling the task, i.e., wake_up() raced with schedule()
+         * and won.
+         */
+        if (mcedf->scheduled != task)
+                requeue(task, edf);
+
+        raw_spin_unlock_irqrestore(&mcedf->slock, flags);
+        TRACE_TASK(task, "wake up done\n");
+}
+
+static void mcedf_task_block(struct task_struct *t)
+{
+        /* only running tasks can block, thus t is in no queue */
+        TRACE_TASK(t, "block at %llu, state=%d\n", litmus_clock(), t->state);
+
+        BUG_ON(!is_realtime(t));
+        BUG_ON(is_queued(t));
+}
+
+static void mcedf_task_exit(struct task_struct * t)
+{
+        unsigned long flags;
+        mcedf_domain_t*         mcedf = task_mcedf(t);
+        rt_domain_t*            edf;
+
+        raw_spin_lock_irqsave(&mcedf->slock, flags);
+        if (is_queued(t)) {
+                /* dequeue */
+                edf  = task_edf(t);
+                remove(edf, t);
+        }
+        if (mcedf->scheduled == t)
+                mcedf->scheduled = NULL;
+
+        TRACE_TASK(t, "RIP, now reschedule\n");
+
+        preempt(mcedf);
+        raw_spin_unlock_irqrestore(&mcedf->slock, flags);
+}
+// I don't need to following for my work
+#ifdef CONFIG_LITMUS_LOCKING
+
+#include <litmus/fdso.h>
+#include <litmus/srp.h>
+
+/* ******************** SRP support ************************ */
+
+static unsigned int mcedf_get_srp_prio(struct task_struct* t)
+{
+        /* assumes implicit deadlines */
+        return get_rt_period(t);
+}
+
+/* ******************** FMLP support ********************** */
+
+/* struct for semaphore with priority inheritance */
+struct fmlp_semaphore {
+        struct litmus_lock litmus_lock;
+
+        /* current resource holder */
+        struct task_struct *owner;
+
+        /* 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);
+}
+int mcedf_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);
+
+                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;
+
+                /* mark the task as priority-boosted. */
+                boost_priority(t);
+
+                spin_unlock_irqrestore(&sem->wait.lock, flags);
+        }
+
+        return 0;
+}
+
+int mcedf_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;
+        }
+
+        /* we lose the benefit of priority boosting */
+
+        unboost_priority(t);
+
+        /* check if there are jobs waiting for this resource */
+        next = __waitqueue_remove_first(&sem->wait);
+        if (next) {
+                /* boost next job */
+                boost_priority(next);
+
+                /* next becomes the resouce holder */
+                sem->owner = next;
+
+                /* wake up next */
+                wake_up_process(next);
+        } else
+                /* resource becomes available */
+                sem->owner = NULL;
+
+out:
+        spin_unlock_irqrestore(&sem->wait.lock, flags);
+        return err;
+}
+
+int mcedf_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)
+                mcedf_fmlp_unlock(l);
+
+        return 0;
+}
+
+void mcedf_fmlp_free(struct litmus_lock* lock)
+{
+        kfree(fmlp_from_lock(lock));
+}
+
+static struct litmus_lock_ops mcedf_fmlp_lock_ops = {
+        .close  = mcedf_fmlp_close,
+        .lock   = mcedf_fmlp_lock,
+        .unlock = mcedf_fmlp_unlock,
+        .deallocate = mcedf_fmlp_free,
+};
+
+static struct litmus_lock* mcedf_new_fmlp(void)
+{
+        struct fmlp_semaphore* sem;
+
+        sem = kmalloc(sizeof(*sem), GFP_KERNEL);
+        if (!sem)
+                return NULL;
+
+        sem->owner   = NULL;
+        init_waitqueue_head(&sem->wait);
+        sem->litmus_lock.ops = &mcedf_fmlp_lock_ops;
+
+        return &sem->litmus_lock;
+}
+
+/* **** lock constructor **** */
+
+
+static long mcedf_allocate_lock(struct litmus_lock **lock, int type,
+                                 void* __user unused)
+{
+        int err = -ENXIO;
+        struct srp_semaphore* srp;
+
+        /* MC-EDF currently supports the SRP for local resources and the FMLP
+         * for global resources. */
+        switch (type) {
+        case FMLP_SEM:
+                /* Flexible Multiprocessor Locking Protocol */
+                *lock = mcedf_new_fmlp();
+                if (*lock)
+                        err = 0;
+                else
+                        err = -ENOMEM;
+                break;
+
+        case SRP_SEM:
+                /* Baker's Stack Resource Policy */
+                srp = allocate_srp_semaphore();
+                if (srp) {
+                        *lock = &srp->litmus_lock;
+                        err = 0;
+                } else
+                        err = -ENOMEM;
+                break;
+        };
+
+        return err;
+}
+
+#endif
+
+
+static long mcedf_activate_plugin(void)
+{
+#ifdef CONFIG_RELEASE_MASTER
+        int cpu;
+
+        for_each_online_cpu(cpu) {
+                remote_edf(cpu)->release_master = atomic_read(&release_master_cpu);
+        }
+#endif
+
+#ifdef CONFIG_LITMUS_LOCKING
+        get_srp_prio = mcedf_get_srp_prio;
+#endif
+
+        return 0;
+}
+
+static long mcedf_admit_task(struct task_struct* tsk)
+{
+        if (task_cpu(tsk) == tsk->rt_param.task_params.cpu
+#ifdef CONFIG_RELEASE_MASTER
+            /* don't allow tasks on release master CPU */
+             && task_cpu(tsk) != remote_edf(task_cpu(tsk))->release_master
+#endif
+                )
+                return 0;
+        else
+                return -EINVAL;
+}
+
+/*      Plugin object   */
+static struct sched_plugin mc_edf_plugin __cacheline_aligned_in_smp = {
+        .plugin_name            = "MC-EDF",
+        .tick                   = mcedf_tick,
+        .task_new               = mcedf_task_new,
+        .complete_job           = complete_job,
+        .task_exit              = mcedf_task_exit,
+        .schedule               = mcedf_schedule,
+        .task_wake_up           = mcedf_task_wake_up,
+        .task_block             = mcedf_task_block,
+        .admit_task             = mcedf_admit_task,
+        .activate_plugin        = mcedf_activate_plugin,
+#ifdef CONFIG_LITMUS_LOCKING
+        .allocate_lock          = mcedf_allocate_lock,
+#endif
+};
+
+
+static int __init init_mc_edf(void)
+{
+        int i;
+
+        /* We do not really want to support cpu hotplug, do we? ;)
+         * However, if we are so crazy to do so,
+         * we cannot use num_online_cpu()
+         */
+        printk("Hello Mixed Criticality\n");
+        for (i = 0; i < num_online_cpus(); i++) {
+                mcedf_domain_init(remote_mcedf(i),
+                                   mcedf_check_resched,
+                                   NULL, i);
+        }
+        return register_sched_plugin(&mc_edf_plugin);
+}
+
+module_init(init_mc_edf);
+