1.1 --- a/xen/common/sched_credit2.c	Fri Dec 24 08:26:59 2010 +0000
     1.2 +++ b/xen/common/sched_credit2.c	Fri Dec 24 08:27:42 2010 +0000
     1.3 @@ -176,10 +176,14 @@ integer_param("sched_credit2_migrate_res
     1.4   */
     1.5  struct csched_runqueue_data {
     1.6      int id;
     1.7 +
     1.8 +    spinlock_t lock;      /* Lock for this runqueue. */
     1.9 +    cpumask_t active;      /* CPUs enabled for this runqueue */
    1.10 +
    1.11      struct list_head runq; /* Ordered list of runnable vms */
    1.12      struct list_head svc;  /* List of all vcpus assigned to this runqueue */
    1.13      int max_weight;
    1.14 -    int cpu_min, cpu_max;  /* Range of physical cpus this runqueue runs */
    1.15 +
    1.16      cpumask_t idle,        /* Currently idle */
    1.17          tickled;           /* Another cpu in the queue is already targeted for this one */
    1.18  };
    1.19 @@ -189,12 +193,12 @@ struct csched_runqueue_data {
    1.20   */
    1.21  struct csched_private {
    1.22      spinlock_t lock;
    1.23 -    uint32_t ncpus;
    1.24 -
    1.25 +    cpumask_t initialized; /* CPU is initialized for this pool */
    1.26 +    
    1.27      struct list_head sdom; /* Used mostly for dump keyhandler. */
    1.28  
    1.29      int runq_map[NR_CPUS];
    1.30 -    uint32_t runq_count;
    1.31 +    cpumask_t active_queues; /* Queues which may have active cpus */
    1.32      struct csched_runqueue_data rqd[NR_CPUS];
    1.33  };
    1.34  
    1.35 @@ -341,7 +345,7 @@ runq_tickle(const struct scheduler *ops,
    1.36      int i, ipid=-1;
    1.37      s_time_t lowest=(1<<30);
    1.38      struct csched_runqueue_data *rqd = RQD(ops, cpu);
    1.39 -    cpumask_t *online, mask;
    1.40 +    cpumask_t mask;
    1.41      struct csched_vcpu * cur;
    1.42  
    1.43      d2printk("rqt d%dv%d cd%dv%d\n",
    1.44 @@ -374,9 +378,7 @@ runq_tickle(const struct scheduler *ops,
    1.45  
    1.46      /* Otherwise, look for the non-idle cpu with the lowest credit,
    1.47       * skipping cpus which have been tickled but not scheduled yet */
    1.48 -    online = CSCHED_CPUONLINE(per_cpu(cpupool, cpu));
    1.49 -
    1.50 -    cpus_andnot(mask, *online, rqd->idle);
    1.51 +    cpus_andnot(mask, rqd->active, rqd->idle);
    1.52      cpus_andnot(mask, mask, rqd->tickled);
    1.53  
    1.54      for_each_cpu_mask(i, mask)
    1.55 @@ -997,7 +999,7 @@ csched_schedule(
    1.56      const struct scheduler *ops, s_time_t now, bool_t tasklet_work_scheduled)
    1.57  {
    1.58      const int cpu = smp_processor_id();
    1.59 -    struct csched_runqueue_data *rqd = RQD(ops, cpu);
    1.60 +    struct csched_runqueue_data *rqd;
    1.61      struct csched_vcpu * const scurr = CSCHED_VCPU(current);
    1.62      struct csched_vcpu *snext = NULL;
    1.63      struct task_slice ret;
    1.64 @@ -1010,6 +1012,10 @@ csched_schedule(
    1.65               scurr->vcpu->vcpu_id,
    1.66               now);
    1.67  
    1.68 +    BUG_ON(!cpu_isset(cpu, CSCHED_PRIV(ops)->initialized));
    1.69 +
    1.70 +    rqd = RQD(ops, cpu);
    1.71 +    BUG_ON(!cpu_isset(cpu, rqd->active));
    1.72  
    1.73      /* Protected by runqueue lock */
    1.74  
    1.75 @@ -1166,14 +1172,22 @@ csched_dump(const struct scheduler *ops)
    1.76  {
    1.77      struct list_head *iter_sdom, *iter_svc;
    1.78      struct csched_private *prv = CSCHED_PRIV(ops);
    1.79 -    int loop;
    1.80 +    int i, loop;
    1.81  
    1.82 -    printk("info:\n"
    1.83 -           "\tncpus              = %u\n"
    1.84 +    printk("Active queues: %d\n"
    1.85             "\tdefault-weight     = %d\n",
    1.86 -           prv->ncpus,
    1.87 +           cpus_weight(prv->active_queues),
    1.88             CSCHED_DEFAULT_WEIGHT);
    1.89 +    for_each_cpu_mask(i, prv->active_queues)
    1.90 +    {
    1.91 +        printk("Runqueue %d:\n"
    1.92 +               "\tncpus              = %u\n"
    1.93 +               "\tmax_weight         = %d\n",
    1.94 +               i,
    1.95 +               cpus_weight(prv->rqd[i].active),
    1.96 +               prv->rqd[i].max_weight);
    1.97  
    1.98 +    }
    1.99      /* FIXME: Locking! */
   1.100  
   1.101      printk("Domain info:\n");
   1.102 @@ -1199,16 +1213,82 @@ csched_dump(const struct scheduler *ops)
   1.103      }
   1.104  }
   1.105  
   1.106 -static void
   1.107 -csched_free_pdata(const struct scheduler *ops, void *pcpu, int cpu)
   1.108 +static void activate_runqueue(struct csched_private *prv, int rqi)
   1.109  {
   1.110 -    unsigned long flags;
   1.111 +    struct csched_runqueue_data *rqd;
   1.112 +
   1.113 +    rqd = prv->rqd + rqi;
   1.114 +
   1.115 +    BUG_ON(!cpus_empty(rqd->active));
   1.116 +
   1.117 +    rqd->max_weight = 1;
   1.118 +    rqd->id = rqi;
   1.119 +    INIT_LIST_HEAD(&rqd->svc);
   1.120 +    INIT_LIST_HEAD(&rqd->runq);
   1.121 +    spin_lock_init(&rqd->lock);
   1.122 +
   1.123 +    cpu_set(rqi, prv->active_queues);
   1.124 +}
   1.125 +
   1.126 +static void deactivate_runqueue(struct csched_private *prv, int rqi)
   1.127 +{
   1.128 +    struct csched_runqueue_data *rqd;
   1.129 +
   1.130 +    rqd = prv->rqd + rqi;
   1.131 +
   1.132 +    BUG_ON(!cpus_empty(rqd->active));
   1.133 +    
   1.134 +    rqd->id = -1;
   1.135 +
   1.136 +    cpu_clear(rqi, prv->active_queues);
   1.137 +}
   1.138 +
   1.139 +static void init_pcpu(const struct scheduler *ops, int cpu)
   1.140 +{
   1.141 +    int rqi, old_rqi, flags;
   1.142      struct csched_private *prv = CSCHED_PRIV(ops);
   1.143 +    struct csched_runqueue_data *rqd;
   1.144 +    spinlock_t *old_lock;
   1.145  
   1.146      spin_lock_irqsave(&prv->lock, flags);
   1.147 -    prv->ncpus--;
   1.148 -    cpu_clear(cpu, RQD(ops, cpu)->idle);
   1.149 -    printk("Removing cpu %d to pool (%d total)\n", cpu, prv->ncpus);
   1.150 +
   1.151 +    if ( cpu_isset(cpu, prv->initialized) )
   1.152 +    {
   1.153 +        printk("%s: Strange, cpu %d already initialized!\n", __func__, cpu);
   1.154 +        spin_unlock_irqrestore(&prv->lock, flags);
   1.155 +        return;
   1.156 +    }
   1.157 +
   1.158 +    old_rqi = prv->runq_map[cpu];
   1.159 +
   1.160 +    /* Figure out which runqueue to put it in */
   1.161 +    rqi = 0;
   1.162 +
   1.163 +    rqd=prv->rqd + rqi;
   1.164 +
   1.165 +    printk("Adding cpu %d to runqueue %d\n", cpu, rqi);
   1.166 +    if ( ! cpu_isset(rqi, prv->active_queues) )
   1.167 +    {
   1.168 +        printk(" First cpu on runqueue, activating\n");
   1.169 +        activate_runqueue(prv, rqi);
   1.170 +    }
   1.171 +    
   1.172 +    /* IRQs already disabled */
   1.173 +    old_lock=pcpu_schedule_lock(cpu);
   1.174 +
   1.175 +    /* Move spinlock to new runq lock.  */
   1.176 +    per_cpu(schedule_data, cpu).schedule_lock = &rqd->lock;
   1.177 +
   1.178 +    /* Set the runqueue map */
   1.179 +    prv->runq_map[cpu]=rqi;
   1.180 +    
   1.181 +    cpu_set(cpu, rqd->idle);
   1.182 +    cpu_set(cpu, rqd->active);
   1.183 +
   1.184 +    spin_unlock(old_lock);
   1.185 +
   1.186 +    cpu_set(cpu, prv->initialized);
   1.187 +
   1.188      spin_unlock_irqrestore(&prv->lock, flags);
   1.189  
   1.190      return;
   1.191 @@ -1217,33 +1297,51 @@ csched_free_pdata(const struct scheduler
   1.192  static void *
   1.193  csched_alloc_pdata(const struct scheduler *ops, int cpu)
   1.194  {
   1.195 -    spinlock_t *new_lock;
   1.196 -    spinlock_t *old_lock = per_cpu(schedule_data, cpu).schedule_lock;
   1.197 -    unsigned long flags;
   1.198 -    struct csched_private *prv = CSCHED_PRIV(ops);
   1.199 -
   1.200 -    spin_lock_irqsave(old_lock, flags);
   1.201 -    new_lock = &per_cpu(schedule_data, prv->runq_map[cpu])._lock;
   1.202 -    per_cpu(schedule_data, cpu).schedule_lock = new_lock;
   1.203 -    spin_unlock_irqrestore(old_lock, flags);
   1.204 -
   1.205 -    spin_lock_irqsave(&prv->lock, flags);
   1.206 -    prv->ncpus++;
   1.207 -    cpu_set(cpu, RQD(ops, cpu)->idle);
   1.208 -    printk("Adding cpu %d to pool (%d total)\n", cpu, prv->ncpus);
   1.209 -    spin_unlock_irqrestore(&prv->lock, flags);
   1.210 +    init_pcpu(ops, cpu);
   1.211  
   1.212      return (void *)1;
   1.213  }
   1.214  
   1.215  static void
   1.216 -make_runq_map(struct csched_private *prv)
   1.217 +csched_free_pdata(const struct scheduler *ops, void *pcpu, int cpu)
   1.218  {
   1.219 -    /* FIXME: Read pcpu layout and do this properly */
   1.220 -    prv->runq_count = 1;
   1.221 -    prv->rqd[0].cpu_min = 0;
   1.222 -    prv->rqd[0].cpu_max = NR_CPUS;
   1.223 -    memset(prv->runq_map, 0, sizeof(prv->runq_map));
   1.224 +    unsigned long flags;
   1.225 +    struct csched_private *prv = CSCHED_PRIV(ops);
   1.226 +    struct csched_runqueue_data *rqd;
   1.227 +    int rqi;
   1.228 +
   1.229 +    spin_lock_irqsave(&prv->lock, flags);
   1.230 +
   1.231 +    BUG_ON( !cpu_isset(cpu, prv->initialized));
   1.232 +    
   1.233 +    /* Find the old runqueue and remove this cpu from it */
   1.234 +    rqi = prv->runq_map[cpu];
   1.235 +
   1.236 +    rqd = prv->rqd + rqi;
   1.237 +
   1.238 +    /* No need to save IRQs here, they're already disabled */
   1.239 +    spin_lock(&rqd->lock);
   1.240 +
   1.241 +    BUG_ON(!cpu_isset(cpu, rqd->idle));
   1.242 +
   1.243 +    printk("Removing cpu %d from runqueue %d\n", cpu, rqi);
   1.244 +
   1.245 +    cpu_clear(cpu, rqd->idle);
   1.246 +    cpu_clear(cpu, rqd->active);
   1.247 +
   1.248 +    if ( cpus_empty(rqd->active) )
   1.249 +    {
   1.250 +        printk(" No cpus left on runqueue, disabling\n");
   1.251 +        deactivate_runqueue(prv, rqi);
   1.252 +    }
   1.253 +
   1.254 +    spin_unlock(&rqd->lock);
   1.255 +
   1.256 +    cpu_clear(cpu, prv->initialized);
   1.257 +
   1.258 +    spin_unlock_irqrestore(&prv->lock, flags);
   1.259 +
   1.260 +    return;
   1.261  }
   1.262  
   1.263  static int
   1.264 @@ -1265,18 +1363,11 @@ csched_init(struct scheduler *ops)
   1.265      spin_lock_init(&prv->lock);
   1.266      INIT_LIST_HEAD(&prv->sdom);
   1.267  
   1.268 -    prv->ncpus = 0;
   1.269 -
   1.270 -    make_runq_map(prv);
   1.271 -
   1.272 -    for ( i=0; i<prv->runq_count ; i++ )
   1.273 +    /* But un-initialize all runqueues */
   1.274 +    for ( i=0; i<NR_CPUS; i++)
   1.275      {
   1.276 -        struct csched_runqueue_data *rqd = prv->rqd + i;
   1.277 -
   1.278 -        rqd->max_weight = 1;
   1.279 -        rqd->id = i;
   1.280 -        INIT_LIST_HEAD(&rqd->svc);
   1.281 -        INIT_LIST_HEAD(&rqd->runq);
   1.282 +        prv->runq_map[i] = -1;
   1.283 +        prv->rqd[i].id = -1;
   1.284      }
   1.285  
   1.286      return 0;