Merged ckrm-E15 CPU controller
authorMarc Fiuczynski <mef@cs.princeton.edu>
Sat, 17 Jul 2004 13:00:36 +0000 (13:00 +0000)
committerMarc Fiuczynski <mef@cs.princeton.edu>
Sat, 17 Jul 2004 13:00:36 +0000 (13:00 +0000)
.config
.config.old
include/linux/autoconf.h
include/linux/sched.h
init/Kconfig
init/main.c
kernel/Makefile
kernel/sched.c

diff --git a/.config b/.config
index 7f9bc8d..a1a300b 100644 (file)
--- a/.config
+++ b/.config
@@ -29,6 +29,8 @@ CONFIG_CKRM=y
 CONFIG_RCFS_FS=y
 CONFIG_CKRM_TYPE_TASKCLASS=y
 CONFIG_CKRM_RES_NUMTASKS=y
+CONFIG_CKRM_CPU_SCHEDULE=y
+CONFIG_CKRM_CPU_MONITOR=y
 CONFIG_CKRM_TYPE_SOCKETCLASS=y
 CONFIG_CKRM_RES_LISTENAQ=m
 CONFIG_CKRM_RBCE=m
index 55ca636..b22fb71 100644 (file)
@@ -29,8 +29,13 @@ CONFIG_CKRM=y
 CONFIG_RCFS_FS=y
 CONFIG_CKRM_TYPE_TASKCLASS=y
 CONFIG_CKRM_RES_NUMTASKS=y
+CONFIG_CKRM_CPU_SCHEDULE=y
+CONFIG_CKRM_CPU_MONITOR=m
 CONFIG_CKRM_TYPE_SOCKETCLASS=y
 CONFIG_CKRM_RES_LISTENAQ=m
+CONFIG_CKRM_RBCE=m
+CONFIG_CKRM_CRBCE=m
+# CONFIG_BSD_PROCESS_ACCT_V3 is not set
 CONFIG_SYSCTL=y
 CONFIG_AUDIT=y
 CONFIG_AUDITSYSCALL=y
@@ -414,7 +419,6 @@ CONFIG_BLK_DEV_IDE=y
 # CONFIG_BLK_DEV_HD_IDE is not set
 CONFIG_BLK_DEV_IDEDISK=y
 CONFIG_IDEDISK_MULTI_MODE=y
-# CONFIG_IDEDISK_STROKE is not set
 CONFIG_BLK_DEV_IDECS=m
 CONFIG_BLK_DEV_IDECD=y
 CONFIG_BLK_DEV_IDETAPE=m
@@ -506,6 +510,7 @@ CONFIG_SCSI_FC_ATTRS=m
 # SCSI low-level drivers
 #
 CONFIG_BLK_DEV_3W_XXXX_RAID=m
+# CONFIG_SCSI_3W_9XXX is not set
 # CONFIG_SCSI_7000FASST is not set
 # CONFIG_SCSI_ACARD is not set
 CONFIG_SCSI_AHA152X=m
@@ -542,7 +547,6 @@ CONFIG_SCSI_SATA_VIA=m
 CONFIG_SCSI_SATA_VITESSE=m
 CONFIG_SCSI_BUSLOGIC=m
 # CONFIG_SCSI_OMIT_FLASHPOINT is not set
-# CONFIG_SCSI_CPQFCTS is not set
 # CONFIG_SCSI_DMX3191D is not set
 # CONFIG_SCSI_DTC3280 is not set
 # CONFIG_SCSI_EATA is not set
@@ -613,6 +617,9 @@ CONFIG_MD_RAID6=m
 CONFIG_MD_MULTIPATH=m
 CONFIG_BLK_DEV_DM=m
 CONFIG_DM_CRYPT=m
+# CONFIG_DM_SNAPSHOT is not set
+# CONFIG_DM_MIRROR is not set
+# CONFIG_DM_ZERO is not set
 
 #
 # Fusion MPT device support
@@ -1083,6 +1090,7 @@ CONFIG_SUNDANCE=m
 CONFIG_TLAN=m
 CONFIG_VIA_RHINE=m
 CONFIG_VIA_RHINE_MMIO=y
+# CONFIG_VIA_VELOCITY is not set
 CONFIG_NET_POCKET=y
 CONFIG_ATP=m
 CONFIG_DE600=m
@@ -1576,6 +1584,7 @@ CONFIG_DRM_SIS=m
 CONFIG_SYNCLINK_CS=m
 CONFIG_MWAVE=m
 # CONFIG_RAW_DRIVER is not set
+# CONFIG_HPET is not set
 CONFIG_HANGCHECK_TIMER=m
 
 #
@@ -1783,6 +1792,7 @@ CONFIG_VIDEO_SELECT=y
 CONFIG_FB_HGA=m
 # CONFIG_FB_HGA_ACCEL is not set
 CONFIG_FB_RIVA=m
+# CONFIG_FB_RIVA_I2C is not set
 CONFIG_FB_I810=m
 CONFIG_FB_I810_GTF=y
 CONFIG_FB_MATROX=m
@@ -1993,6 +2003,7 @@ CONFIG_USB_ACM=m
 CONFIG_USB_PRINTER=m
 CONFIG_USB_STORAGE=m
 # CONFIG_USB_STORAGE_DEBUG is not set
+# CONFIG_USB_STORAGE_RW_DETECT is not set
 CONFIG_USB_STORAGE_DATAFAB=y
 CONFIG_USB_STORAGE_FREECOM=y
 CONFIG_USB_STORAGE_ISD200=y
@@ -2284,7 +2295,6 @@ CONFIG_MINIX_SUBPARTITION=y
 CONFIG_SOLARIS_X86_PARTITION=y
 CONFIG_UNIXWARE_DISKLABEL=y
 # CONFIG_LDM_PARTITION is not set
-# CONFIG_NEC98_PARTITION is not set
 CONFIG_SGI_PARTITION=y
 # CONFIG_ULTRIX_PARTITION is not set
 CONFIG_SUN_PARTITION=y
@@ -2411,5 +2421,4 @@ CONFIG_LIBCRC32C=m
 CONFIG_ZLIB_INFLATE=y
 CONFIG_ZLIB_DEFLATE=m
 CONFIG_X86_BIOS_REBOOT=y
-CONFIG_X86_STD_RESOURCES=y
 CONFIG_PC=y
index 5ee385a..307d630 100644 (file)
@@ -30,6 +30,8 @@
 #define CONFIG_RCFS_FS 1
 #define CONFIG_CKRM_TYPE_TASKCLASS 1
 #define CONFIG_CKRM_RES_NUMTASKS 1
+#define CONFIG_CKRM_CPU_SCHEDULE 1
+#define CONFIG_CKRM_CPU_MONITOR 1
 #define CONFIG_CKRM_TYPE_SOCKETCLASS 1
 #define CONFIG_CKRM_RES_LISTENAQ_MODULE 1
 #define CONFIG_CKRM_RBCE_MODULE 1
index 385b390..0b1efea 100644 (file)
@@ -551,6 +551,9 @@ struct task_struct {
        // .. Hubertus should change to CONFIG_CKRM_TYPE_TASKCLASS 
        struct ckrm_task_class *taskclass;
        struct list_head        taskclass_link;
+#ifdef CONFIG_CKRM_CPU_SCHEDULE
+        struct ckrm_cpu_class *cpu_class;
+#endif
 #endif // CONFIG_CKRM_TYPE_TASKCLASS
 #endif // CONFIG_CKRM
 
@@ -1009,8 +1012,7 @@ static inline struct mm_struct * get_task_mm(struct task_struct * task)
 
        return mm;
 }
+
 /* set thread flags in other task's structures
  * - see asm/thread_info.h for TIF_xxxx flags available
  */
index d9153c5..94af58d 100644 (file)
@@ -168,6 +168,24 @@ config CKRM_RES_NUMTASKS
        
          Say N if unsure, Y to use the feature.
 
+config CKRM_CPU_SCHEDULE
+       bool "CKRM CPU scheduler"
+       depends on CKRM_TYPE_TASKCLASS
+       default m
+       help
+         Use CKRM CPU scheduler instead of Linux Scheduler
+       
+         Say N if unsure, Y to use the feature.
+
+config CKRM_CPU_MONITOR
+       tristate  "CKRM CPU Resoure Monitor"
+       depends on CKRM_CPU_SCHEDULE
+       default m
+       help
+         Monitor CPU Resource Usage of the classes
+       
+         Say N if unsure, Y to use the feature.
+
 config CKRM_TYPE_SOCKETCLASS
        bool "Class Manager for socket groups"
        depends on CKRM
index 155d05f..0dfaf47 100644 (file)
@@ -49,6 +49,8 @@
 #include <asm/bugs.h>
 
 #include <linux/ckrm.h>
+int __init init_ckrm_sched_res(void);
+
 
 /*
  * This is one of the first .c files built. Error out early
@@ -419,7 +421,6 @@ asmlinkage void __init start_kernel(void)
         * printk() and can access its per-cpu storage.
         */
        smp_prepare_boot_cpu();
-
        /*
         * Set up the scheduler prior starting any interrupts (such as the
         * timer interrupt). Full topology setup happens at smp_init()
@@ -638,8 +639,8 @@ static int init(void * unused)
         * firmware files.
         */
        populate_rootfs();
-
        do_basic_setup();
+       init_ckrm_sched_res();
 
        /*
         * check if there is an early userspace init.  If yes, let it do all
index e1b6501..e0a5feb 100644 (file)
@@ -26,8 +26,12 @@ obj-$(CONFIG_COMPAT) += compat.o
 obj-$(CONFIG_IKCONFIG) += configs.o
 obj-$(CONFIG_IKCONFIG_PROC) += configs.o
 obj-$(CONFIG_STOP_MACHINE) += stop_machine.o
+obj-$(CONFIG_CKRM_CPU_SCHEDULE) += ckrm_classqueue.o
+obj-$(CONFIG_CKRM_CPU_SCHEDULE) += ckrm_sched.o
 obj-$(CONFIG_AUDIT) += audit.o
 obj-$(CONFIG_AUDITSYSCALL) += auditsc.o
+obj-$(CONFIG_KGDB) += kgdbstub.o
+
 
 ifneq ($(CONFIG_IA64),y)
 # According to Alan Modra <alan@linuxcare.com.au>, the -fno-omit-frame-pointer is
index a5058d6..9667edb 100644 (file)
@@ -17,7 +17,6 @@
  *  2003-09-03 Interactivity tuning by Con Kolivas.
  *  2004-04-02 Scheduler domains code by Nick Piggin
  */
-
 #include <linux/mm.h>
 #include <linux/module.h>
 #include <linux/nmi.h>
 #define LOW_CREDIT(p) \
        ((p)->interactive_credit < -CREDIT_LIMIT)
 
-#define TASK_PREEMPTS_CURR(p, rq) \
-       ((p)->prio < (rq)->curr->prio)
-
 /*
  * BASE_TIMESLICE scales user-nice values [ -20 ... 19 ]
  * to time slice values.
@@ -189,16 +185,24 @@ static unsigned int task_timeslice(task_t *p)
 /*
  * These are the runqueue data structures:
  */
+typedef struct runqueue runqueue_t;
 
-#define BITMAP_SIZE ((((MAX_PRIO+1+7)/8)+sizeof(long)-1)/sizeof(long))
+#ifdef CONFIG_CKRM_CPU_SCHEDULE
+#include <linux/ckrm_classqueue.h>
+#endif
 
-typedef struct runqueue runqueue_t;
+#ifdef CONFIG_CKRM_CPU_SCHEDULE
 
-struct prio_array {
-       unsigned int nr_active;
-       unsigned long bitmap[BITMAP_SIZE];
-       struct list_head queue[MAX_PRIO];
-};
+/**
+ *  if belong to different class, compare class priority
+ *  otherwise compare task priority 
+ */
+#define TASK_PREEMPTS_CURR(p, rq) \
+       (((p)->cpu_class != (rq)->curr->cpu_class) && ((rq)->curr != (rq)->idle))? class_preempts_curr((p),(rq)->curr) : ((p)->prio < (rq)->curr->prio)
+#else
+#define TASK_PREEMPTS_CURR(p, rq) \
+       ((p)->prio < (rq)->curr->prio)
+#endif
 
 /*
  * This is the main, per-CPU runqueue data structure.
@@ -215,7 +219,7 @@ struct runqueue {
         * remote CPUs use both these fields when doing load calculation.
         */
        unsigned long nr_running;
-#ifdef CONFIG_SMP
+#if defined(CONFIG_SMP)
        unsigned long cpu_load;
 #endif
        unsigned long long nr_switches;
@@ -223,7 +227,12 @@ struct runqueue {
        unsigned long long timestamp_last_tick;
        task_t *curr, *idle;
        struct mm_struct *prev_mm;
-       prio_array_t *active, *expired, arrays[2];
+#ifdef CONFIG_CKRM_CPU_SCHEDULE
+       unsigned long ckrm_cpu_load;
+       struct classqueue_struct classqueue;   
+#else
+        prio_array_t *active, *expired, arrays[2];
+#endif
        int best_expired_prio;
        atomic_t nr_iowait;
 
@@ -260,12 +269,158 @@ static DEFINE_PER_CPU(struct runqueue, runqueues);
 # define task_running(rq, p)           ((rq)->curr == (p))
 #endif
 
+#ifdef CONFIG_CKRM_CPU_SCHEDULE
+#include <linux/ckrm_sched.h>
+spinlock_t cvt_lock        = SPIN_LOCK_UNLOCKED;
+rwlock_t   class_list_lock = RW_LOCK_UNLOCKED;
+LIST_HEAD(active_cpu_classes);   // list of active cpu classes; anchor
+struct ckrm_cpu_class default_cpu_class_obj;
+
+/*
+ * the minimum CVT allowed is the base_cvt
+ * otherwise, it will starve others
+ */
+CVT_t get_min_cvt(int cpu)
+{
+       cq_node_t *node;
+       struct ckrm_local_runqueue * lrq;
+       CVT_t min_cvt;
+
+       node = classqueue_get_head(bpt_queue(cpu));
+       lrq =  (node) ? class_list_entry(node) : NULL;
+       
+       if (lrq) 
+               min_cvt = lrq->local_cvt;
+       else 
+               min_cvt = 0;
+               
+       return min_cvt;
+}
+
+/*
+ * update the classueue base for all the runqueues
+ * TODO: we can only update half of the min_base to solve the movebackward issue
+ */
+static inline void check_update_class_base(int this_cpu) {
+       unsigned long min_base = 0xFFFFFFFF; 
+       cq_node_t *node;
+       int i;
+
+       if (! cpu_online(this_cpu)) return;
+
+       /*
+        * find the min_base across all the processors
+        */
+       for_each_online_cpu(i) {
+               /*
+                * I should change it to directly use bpt->base
+                */
+               node = classqueue_get_head(bpt_queue(i));
+               if (node && node->prio < min_base) {
+                       min_base = node->prio;
+               }
+       }
+       if (min_base != 0xFFFFFFFF) 
+               classqueue_update_base(bpt_queue(this_cpu),min_base);
+}
+
+static inline void ckrm_rebalance_tick(int j,int this_cpu)
+{
+#ifdef CONFIG_CKRM_CPU_SCHEDULE
+       read_lock(&class_list_lock);
+       if (!(j % CVT_UPDATE_TICK))
+               update_global_cvts(this_cpu);
+
+#define CKRM_BASE_UPDATE_RATE 400
+       if (! (jiffies % CKRM_BASE_UPDATE_RATE))
+               check_update_class_base(this_cpu);
+
+       read_unlock(&class_list_lock);
+#endif
+}
+
+static inline struct ckrm_local_runqueue *rq_get_next_class(struct runqueue *rq)
+{
+       cq_node_t *node = classqueue_get_head(&rq->classqueue);
+       return ((node) ? class_list_entry(node) : NULL);
+}
+
+static inline struct task_struct * rq_get_next_task(struct runqueue* rq) 
+{
+       prio_array_t               *array;
+       struct task_struct         *next;
+       struct ckrm_local_runqueue *queue;
+       int cpu = smp_processor_id();
+       
+       next = rq->idle;
+ retry_next_class:
+       if ((queue = rq_get_next_class(rq))) {
+               array = queue->active;
+               //check switch active/expired queue
+               if (unlikely(!queue->active->nr_active)) {
+                       queue->active = queue->expired;
+                       queue->expired = array;
+                       queue->expired_timestamp = 0;
+
+                       if (queue->active->nr_active)
+                               set_top_priority(queue,
+                                                find_first_bit(queue->active->bitmap, MAX_PRIO));
+                       else {
+                               classqueue_dequeue(queue->classqueue,
+                                                  &queue->classqueue_linkobj);
+                               cpu_demand_event(get_rq_local_stat(queue,cpu),CPU_DEMAND_DEQUEUE,0);
+                       }
+
+                       goto retry_next_class;                          
+               }
+               BUG_ON(!queue->active->nr_active);
+               next = task_list_entry(array->queue[queue->top_priority].next);
+       }
+       return next;
+}
+
+static inline void rq_load_inc(runqueue_t *rq, struct task_struct *p) { rq->ckrm_cpu_load += cpu_class_weight(p->cpu_class); }
+static inline void rq_load_dec(runqueue_t *rq, struct task_struct *p) { rq->ckrm_cpu_load -= cpu_class_weight(p->cpu_class); }
+
+#else /*CONFIG_CKRM_CPU_SCHEDULE*/
+
+static inline struct task_struct * rq_get_next_task(struct runqueue* rq) 
+{
+       prio_array_t *array;
+        struct list_head *queue;
+       int idx;
+
+       array = rq->active;
+       if (unlikely(!array->nr_active)) {
+               /*
+                * Switch the active and expired arrays.
+                */
+               rq->active = rq->expired;
+               rq->expired = array;
+               array = rq->active;
+               rq->expired_timestamp = 0;
+               rq->best_expired_prio = MAX_PRIO;
+       }
+
+       idx = sched_find_first_bit(array->bitmap);
+       queue = array->queue + idx;
+       return list_entry(queue->next, task_t, run_list);
+}
+
+static inline void class_enqueue_task(struct task_struct* p, prio_array_t *array) { }
+static inline void class_dequeue_task(struct task_struct* p, prio_array_t *array) { }
+static inline void init_cpu_classes(void) { }
+static inline void rq_load_inc(runqueue_t *rq, struct task_struct *p) { }
+static inline void rq_load_dec(runqueue_t *rq, struct task_struct *p) { }
+#endif  /* CONFIG_CKRM_CPU_SCHEDULE */
+
+
 /*
  * task_rq_lock - lock the runqueue a given task resides on and disable
  * interrupts.  Note the ordering: we can safely lookup the task_rq without
  * explicitly disabling preemption.
  */
-static runqueue_t *task_rq_lock(task_t *p, unsigned long *flags)
+runqueue_t *task_rq_lock(task_t *p, unsigned long *flags)
 {
        struct runqueue *rq;
 
@@ -280,7 +435,7 @@ repeat_lock_task:
        return rq;
 }
 
-static inline void task_rq_unlock(runqueue_t *rq, unsigned long *flags)
+void task_rq_unlock(runqueue_t *rq, unsigned long *flags)
 {
        spin_unlock_irqrestore(&rq->lock, *flags);
 }
@@ -307,20 +462,23 @@ static inline void rq_unlock(runqueue_t *rq)
 /*
  * Adding/removing a task to/from a priority array:
  */
-static void dequeue_task(struct task_struct *p, prio_array_t *array)
+void dequeue_task(struct task_struct *p, prio_array_t *array)
 {
+       BUG_ON(! array);
        array->nr_active--;
        list_del(&p->run_list);
        if (list_empty(array->queue + p->prio))
                __clear_bit(p->prio, array->bitmap);
+       class_dequeue_task(p,array);
 }
 
-static void enqueue_task(struct task_struct *p, prio_array_t *array)
+void enqueue_task(struct task_struct *p, prio_array_t *array)
 {
        list_add_tail(&p->run_list, array->queue + p->prio);
        __set_bit(p->prio, array->bitmap);
        array->nr_active++;
        p->array = array;
+       class_enqueue_task(p,array);
 }
 
 /*
@@ -334,6 +492,7 @@ static inline void enqueue_task_head(struct task_struct *p, prio_array_t *array)
        __set_bit(p->prio, array->bitmap);
        array->nr_active++;
        p->array = array;
+       class_enqueue_task(p,array);
 }
 
 /*
@@ -375,8 +534,9 @@ static int effective_prio(task_t *p)
  */
 static inline void __activate_task(task_t *p, runqueue_t *rq)
 {
-       enqueue_task(p, rq->active);
+       enqueue_task(p, rq_active(p,rq));
        rq->nr_running++;
+       rq_load_inc(rq,p);
 }
 
 /*
@@ -384,8 +544,9 @@ static inline void __activate_task(task_t *p, runqueue_t *rq)
  */
 static inline void __activate_idle_task(task_t *p, runqueue_t *rq)
 {
-       enqueue_task_head(p, rq->active);
+       enqueue_task_head(p, rq_active(p,rq));
        rq->nr_running++;
+       rq_load_inc(rq,p);
 }
 
 static void recalc_task_prio(task_t *p, unsigned long long now)
@@ -517,6 +678,7 @@ static void activate_task(task_t *p, runqueue_t *rq, int local)
 static void deactivate_task(struct task_struct *p, runqueue_t *rq)
 {
        rq->nr_running--;
+       rq_load_dec(rq,p);
        if (p->state == TASK_UNINTERRUPTIBLE)
                rq->nr_uninterruptible++;
        dequeue_task(p, p->array);
@@ -966,6 +1128,7 @@ void fastcall wake_up_forked_process(task_t * p)
                p->array = current->array;
                p->array->nr_active++;
                rq->nr_running++;
+               rq_load_inc(rq,p);
        }
        task_rq_unlock(rq, &flags);
 }
@@ -1288,6 +1451,7 @@ lock_again:
                        p->array = current->array;
                        p->array->nr_active++;
                        rq->nr_running++;
+                       rq_load_inc(rq,p);
                }
        } else {
                /* Not the local CPU - must adjust timestamp */
@@ -1392,9 +1556,13 @@ void pull_task(runqueue_t *src_rq, prio_array_t *src_array, task_t *p,
 {
        dequeue_task(p, src_array);
        src_rq->nr_running--;
+       rq_load_dec(src_rq,p);
+
        set_task_cpu(p, this_cpu);
        this_rq->nr_running++;
+       rq_load_inc(this_rq,p);
        enqueue_task(p, this_array);
+
        p->timestamp = (p->timestamp - src_rq->timestamp_last_tick)
                                + this_rq->timestamp_last_tick;
        /*
@@ -1433,6 +1601,194 @@ int can_migrate_task(task_t *p, runqueue_t *rq, int this_cpu,
        return 1;
 }
 
+#ifdef CONFIG_CKRM_CPU_SCHEDULE
+
+struct ckrm_cpu_class *find_unbalanced_class(int busiest_cpu, int this_cpu, unsigned long *cls_imbalance)
+{
+       struct ckrm_cpu_class *most_unbalanced_class = NULL;
+       struct ckrm_cpu_class *clsptr;
+       int max_unbalance = 0;
+
+       list_for_each_entry(clsptr,&active_cpu_classes,links) {
+               struct ckrm_local_runqueue *this_lrq    = get_ckrm_local_runqueue(clsptr,this_cpu);
+               struct ckrm_local_runqueue *busiest_lrq = get_ckrm_local_runqueue(clsptr,busiest_cpu);
+               int unbalance_degree;
+               
+               unbalance_degree = (local_queue_nr_running(busiest_lrq) - local_queue_nr_running(this_lrq)) * cpu_class_weight(clsptr);
+               if (unbalance_degree >= *cls_imbalance) 
+                       continue;  // already looked at this class
+
+               if (unbalance_degree > max_unbalance) {
+                       max_unbalance = unbalance_degree;
+                       most_unbalanced_class = clsptr;
+               }
+       }
+       *cls_imbalance = max_unbalance;
+       return most_unbalanced_class;
+}
+
+
+/*
+ * find_busiest_queue - find the busiest runqueue among the cpus in cpumask.
+ */
+static int find_busiest_cpu(runqueue_t *this_rq, int this_cpu, int idle, 
+                           int *imbalance)
+{
+       int cpu_load, load, max_load, i, busiest_cpu;
+       runqueue_t *busiest, *rq_src;
+
+
+       /*Hubertus ... the concept of nr_running is replace with cpu_load */
+       cpu_load = this_rq->ckrm_cpu_load;
+
+       busiest = NULL;
+       busiest_cpu = -1;
+
+       max_load = -1;
+       for_each_online_cpu(i) {
+               rq_src = cpu_rq(i);
+               load = rq_src->ckrm_cpu_load;
+
+               if ((load > max_load) && (rq_src != this_rq)) {
+                       busiest = rq_src;
+                       busiest_cpu = i;
+                       max_load = load;
+               }
+       }
+
+       if (likely(!busiest))
+               goto out;
+
+       *imbalance = max_load - cpu_load;
+
+       /* It needs an at least ~25% imbalance to trigger balancing. */
+       if (!idle && ((*imbalance)*4 < max_load)) {
+               busiest = NULL;
+               goto out;
+       }
+
+       double_lock_balance(this_rq, busiest);
+       /*
+        * Make sure nothing changed since we checked the
+        * runqueue length.
+        */
+       if (busiest->ckrm_cpu_load <= cpu_load) {
+               spin_unlock(&busiest->lock);
+               busiest = NULL;
+       }
+out:
+       return (busiest ? busiest_cpu : -1);
+}
+
+static int load_balance(int this_cpu, runqueue_t *this_rq,
+                       struct sched_domain *sd, enum idle_type idle)
+{
+       int imbalance, idx;
+       int busiest_cpu;
+       runqueue_t *busiest;
+       prio_array_t *array;
+       struct list_head *head, *curr;
+       task_t *tmp;
+        struct ckrm_local_runqueue * busiest_local_queue;
+       struct ckrm_cpu_class *clsptr;
+       int weight;
+       unsigned long cls_imbalance;      // so we can retry other classes
+
+       // need to update global CVT based on local accumulated CVTs
+       read_lock(&class_list_lock);
+       busiest_cpu = find_busiest_cpu(this_rq, this_cpu, idle, &imbalance);
+       if (busiest_cpu == -1)
+               goto out;
+
+       busiest = cpu_rq(busiest_cpu);
+
+       /*
+        * We only want to steal a number of tasks equal to 1/2 the imbalance,
+        * otherwise we'll just shift the imbalance to the new queue:
+        */
+       imbalance /= 2;
+               
+       /* now find class on that runqueue with largest inbalance */
+       cls_imbalance = 0xFFFFFFFF; 
+
+ retry_other_class:
+       clsptr = find_unbalanced_class(busiest_cpu, this_cpu, &cls_imbalance);
+       if (!clsptr) 
+               goto out_unlock;
+
+       busiest_local_queue = get_ckrm_local_runqueue(clsptr,busiest_cpu);
+       weight = cpu_class_weight(clsptr);
+
+       /*
+        * We first consider expired tasks. Those will likely not be
+        * executed in the near future, and they are most likely to
+        * be cache-cold, thus switching CPUs has the least effect
+        * on them.
+        */
+       if (busiest_local_queue->expired->nr_active)
+               array = busiest_local_queue->expired;
+       else
+               array = busiest_local_queue->active;
+       
+ new_array:
+       /* Start searching at priority 0: */
+       idx = 0;
+ skip_bitmap:
+       if (!idx)
+               idx = sched_find_first_bit(array->bitmap);
+       else
+               idx = find_next_bit(array->bitmap, MAX_PRIO, idx);
+       if (idx >= MAX_PRIO) {
+               if (array == busiest_local_queue->expired && busiest_local_queue->active->nr_active) {
+                       array = busiest_local_queue->active;
+                       goto new_array;
+               }
+               goto retry_other_class;
+       }
+       
+       head = array->queue + idx;
+       curr = head->prev;
+ skip_queue:
+       tmp = list_entry(curr, task_t, run_list);
+       
+       curr = curr->prev;
+       
+       if (!can_migrate_task(tmp, busiest, this_cpu, sd,idle)) {
+               if (curr != head)
+                       goto skip_queue;
+               idx++;
+               goto skip_bitmap;
+       }
+       pull_task(busiest, array, tmp, this_rq, rq_active(tmp,this_rq),this_cpu);
+       /*
+        * tmp BUG FIX: hzheng
+        * load balancing can make the busiest local queue empty
+        * thus it should be removed from bpt
+        */
+       if (! local_queue_nr_running(busiest_local_queue)) {
+               classqueue_dequeue(busiest_local_queue->classqueue,&busiest_local_queue->classqueue_linkobj);
+               cpu_demand_event(get_rq_local_stat(busiest_local_queue,busiest_cpu),CPU_DEMAND_DEQUEUE,0);              
+       }
+
+       imbalance -= weight;
+       if (!idle && (imbalance>0)) {
+               if (curr != head)
+                       goto skip_queue;
+               idx++;
+               goto skip_bitmap;
+       }
+ out_unlock:
+       spin_unlock(&busiest->lock);
+ out:
+       read_unlock(&class_list_lock);
+       return 0;
+}
+
+
+static inline void idle_balance(int this_cpu, runqueue_t *this_rq)
+{
+}
+#else /* CONFIG_CKRM_CPU_SCHEDULE */
 /*
  * move_tasks tries to move up to max_nr_move tasks from busiest to this_rq,
  * as part of a balancing operation within "domain". Returns the number of
@@ -1883,6 +2239,7 @@ next_group:
        } while (group != sd->groups);
 >>>>>>> 1.1.9.3
 }
+#endif /* CONFIG_CKRM_CPU_SCHEDULE*/
 
 /*
  * rebalance_tick will get called every timer tick, on every CPU.
@@ -1903,6 +2260,8 @@ static void rebalance_tick(int this_cpu, runqueue_t *this_rq,
        unsigned long j = jiffies + CPU_OFFSET(this_cpu);
        struct sched_domain *sd;
 
+       ckrm_rebalance_tick(j,this_cpu);
+
        /* Update our load */
        old_load = this_rq->cpu_load;
        this_load = this_rq->nr_running * SCHED_LOAD_SCALE;
@@ -1935,13 +2294,15 @@ static void rebalance_tick(int this_cpu, runqueue_t *this_rq,
                }
        }
 }
-#else
+#else /* SMP*/
 /*
  * on UP we do not need to balance between CPUs:
  */
 static inline void rebalance_tick(int cpu, runqueue_t *rq, enum idle_type idle)
 {
+       ckrm_rebalance_tick(jiffies,cpu);
 }
+
 static inline void idle_balance(int cpu, runqueue_t *rq)
 {
 }
@@ -1962,7 +2323,7 @@ static inline int wake_priority_sleeper(runqueue_t *rq)
        return 0;
 }
 
-DEFINE_PER_CPU(struct kernel_stat, kstat);
+DEFINE_PER_CPU(struct kernel_stat, kstat) = { { 0 } };
 
 EXPORT_PER_CPU_SYMBOL(kstat);
 
@@ -1976,11 +2337,19 @@ EXPORT_PER_CPU_SYMBOL(kstat);
  * increasing number of running tasks. We also ignore the interactivity
  * if a better static_prio task has expired:
  */
+
+#ifndef CONFIG_CKRM_CPU_SCHEDULE
 #define EXPIRED_STARVING(rq) \
        ((STARVATION_LIMIT && ((rq)->expired_timestamp && \
                (jiffies - (rq)->expired_timestamp >= \
                        STARVATION_LIMIT * ((rq)->nr_running) + 1))) || \
                        ((rq)->curr->static_prio > (rq)->best_expired_prio))
+#else
+#define EXPIRED_STARVING(rq) \
+               (STARVATION_LIMIT && ((rq)->expired_timestamp && \
+               (jiffies - (rq)->expired_timestamp >= \
+                       STARVATION_LIMIT * (local_queue_nr_running(rq)) + 1)))
+#endif
 
 /*
  * This function gets called by the timer code, with HZ frequency.
@@ -2030,7 +2399,7 @@ void scheduler_tick(int user_ticks, int sys_ticks)
        cpustat->system += sys_ticks;
 
        /* Task might have expired already, but not scheduled off yet */
-       if (p->array != rq->active) {
+       if (p->array != rq_active(p,rq)) {
                set_tsk_need_resched(p);
                goto out;
        }
@@ -2053,12 +2422,17 @@ void scheduler_tick(int user_ticks, int sys_ticks)
                        set_tsk_need_resched(p);
 
                        /* put it at the end of the queue: */
-                       dequeue_task(p, rq->active);
-                       enqueue_task(p, rq->active);
+                       dequeue_task(p, rq_active(p,rq));
+                       enqueue_task(p, rq_active(p,rq));
                }
                goto out_unlock;
        }
+#warning MEF PLANETLAB: "if (vx_need_resched(p)) was if (!--p->time_slice) */"
        if (vx_need_resched(p)) {
+#ifdef CONFIG_CKRM_CPU_SCHEDULE
+               /* Hubertus ... we can abstract this out */
+               struct ckrm_local_runqueue* rq = get_task_class_queue(p);
+#endif
                dequeue_task(p, rq->active);
                set_tsk_need_resched(p);
                p->prio = effective_prio(p);
@@ -2069,8 +2443,8 @@ void scheduler_tick(int user_ticks, int sys_ticks)
                        rq->expired_timestamp = jiffies;
                if (!TASK_INTERACTIVE(p) || EXPIRED_STARVING(rq)) {
                        enqueue_task(p, rq->expired);
-                       if (p->static_prio < rq->best_expired_prio)
-                               rq->best_expired_prio = p->static_prio;
+                       if (p->static_prio < this_rq()->best_expired_prio)
+                               this_rq()->best_expired_prio = p->static_prio;
                } else
                        enqueue_task(p, rq->active);
        } else {
@@ -2093,12 +2467,12 @@ void scheduler_tick(int user_ticks, int sys_ticks)
                if (TASK_INTERACTIVE(p) && !((task_timeslice(p) -
                        p->time_slice) % TIMESLICE_GRANULARITY(p)) &&
                        (p->time_slice >= TIMESLICE_GRANULARITY(p)) &&
-                       (p->array == rq->active)) {
+                       (p->array == rq_active(p,rq))) {
 
-                       dequeue_task(p, rq->active);
+                       dequeue_task(p, rq_active(p,rq));
                        set_tsk_need_resched(p);
                        p->prio = effective_prio(p);
-                       enqueue_task(p, rq->active);
+                       enqueue_task(p, rq_active(p,rq));
                }
        }
 out_unlock:
@@ -2201,10 +2575,9 @@ asmlinkage void __sched schedule(void)
        task_t *prev, *next;
        runqueue_t *rq;
        prio_array_t *array;
-       struct list_head *queue;
        unsigned long long now;
        unsigned long run_time;
-       int cpu, idx;
+       int cpu;
 #ifdef CONFIG_VSERVER_HARDCPU          
        struct vx_info *vxi;
        int maxidle = -HZ;
@@ -2300,21 +2673,9 @@ pick_next:
                }
        }
 
-       array = rq->active;
-       if (unlikely(!array->nr_active)) {
-               /*
-                * Switch the active and expired arrays.
-                */
-               rq->active = rq->expired;
-               rq->expired = array;
-               array = rq->active;
-               rq->expired_timestamp = 0;
-               rq->best_expired_prio = MAX_PRIO;
-       }
-
-       idx = sched_find_first_bit(array->bitmap);
-       queue = array->queue + idx;
-       next = list_entry(queue->next, task_t, run_list);
+       next = rq_get_next_task(rq);
+       if (next == rq->idle) 
+               goto switch_tasks;
 
        if (dependent_sleeper(cpu, rq, next)) {
                next = rq->idle;
@@ -2355,6 +2716,14 @@ switch_tasks:
        clear_tsk_need_resched(prev);
        RCU_qsctr(task_cpu(prev))++;
 
+#ifdef CONFIG_CKRM_CPU_SCHEDULE
+       if (prev != rq->idle) {
+               unsigned long long run = now - prev->timestamp;
+               cpu_demand_event(get_task_local_stat(prev),CPU_DEMAND_DESCHEDULE,run);
+               update_local_cvt(prev, run);
+       }
+#endif
+
        prev->sleep_avg -= run_time;
        if ((long)prev->sleep_avg <= 0) {
                prev->sleep_avg = 0;
@@ -3062,7 +3431,7 @@ asmlinkage long sys_sched_yield(void)
 {
        runqueue_t *rq = this_rq_lock();
        prio_array_t *array = current->array;
-       prio_array_t *target = rq->expired;
+       prio_array_t *target = rq_expired(current,rq);
 
        /*
         * We implement yielding by moving the task into the expired
@@ -3072,7 +3441,7 @@ asmlinkage long sys_sched_yield(void)
         *  array.)
         */
        if (unlikely(rt_task(current)))
-               target = rq->active;
+               target = rq_active(current,rq);
 
        dequeue_task(current, array);
        enqueue_task(current, target);
@@ -3497,7 +3866,9 @@ static int migration_thread(void * data)
                }
 
                if (rq->active_balance) {
+#ifndef CONFIG_CKRM_CPU_SCHEDULE
                        active_load_balance(rq, cpu);
+#endif
                        rq->active_balance = 0;
                }
 
@@ -3973,7 +4344,10 @@ int in_sched_functions(unsigned long addr)
 void __init sched_init(void)
 {
        runqueue_t *rq;
-       int i, j, k;
+       int i;
+#ifndef CONFIG_CKRM_CPU_SCHEDULE
+       int j, k;
+#endif
 
 #ifdef CONFIG_SMP
        /* Set up an initial dummy domain for early boot */
@@ -3993,13 +4367,21 @@ void __init sched_init(void)
        sched_group_init.cpu_power = SCHED_LOAD_SCALE;
 #endif
 
+       init_cpu_classes();
+
        for (i = 0; i < NR_CPUS; i++) {
+#ifndef CONFIG_CKRM_CPU_SCHEDULE
                prio_array_t *array;
-
+#endif
                rq = cpu_rq(i);
                spin_lock_init(&rq->lock);
+
+#ifndef CONFIG_CKRM_CPU_SCHEDULE
                rq->active = rq->arrays;
                rq->expired = rq->arrays + 1;
+#else
+               rq->ckrm_cpu_load = 0;
+#endif
                rq->best_expired_prio = MAX_PRIO;
 
 #ifdef CONFIG_SMP
@@ -4013,6 +4395,7 @@ void __init sched_init(void)
                INIT_LIST_HEAD(&rq->hold_queue);
                atomic_set(&rq->nr_iowait, 0);
 
+#ifndef CONFIG_CKRM_CPU_SCHEDULE
                for (j = 0; j < 2; j++) {
                        array = rq->arrays + j;
                        for (k = 0; k < MAX_PRIO; k++) {
@@ -4022,7 +4405,9 @@ void __init sched_init(void)
                        // delimiter for bitsearch
                        __set_bit(MAX_PRIO, array->bitmap);
                }
+#endif
        }
+
        /*
         * We have to do a little magic to get the first
         * thread right in SMP mode.
@@ -4031,6 +4416,10 @@ void __init sched_init(void)
        rq->curr = current;
        rq->idle = current;
        set_task_cpu(current, smp_processor_id());
+#ifdef CONFIG_CKRM_CPU_SCHEDULE
+       current->cpu_class = default_cpu_class;
+       current->array = NULL;
+#endif
        wake_up_forked_process(current);
 
        /*
@@ -4116,3 +4505,13 @@ int task_running_sys(struct task_struct *p)
 EXPORT_SYMBOL(task_running_sys);
 #endif
 
+#ifdef CONFIG_CKRM_CPU_SCHEDULE
+/**
+ * return the classqueue object of a certain processor
+ * Note: not supposed to be used in performance sensitive functions
+ */
+struct classqueue_struct * get_cpu_classqueue(int cpu)
+{
+       return (& (cpu_rq(cpu)->classqueue) );
+}
+#endif