8 files changed, 483 insertions, 348 deletions

diff --git a/arch/x86/include/asm/processor.h b/arch/x86/include/asm/processor.h
index c9786480f0fe..e75daac64962 100644
--- a/arch/x86/include/asm/processor.h
+++ b/arch/x86/include/asm/processor.h
@@ -172,6 +172,8 @@ extern void print_cpu_info(struct cpuinfo_x86 *);
 extern void init_scattered_cpuid_features(struct cpuinfo_x86 *c);
 extern unsigned int init_intel_cacheinfo(struct cpuinfo_x86 *c);
 extern unsigned short num_cache_leaves;
+extern int get_shared_cpu_map(cpumask_var_t mask,
+                               unsigned int cpu, int index);
 
 extern void detect_extended_topology(struct cpuinfo_x86 *c);
 extern void detect_ht(struct cpuinfo_x86 *c);
diff --git a/arch/x86/kernel/cpu/intel_cacheinfo.c b/arch/x86/kernel/cpu/intel_cacheinfo.c
index 804c40e2bc3e..3167c3d72596 100644
--- a/arch/x86/kernel/cpu/intel_cacheinfo.c
+++ b/arch/x86/kernel/cpu/intel_cacheinfo.c
@@ -515,6 +515,23 @@ unsigned int __cpuinit init_intel_cacheinfo(struct cpuinfo_x86 *c)
 static DEFINE_PER_CPU(struct _cpuid4_info *, cpuid4_info);
 #define CPUID4_INFO_IDX(x, y)   (&((per_cpu(cpuid4_info, x))[y]))
 
+/* returns CPUs that share the index cache with cpu */
+int get_shared_cpu_map(cpumask_var_t mask, unsigned int cpu, int index)
+{
+        int ret = 0;
+        struct _cpuid4_info *this_leaf;
+
+        if (index >= num_cache_leaves) {
+                index = num_cache_leaves - 1;
+                ret = index;
+        }
+
+        this_leaf = CPUID4_INFO_IDX(cpu,index);
+        cpumask_copy(mask, to_cpumask(this_leaf->shared_cpu_map));
+
+        return ret;
+}
+
 #ifdef CONFIG_SMP
 static void __cpuinit cache_shared_cpu_map_setup(unsigned int cpu, int index)
 {
diff --git a/include/litmus/sched_plugin.h b/include/litmus/sched_plugin.h
index 2d856d587041..9c1c9f28ba79 100644
--- a/include/litmus/sched_plugin.h
+++ b/include/litmus/sched_plugin.h
@@ -133,6 +133,9 @@ struct sched_plugin {
 
 extern struct sched_plugin *litmus;
 
+/* cluster size: cache_index = 2 L2, cache_index = 3 L3 */
+extern int cluster_cache_index;
+
 int register_sched_plugin(struct sched_plugin* plugin);
 struct sched_plugin* find_sched_plugin(const char* name);
 int print_sched_plugins(char* buf, int max);
diff --git a/kernel/sched.c b/kernel/sched.c
index 1701eaebb79c..adb5e923cc61 100644
--- a/kernel/sched.c
+++ b/kernel/sched.c
@@ -5520,20 +5520,16 @@ need_resched_nonpreemptible:
                 spin_unlock_irq(&rq->lock);
         }
 
-        TS_SCHED2_START(current);
         sched_trace_task_switch_to(current);
 
         post_schedule(rq);
 
         if (unlikely(reacquire_kernel_lock(current) < 0)) {
-                TS_SCHED2_END(current);
                 goto need_resched_nonpreemptible;
         }
 
         preempt_enable_no_resched();
 
-        TS_SCHED2_END(current);
-
         if (need_resched())
                 goto need_resched;
 
diff --git a/litmus/litmus.c b/litmus/litmus.c
index 3cf7cb9e8a9f..e43596a5104c 100644
--- a/litmus/litmus.c
+++ b/litmus/litmus.c
@@ -24,6 +24,8 @@
 /* Number of RT tasks that exist in the system */
 atomic_t rt_task_count          = ATOMIC_INIT(0);
 static DEFINE_SPINLOCK(task_transition_lock);
+/* synchronize plugin switching */
+atomic_t cannot_use_plugin      = ATOMIC_INIT(0);
 
 /* Give log messages sequential IDs. */
 atomic_t __log_seq_no = ATOMIC_INIT(0);
@@ -369,13 +371,17 @@ void litmus_exit_task(struct task_struct* tsk)
         }
 }
 
+/* IPI callback to synchronize plugin switching */
+static void synch_on_plugin_switch(void* info)
+{
+        while (atomic_read(&cannot_use_plugin))
+                cpu_relax();
+}
+
 /* Switching a plugin in use is tricky.
  * We must watch out that no real-time tasks exists
  * (and that none is created in parallel) and that the plugin is not
  * currently in use on any processor (in theory).
- *
- * For now, we don't enforce the second part since it is unlikely to cause
- * any trouble by itself as long as we don't unload modules.
  */
 int switch_sched_plugin(struct sched_plugin* plugin)
 {
@@ -384,6 +390,11 @@ int switch_sched_plugin(struct sched_plugin* plugin)
 
         BUG_ON(!plugin);
 
+        /* forbid other cpus to use the plugin */
+        atomic_set(&cannot_use_plugin, 1);
+        /* send IPI to force other CPUs to synch with us */
+        smp_call_function(synch_on_plugin_switch, NULL, 0);
+
         /* stop task transitions */
         spin_lock_irqsave(&task_transition_lock, flags);
 
@@ -404,6 +415,7 @@ int switch_sched_plugin(struct sched_plugin* plugin)
                 ret = -EBUSY;
 out:
         spin_unlock_irqrestore(&task_transition_lock, flags);
+        atomic_set(&cannot_use_plugin, 0);
         return ret;
 }
 
@@ -554,6 +566,55 @@ static int proc_write_curr(struct file *file,
         return len;
 }
 
+static int proc_read_cluster_size(char *page, char **start,
+                          off_t off, int count,
+                          int *eof, void *data)
+{
+        int len;
+        if (cluster_cache_index == 2)
+                len = snprintf(page, PAGE_SIZE, "L2\n");
+        else if (cluster_cache_index == 3)
+                len = snprintf(page, PAGE_SIZE, "L3\n");
+        else /* (cluster_cache_index == 1) */
+                len = snprintf(page, PAGE_SIZE, "L1\n");
+
+        return len;
+}
+
+static int proc_write_cluster_size(struct file *file,
+                           const char *buffer,
+                           unsigned long count,
+                           void *data)
+{
+        int len;
+        /* L2, L3 */
+        char cache_name[33];
+
+        if(count > 32)
+                len = 32;
+        else
+                len = count;
+
+        if(copy_from_user(cache_name, buffer, len))
+                return -EFAULT;
+
+        cache_name[len] = '\0';
+        /* chomp name */
+        if (len > 1 && cache_name[len - 1] == '\n')
+                cache_name[len - 1] = '\0';
+
+        /* do a quick and dirty comparison to find the cluster size */
+        if (!strcmp(cache_name, "L2"))
+                cluster_cache_index = 2;
+        else if (!strcmp(cache_name, "L3"))
+                cluster_cache_index = 3;
+        else if (!strcmp(cache_name, "L1"))
+                cluster_cache_index = 1;
+        else
+                printk(KERN_INFO "Cluster '%s' is unknown.\n", cache_name);
+
+        return len;
+}
 
 static int proc_read_release_master(char *page, char **start,
                                     off_t off, int count,
@@ -609,6 +670,7 @@ static struct proc_dir_entry *litmus_dir = NULL,
         *curr_file = NULL,
         *stat_file = NULL,
         *plugs_file = NULL,
+        *clus_cache_idx_file = NULL,
         *release_master_file = NULL;
 
 static int __init init_litmus_proc(void)
@@ -639,6 +701,16 @@ static int __init init_litmus_proc(void)
         release_master_file->read_proc = proc_read_release_master;
         release_master_file->write_proc  = proc_write_release_master;
 
+        clus_cache_idx_file = create_proc_entry("cluster_cache",
+                                                0644, litmus_dir);
+        if (!clus_cache_idx_file) {
+                printk(KERN_ERR "Could not allocate cluster_cache "
+                       "procfs entry.\n");
+                return -ENOMEM;
+        }
+        clus_cache_idx_file->read_proc = proc_read_cluster_size;
+        clus_cache_idx_file->write_proc = proc_write_cluster_size;
+
         stat_file = create_proc_read_entry("stats", 0444, litmus_dir,
                                            proc_read_stats, NULL);
 
@@ -656,6 +728,10 @@ static void exit_litmus_proc(void)
                 remove_proc_entry("stats", litmus_dir);
         if (curr_file)
                 remove_proc_entry("active_plugin", litmus_dir);
+        if (clus_cache_idx_file)
+                remove_proc_entry("cluster_cache", litmus_dir);
+        if (release_master_file)
+                remove_proc_entry("release_master", litmus_dir);
         if (litmus_dir)
                 remove_proc_entry("litmus", NULL);
 }
diff --git a/litmus/rt_domain.c b/litmus/rt_domain.c
index 0ed6d5cbbfc5..609ff0f82abb 100644
--- a/litmus/rt_domain.c
+++ b/litmus/rt_domain.c
@@ -301,6 +301,10 @@ void __add_release(rt_domain_t* rt, struct task_struct *task)
         task->rt_param.domain = rt;
 
         /* start release timer */
+        TS_SCHED2_START(task);
+
         arm_release_timer(rt);
+
+        TS_SCHED2_END(task);
 }
 
diff --git a/litmus/sched_cedf.c b/litmus/sched_cedf.c
index d0767ce9e178..da44b451c9ad 100644
--- a/litmus/sched_cedf.c
+++ b/litmus/sched_cedf.c
@@ -1,9 +1,20 @@
 /*
- * kernel/sched_cedf.c
+ * litmus/sched_cedf.c
  *
- * Implementation of the Clustered EDF (C-EDF) scheduling algorithm.
- * Linking is included so that support for synchronization (e.g., through
- * the implementation of a "CSN-EDF" algorithm) can be added later if desired.
+ * 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
+ *
+ *   For details on functions, take a look at sched_gsn_edf.c
  *
  * This version uses the simple approach and serializes all scheduling
  * decisions by the use of a queue lock. This is probably not the
@@ -13,92 +24,36 @@
 #include <linux/spinlock.h>
 #include <linux/percpu.h>
 #include <linux/sched.h>
-#include <linux/list.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/bheap.h>
 
 #include <linux/module.h>
 
-/* Overview of C-EDF operations.
- *
- * link_task_to_cpu(T, cpu)     - Low-level operation to update the linkage
- *                                structure (NOT the actually scheduled
- *                                task). If there is another linked task To
- *                                already it will set To->linked_on = NO_CPU
- *                                (thereby removing its association with this
- *                                CPU). However, it will not requeue the
- *                                previously linked task (if any). It will set
- *                                T's state to RT_F_RUNNING and check whether
- *                                it is already running somewhere else. If T
- *                                is scheduled somewhere else it will link
- *                                it to that CPU instead (and pull the linked
- *                                task to cpu). T may be NULL.
- *
- * unlink(T)                    - Unlink removes T from all scheduler data
- *                                structures. If it is linked to some CPU it
- *                                will link NULL to that CPU. If it is
- *                                currently queued in the cedf queue for
- *                                a partition, it will be removed from
- *                                the rt_domain. It is safe to call
- *                                unlink(T) if T is not linked. T may not
- *                                be NULL.
- *
- * requeue(T)                   - Requeue will insert T into the appropriate
- *                                queue. If the system is in real-time mode and
- *                                the T is released already, it will go into the
- *                                ready queue. If the system is not in
- *                                real-time mode is T, then T will go into the
- *                                release queue. If T's release time is in the
- *                                future, it will go into the release
- *                                queue. That means that T's release time/job
- *                                no/etc. has to be updated before requeue(T) is
- *                                called. It is not safe to call requeue(T)
- *                                when T is already queued. T may not be NULL.
- *
- * cedf_job_arrival(T)          - This is the catch-all function when T enters
- *                                the system after either a suspension or at a
- *                                job release. It will queue T (which means it
- *                                is not safe to call cedf_job_arrival(T) if
- *                                T is already queued) and then check whether a
- *                                preemption is necessary. If a preemption is
- *                                necessary it will update the linkage
- *                                accordingly and cause scheduled to be called
- *                                (either with an IPI or need_resched). It is
- *                                safe to call cedf_job_arrival(T) if T's
- *                                next job has not been actually released yet
- *                                (release time in the future). T will be put
- *                                on the release queue in that case.
- *
- * job_completion(T)            - Take care of everything that needs to be done
- *                                to prepare T for its next release and place
- *                                it in the right queue with
- *                                cedf_job_arrival().
- *
- *
- * When we now that T is linked to CPU then link_task_to_cpu(NULL, CPU) is
- * equivalent to unlink(T). Note that if you unlink a task from a CPU none of
- * the functions will automatically propagate pending task from the ready queue
- * to a linked task. This is the job of the calling function ( by means of
- * __take_ready).
- */
+/* forward declaration... a funny thing with C ;) */
+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 */
-        struct list_head        list;
         atomic_t                will_schedule;  /* prevent unneeded IPIs */
+        struct bheap_node*      hn;
 } cpu_entry_t;
-DEFINE_PER_CPU(cpu_entry_t, cedf_cpu_entries);
 
-cpu_entry_t* *cedf_cpu_entries_array;
+/* 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))
@@ -107,75 +62,73 @@ cpu_entry_t* *cedf_cpu_entries_array;
 #define test_will_schedule(cpu) \
         (atomic_read(&per_cpu(cedf_cpu_entries, cpu).will_schedule))
 
-/* Cluster size -- currently four. This is a variable to allow for
- * the possibility of changing the cluster size online in the future.
+/*
+ * 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
  */
-int cluster_size = 4;
-
-int do_cleanup = 1;
-
-typedef struct {
-        rt_domain_t             domain;
-        int                     first_cpu;
-        int                     last_cpu;
-
+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 list_head        cedf_cpu_queue;
-
-        /* per-partition spinlock: protects the domain and
-         * serializes scheduling decisions
-         */
-#define slock domain.ready_lock
+        struct bheap_node *heap_node;
+        struct bheap      cpu_heap;
+        /* lock for this cluster */
+#define lock domain.ready_lock
 } cedf_domain_t;
 
-DEFINE_PER_CPU(cedf_domain_t*, cedf_domains) = NULL;
-
-cedf_domain_t* *cedf_domains_array;
+/* 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))
 
-/* These are defined similarly to partitioning, except that a
- * tasks partition is any cpu of the cluster to which it
- * is assigned, typically the lowest-numbered cpu.
+/* 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 local_edf               (&__get_cpu_var(cedf_domains)->domain)
-#define local_cedf              __get_cpu_var(cedf_domains)
-#define remote_edf(cpu)         (&per_cpu(cedf_domains, cpu)->domain)
-#define remote_cedf(cpu)        per_cpu(cedf_domains, cpu)
-#define task_edf(task)          remote_edf(get_partition(task))
-#define task_cedf(task)         remote_cedf(get_partition(task))
+#define VERBOSE_INIT
+
+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 cedf lock.
- *
- *                                              This really should be a heap.
  */
 static void update_cpu_position(cpu_entry_t *entry)
 {
-        cpu_entry_t *other;
-        struct list_head *cedf_cpu_queue =
-                &(remote_cedf(entry->cpu))->cedf_cpu_queue;
-        struct list_head *pos;
+        cedf_domain_t *cluster = entry->cluster;
 
-        BUG_ON(!cedf_cpu_queue);
+        if (likely(bheap_node_in_heap(entry->hn)))
+                bheap_delete(cpu_lower_prio,
+                                &cluster->cpu_heap,
+                                entry->hn);
 
-        if (likely(in_list(&entry->list)))
-                list_del(&entry->list);
-        /* if we do not execute real-time jobs we just move
-         * to the end of the queue
-         */
-        if (entry->linked) {
-                list_for_each(pos, cedf_cpu_queue) {
-                        other = list_entry(pos, cpu_entry_t, list);
-                        if (edf_higher_prio(entry->linked, other->linked)) {
-                                __list_add(&entry->list, pos->prev, pos);
-                                return;
-                        }
-                }
-        }
-        /* if we get this far we have the lowest priority job */
-        list_add_tail(&entry->list, cedf_cpu_queue);
+        bheap_insert(cpu_lower_prio, &cluster->cpu_heap, entry->hn);
+}
+
+/* caller must hold cedf lock */
+static cpu_entry_t* lowest_prio_cpu(cedf_domain_t *cluster)
+{
+        struct bheap_node* hn;
+        hn = bheap_peek(cpu_lower_prio, &cluster->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.
@@ -189,9 +142,6 @@ static noinline void link_task_to_cpu(struct task_struct* linked,
 
         BUG_ON(linked && !is_realtime(linked));
 
-        /* Cannot link task to a CPU that doesn't belong to its partition... */
-        BUG_ON(linked && remote_cedf(entry->cpu) != task_cedf(linked));
-
         /* Currently linked task is set to be unlinked. */
         if (entry->linked) {
                 entry->linked->rt_param.linked_on = NO_CPU;
@@ -213,6 +163,9 @@ static noinline void link_task_to_cpu(struct task_struct* linked,
                          * 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;
@@ -224,13 +177,12 @@ static noinline void link_task_to_cpu(struct task_struct* linked,
                         linked->rt_param.linked_on = entry->cpu;
         }
         entry->linked = linked;
-
-        if (entry->linked)
-                TRACE_TASK(entry->linked, "linked to CPU %d, state:%d\n",
-                           entry->cpu, entry->linked->state);
+#ifdef WANT_ALL_SCHED_EVENTS
+        if (linked)
+                TRACE_TASK(linked, "linked to %d.\n", entry->cpu);
         else
-                TRACE("NULL linked to CPU %d\n", entry->cpu);
-
+                TRACE("NULL linked to %d.\n", entry->cpu);
+#endif
         update_cpu_position(entry);
 }
 
@@ -246,6 +198,7 @@ static noinline void unlink(struct task_struct* t)
                 return;
         }
 
+
         if (t->rt_param.linked_on != NO_CPU) {
                 /* unlink */
                 entry = &per_cpu(cedf_cpu_entries, t->rt_param.linked_on);
@@ -258,95 +211,105 @@ static noinline void unlink(struct task_struct* t)
                  * 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_edf(t), t);
+                remove(&(task_cpu_cluster(t))->domain, t);
         }
 }
 
 
 /* preempt - force a CPU to reschedule
  */
-static noinline void preempt(cpu_entry_t *entry)
+static void preempt(cpu_entry_t *entry)
 {
         preempt_if_preemptable(entry->scheduled, entry->cpu);
 }
 
-/* requeue - Put an unlinked task into c-edf domain.
+/* requeue - Put an unlinked task into gsn-edf domain.
  *           Caller must hold cedf_lock.
  */
 static noinline void requeue(struct task_struct* task)
 {
-        cedf_domain_t* cedf;
-        rt_domain_t* edf;
-
+        cedf_domain_t *cluster = task_cpu_cluster(task);
         BUG_ON(!task);
-        /* sanity check rt_list before insertion */
+        /* sanity check before insertion */
         BUG_ON(is_queued(task));
 
-        /* Get correct real-time domain. */
-        cedf = task_cedf(task);
-        edf = &cedf->domain;
-
         if (is_released(task, litmus_clock()))
-                __add_ready(edf, task);
+                __add_ready(&cluster->domain, task);
         else {
                 /* it has got to wait */
-                add_release(edf, task);
+                add_release(&cluster->domain, task);
         }
 }
 
-static void check_for_preemptions(cedf_domain_t* cedf)
+/* check for any necessary preemptions */
+static void check_for_preemptions(cedf_domain_t *cluster)
 {
-        cpu_entry_t *last;
         struct task_struct *task;
-        struct list_head *cedf_cpu_queue;
-        cedf_cpu_queue = &cedf->cedf_cpu_queue;
+        cpu_entry_t* last;
 
-        for(last = list_entry(cedf_cpu_queue->prev, cpu_entry_t, list);
-            edf_preemption_needed(&cedf->domain, last->linked);
-            last = list_entry(cedf_cpu_queue->prev, cpu_entry_t, list)) {
+        for(last = lowest_prio_cpu(cluster);
+            edf_preemption_needed(&cluster->domain, last->linked);
+            last = lowest_prio_cpu(cluster)) {
                 /* preemption necessary */
-                task = __take_ready(&cedf->domain);
-                TRACE("check_for_preemptions: task %d linked to %d, state:%d\n",
-                      task->pid, last->cpu, task->state);
+                task = __take_ready(&cluster->domain);
+                TRACE("check_for_preemptions: attempting to link task %d to %d\n",
+                      task->pid, last->cpu);
                 if (last->linked)
                         requeue(last->linked);
                 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* cedf;
-        rt_domain_t* edf;
-
+        cedf_domain_t *cluster = task_cpu_cluster(task);
         BUG_ON(!task);
 
-        /* Get correct real-time domain. */
-        cedf = task_cedf(task);
-        edf = &cedf->domain;
-
-        /* first queue arriving job */
         requeue(task);
-
-        /* then check for any necessary preemptions */
-        check_for_preemptions(cedf);
+        check_for_preemptions(cluster);
 }
 
-/* check for current job releases */
 static void cedf_release_jobs(rt_domain_t* rt, struct bheap* tasks)
 {
-        cedf_domain_t*          cedf = container_of(rt, cedf_domain_t, domain);
-        unsigned long           flags;
+        cedf_domain_t* cluster = container_of(rt, cedf_domain_t, domain);
+        unsigned long flags;
+
+        spin_lock_irqsave(&cluster->lock, flags);
+
+        __merge_ready(&cluster->domain, tasks);
+        check_for_preemptions(cluster);
+
+        spin_unlock_irqrestore(&cluster->lock, flags);
+}
+
+/* caller holds cedf_lock */
+static noinline void job_completion(struct task_struct *t, int forced)
+{
+        BUG_ON(!t);
 
-        spin_lock_irqsave(&cedf->slock, flags);
+        sched_trace_task_completion(t, forced);
+
+        TRACE_TASK(t, "job_completion().\n");
 
-        __merge_ready(&cedf->domain, tasks);
-        check_for_preemptions(cedf);
-        spin_unlock_irqrestore(&cedf->slock, flags);
+        /* 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
@@ -357,8 +320,6 @@ static void cedf_release_jobs(rt_domain_t* rt, struct bheap* tasks)
  */
 static void cedf_tick(struct task_struct* t)
 {
-        BUG_ON(!t);
-
         if (is_realtime(t) && budget_exhausted(t)) {
                 if (!is_np(t)) {
                         /* np tasks will be preempted when they become
@@ -367,38 +328,17 @@ static void cedf_tick(struct task_struct* t)
                         set_tsk_need_resched(t);
                         set_will_schedule();
                         TRACE("cedf_scheduler_tick: "
-                              "%d is preemptable (state:%d) "
-                              " => FORCE_RESCHED\n", t->pid, t->state);
-                } else if(is_user_np(t)) {
+                              "%d is preemptable "
+                              " => FORCE_RESCHED\n", t->pid);
+                } else if (is_user_np(t)) {
                         TRACE("cedf_scheduler_tick: "
-                              "%d is non-preemptable (state:%d), "
-                              "preemption delayed.\n", t->pid, t->state);
+                              "%d is non-preemptable, "
+                              "preemption delayed.\n", t->pid);
                         request_exit_np(t);
                 }
         }
 }
 
-/* caller holds cedf_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(). [state:%d]\n", t->state);
-
-        /* set flags */
-        set_rt_flags(t, RT_F_SLEEP);
-        /* prepare for next period */
-        prepare_for_next_period(t);
-        /* unlink */
-        unlink(t);
-        /* requeue
-         * But don't requeue a blocking task. */
-        if (is_running(t))
-                cedf_job_arrival(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
@@ -422,22 +362,12 @@ static noinline void job_completion(struct task_struct *t, int forced)
  */
 static struct task_struct* cedf_schedule(struct task_struct * prev)
 {
-        cedf_domain_t*          cedf = local_cedf;
-        rt_domain_t*            edf  = &cedf->domain;
-        cpu_entry_t*            entry = &__get_cpu_var(cedf_cpu_entries);
-        int                     out_of_time, sleep, preempt, np,
-                                exists, blocks;
-        struct task_struct*     next = NULL;
-
-        BUG_ON(!prev);
-        BUG_ON(!cedf);
-        BUG_ON(!edf);
-        BUG_ON(!entry);
-        BUG_ON(cedf != remote_cedf(entry->cpu));
-        BUG_ON(is_realtime(prev) && cedf != task_cedf(prev));
-
-        /* Will be released in finish_switch. */
-        spin_lock(&cedf->slock);
+        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;
+
+        spin_lock(&cluster->lock);
         clear_will_schedule();
 
         /* sanity checking */
@@ -453,6 +383,21 @@ static struct task_struct* cedf_schedule(struct task_struct * prev)
         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)
@@ -470,8 +415,8 @@ static struct task_struct* cedf_schedule(struct task_struct * prev)
 
         /* 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 blocks (can't have timers
-         * running for blocked jobs). Preemption go first for the same reason.
+         * 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);
@@ -479,10 +424,10 @@ static struct task_struct* cedf_schedule(struct task_struct * prev)
         /* Link pending task if we became unlinked.
          */
         if (!entry->linked)
-                link_task_to_cpu(__take_ready(edf), entry);
+                link_task_to_cpu(__take_ready(&cluster->domain), entry);
 
         /* The final scheduling decision. Do we need to switch for some reason?
-         * If linked different from scheduled select linked as next.
+         * If linked is different from scheduled, then select linked as next.
          */
         if ((!np || blocks) &&
             entry->linked != entry->scheduled) {
@@ -491,76 +436,91 @@ static struct task_struct* cedf_schedule(struct task_struct * prev)
                         entry->linked->rt_param.scheduled_on = entry->cpu;
                         next = entry->linked;
                 }
-                if (entry->scheduled) {
+                if (entry->scheduled) {
                         /* not gonna be scheduled soon */
                         entry->scheduled->rt_param.scheduled_on = NO_CPU;
-                        TRACE_TASK(entry->scheduled, "cedf_schedule: scheduled_on = NO_CPU\n");
+                        TRACE_TASK(entry->scheduled, "scheduled_on = NO_CPU\n");
                 }
         } else
-                /* Only override Linux scheduler if we have real-time task
+                /* Only override Linux scheduler if we have a real-time task
                  * scheduled that needs to continue.
                  */
                 if (exists)
                         next = prev;
 
-        spin_unlock(&cedf->slock);
+        spin_unlock(&cluster->lock);
+
+#ifdef WANT_ALL_SCHED_EVENTS
+        TRACE("cedf_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);
-
-        BUG_ON(!prev);
-        BUG_ON(!entry);
+        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)
+static void cedf_task_new(struct task_struct * t, int on_rq, int running)
 {
         unsigned long           flags;
-        cedf_domain_t*          cedf = task_cedf(t);
         cpu_entry_t*            entry;
+        cedf_domain_t*          cluster;
+
+        TRACE("gsn edf: task new %d\n", t->pid);
+
+        /* the cluster doesn't change even if t is running */
+        cluster = task_cpu_cluster(t);
 
-        BUG_ON(!cedf);
+        spin_lock_irqsave(&cluster->domain.ready_lock, flags);
+
+        /* setup job params */
+        release_at(t, litmus_clock());
 
-        spin_lock_irqsave(&cedf->slock, flags);
         if (running) {
                 entry = &per_cpu(cedf_cpu_entries, task_cpu(t));
-                BUG_ON(!entry);
                 BUG_ON(entry->scheduled);
+
                 entry->scheduled = t;
-                t->rt_param.scheduled_on = task_cpu(t);
-        } else
+                tsk_rt(t)->scheduled_on = task_cpu(t);
+        } else {
                 t->rt_param.scheduled_on = NO_CPU;
-        t->rt_param.linked_on = NO_CPU;
-
-        /* setup job params */
-        release_at(t, litmus_clock());
+        }
+        t->rt_param.linked_on          = NO_CPU;
 
         cedf_job_arrival(t);
-        spin_unlock_irqrestore(&cedf->slock, flags);
+        spin_unlock_irqrestore(&(cluster->domain.ready_lock), flags);
 }
 
-
 static void cedf_task_wake_up(struct task_struct *task)
 {
-        unsigned long           flags;
-        cedf_domain_t*          cedf;
-        lt_t                    now;
+        unsigned long flags;
+        lt_t now;
+        cedf_domain_t *cluster;
 
-        BUG_ON(!task);
+        TRACE_TASK(task, "wake_up at %llu\n", litmus_clock());
 
-        cedf = task_cedf(task);
-        BUG_ON(!cedf);
+        cluster = task_cpu_cluster(task);
 
-        spin_lock_irqsave(&cedf->slock, flags);
+        spin_lock_irqsave(&cluster->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.
@@ -574,48 +534,49 @@ static void cedf_task_wake_up(struct task_struct *task)
                         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 {
+                        if (task->rt.time_slice) {
+                                /* came back in time before deadline
+                                */
+                                set_rt_flags(task, RT_F_RUNNING);
+                        }
+                }
         }
         cedf_job_arrival(task);
-        spin_unlock_irqrestore(&cedf->slock, flags);
+        spin_unlock_irqrestore(&cluster->lock, flags);
 }
 
-
 static void cedf_task_block(struct task_struct *t)
 {
         unsigned long flags;
+        cedf_domain_t *cluster;
 
-        BUG_ON(!t);
+        TRACE_TASK(t, "block at %llu\n", litmus_clock());
 
-        /* unlink if necessary */
-        spin_lock_irqsave(&task_cedf(t)->slock, flags);
+        cluster = task_cpu_cluster(t);
 
-        t->rt_param.scheduled_on = NO_CPU;
+        /* unlink if necessary */
+        spin_lock_irqsave(&cluster->lock, flags);
         unlink(t);
-
-        spin_unlock_irqrestore(&task_cedf(t)->slock, flags);
+        spin_unlock_irqrestore(&cluster->lock, flags);
 
         BUG_ON(!is_realtime(t));
 }
 
+
 static void cedf_task_exit(struct task_struct * t)
 {
         unsigned long flags;
-
-        BUG_ON(!t);
+        cedf_domain_t *cluster = task_cpu_cluster(t);
 
         /* unlink if necessary */
-        spin_lock_irqsave(&task_cedf(t)->slock, flags);
+        spin_lock_irqsave(&cluster->lock, flags);
         unlink(t);
         if (tsk_rt(t)->scheduled_on != NO_CPU) {
-                cedf_cpu_entries_array[tsk_rt(t)->scheduled_on]->
-                        scheduled = NULL;
+                cluster->cpus[tsk_rt(t)->scheduled_on]->scheduled = NULL;
                 tsk_rt(t)->scheduled_on = NO_CPU;
         }
-        spin_unlock_irqrestore(&task_cedf(t)->slock, flags);
+        spin_unlock_irqrestore(&cluster->lock, flags);
 
         BUG_ON(!is_realtime(t));
         TRACE_TASK(t, "RIP\n");
@@ -623,10 +584,147 @@ static void cedf_task_exit(struct task_struct * t)
 
 static long cedf_admit_task(struct task_struct* tsk)
 {
-        return (task_cpu(tsk) >= task_cedf(tsk)->first_cpu &&
-                task_cpu(tsk) <= task_cedf(tsk)->last_cpu) ? 0 : -EINVAL;
+        return task_cpu(tsk) == tsk->rt_param.task_params.cpu ? 0 : -EINVAL;
+}
+
+/* total number of cluster */
+static int num_clusters;
+/* we do not support cluster of different sizes */
+static unsigned int cluster_size;
+
+#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 int clusters_allocated = 0;
+
+static void cleanup_cedf(void)
+{
+        int i;
+
+        if (clusters_allocated) {
+                for (i = 0; i < num_clusters; i++) {
+                        kfree(cedf[i].cpus);
+                        kfree(cedf[i].heap_node);
+                        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, cache index = %d\n",
+                        cluster_cache_index);
+
+        /* need to get cluster_size first */
+        if(!zalloc_cpumask_var(&mask, GFP_ATOMIC))
+                return -ENOMEM;
+
+        chk = get_shared_cpu_map(mask, 0, cluster_cache_index);
+        if (chk) {
+                /* if chk != 0 then it is the max allowed index */
+                printk(KERN_INFO "C-EDF: Cannot support cache index = %d\n",
+                                cluster_cache_index);
+                printk(KERN_INFO "C-EDF: Using cache index = %d\n",
+                                chk);
+                cluster_cache_index = chk;
+        }
+
+        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);
+                cedf[i].heap_node = kmalloc(
+                                cluster_size * sizeof(struct bheap_node),
+                                GFP_ATOMIC);
+                bheap_init(&(cedf[i].cpu_heap));
+                edf_domain_init(&(cedf[i].domain), NULL, cedf_release_jobs);
+
+                if(!zalloc_cpumask_var(&cedf[i].cpu_map, GFP_ATOMIC))
+                        return -ENOMEM;
+        }
+
+        /* cycle through cluster and add cpus to them */
+        for (i = 0; i < num_clusters; i++) {
+
+                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 (cpumask_test_cpu(cpu, cedf[j].cpu_map))
+                                continue;
+
+                        /* this cpu isn't in any cluster */
+                        /* get the shared cpus */
+                        get_shared_cpu_map(mask, cpu, cluster_cache_index);
+                        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];
+                                entry->hn = &(cedf[i].heap_node[cpu_count]);
+                                bheap_node_init(&entry->hn, entry);
+
+                                cpu_count++;
+
+                                entry->linked = NULL;
+                                entry->scheduled = NULL;
+                                update_cpu_position(entry);
+                        }
+                        /* done with this cluster */
+                        break;
+                }
+        }
+
+        free_cpumask_var(mask);
+        clusters_allocated = 1;
+        return 0;
+}
 
 /*      Plugin object   */
 static struct sched_plugin cedf_plugin __cacheline_aligned_in_smp = {
@@ -639,89 +737,20 @@ static struct sched_plugin cedf_plugin __cacheline_aligned_in_smp = {
         .schedule               = cedf_schedule,
         .task_wake_up           = cedf_task_wake_up,
         .task_block             = cedf_task_block,
-        .admit_task             = cedf_admit_task
+        .admit_task             = cedf_admit_task,
+        .activate_plugin        = cedf_activate_plugin,
 };
 
-static void cedf_domain_init(int first_cpu, int last_cpu)
-{
-        int cpu;
-
-        /* Create new domain for this cluster. */
-        cedf_domain_t *new_cedf_domain = kmalloc(sizeof(*new_cedf_domain),
-                                                     GFP_KERNEL);
-
-        /* Initialize cluster domain. */
-        edf_domain_init(&new_cedf_domain->domain, NULL,
-                        cedf_release_jobs);
-        new_cedf_domain->first_cpu      = first_cpu;
-        new_cedf_domain->last_cpu       = last_cpu;
-        INIT_LIST_HEAD(&new_cedf_domain->cedf_cpu_queue);
-
-        /* Assign all cpus in cluster to point to this domain. */
-        for (cpu = first_cpu; cpu <= last_cpu; cpu++) {
-                remote_cedf(cpu) = new_cedf_domain;
-                cedf_domains_array[cpu] = new_cedf_domain;
-        }
-}
 
 static int __init init_cedf(void)
 {
-        int cpu;
-        cpu_entry_t *entry;
-
-        /* num_online_cpus() should have been set already
-         * if the number of available cpus is less then the cluster
-         * size (currently 4) then it is pointless trying to use
-         * CEDF, so we disable this plugin
-         */
-        if(num_online_cpus() < cluster_size) {
-                printk(KERN_INFO "Not registering C-EDF plugin: "
-                        "Num Online Cpus (%d) < Min Cluster Size (%d)\n",
-                        num_online_cpus(), cluster_size);
-                do_cleanup = 0;
-                return 0;
-        }
-
-        /*
-         * initialize short_cut for per-cpu cedf state;
-         * there may be a problem here if someone removes a cpu
-         * while we are doing this initialization... and if cpus
-         * are added / removed later... is it a _real_ problem for cedf?
-         */
-        cedf_cpu_entries_array = kmalloc(
-                        sizeof(cpu_entry_t *) * num_online_cpus(),
-                        GFP_KERNEL);
-
-        cedf_domains_array = kmalloc(
-                        sizeof(cedf_domain_t *) * num_online_cpus(),
-                        GFP_KERNEL);
-
-        /* initialize CPU state */
-        for (cpu = 0; cpu < num_online_cpus(); cpu++)  {
-                entry = &per_cpu(cedf_cpu_entries, cpu);
-                cedf_cpu_entries_array[cpu] = entry;
-                atomic_set(&entry->will_schedule, 0);
-                entry->linked    = NULL;
-                entry->scheduled = NULL;
-                entry->cpu       = cpu;
-                INIT_LIST_HEAD(&entry->list);
-        }
-
-        /* initialize all cluster domains */
-        for (cpu = 0; cpu < num_online_cpus(); cpu += cluster_size)
-                cedf_domain_init(cpu, cpu+cluster_size-1);
-
         return register_sched_plugin(&cedf_plugin);
 }
 
 static void clean_cedf(void)
 {
-        if(do_cleanup) {
-                kfree(cedf_cpu_entries_array);
-                kfree(cedf_domains_array);
-        }
+        cleanup_cedf();
 }
 
 module_init(init_cedf);
 module_exit(clean_cedf);
-
diff --git a/litmus/sched_plugin.c b/litmus/sched_plugin.c
index bc7c0e93fb18..3767b30e610a 100644
--- a/litmus/sched_plugin.c
+++ b/litmus/sched_plugin.c
@@ -171,6 +171,14 @@ struct sched_plugin linux_sched_plugin = {
 };
 
 /*
+ * The cluster size is needed in C-EDF: it makes sense only to cluster
+ * around L2 or L3, so if cluster_cache_index = 2 (default) we cluster
+ * all the CPUs that shares a L2 cache, while cluster_cache_index = 3
+ * we cluster all CPs that shares a L3 cache
+ */
+int cluster_cache_index = 2;
+
+/*
  *      The reference to current plugin that is used to schedule tasks within
  *      the system. It stores references to actual function implementations
  *      Should be initialized by calling "init_***_plugin()"