Final GPUSync implementation.wip-gpu-rtss12

1 files changed, 31 insertions, 8 deletions

diff --git a/litmus/ikglp_lock.c b/litmus/ikglp_lock.c
index 023443014d4b..83b708ab85cb 100644
--- a/litmus/ikglp_lock.c
+++ b/litmus/ikglp_lock.c
@@ -1346,6 +1346,10 @@ int ikglp_unlock(struct litmus_lock* l)
 #ifdef CONFIG_LITMUS_AFFINITY_LOCKING
                 if(sem->aff_obs && sem->aff_obs->relax_max_fifo_len) {
                         fq_of_new_on_fq = sem->aff_obs->ops->advise_enqueue(sem->aff_obs, new_on_fq);
+                        if(fq_of_new_on_fq->count == 0) {
+                                // ignore it?
+//                              fq_of_new_on_fq = fq;
+                        }
                 }
                 else {
                         fq_of_new_on_fq = fq;
@@ -1383,6 +1387,10 @@ int ikglp_unlock(struct litmus_lock* l)
 #ifdef CONFIG_LITMUS_AFFINITY_LOCKING
                 if(sem->aff_obs && sem->aff_obs->relax_max_fifo_len) {
                         fq_of_new_on_fq = sem->aff_obs->ops->advise_enqueue(sem->aff_obs, new_on_fq);
+                        if(fq_of_new_on_fq->count == 0) {
+                                // ignore it?
+//                              fq_of_new_on_fq = fq;
+                        }
                 }
                 else {
                         fq_of_new_on_fq = fq;
@@ -1409,6 +1417,10 @@ int ikglp_unlock(struct litmus_lock* l)
 #ifdef CONFIG_LITMUS_AFFINITY_LOCKING
                 if(sem->aff_obs && sem->aff_obs->relax_max_fifo_len) {
                         fq_of_new_on_fq = sem->aff_obs->ops->advise_enqueue(sem->aff_obs, new_on_fq);
+                        if(fq_of_new_on_fq->count == 0) {
+                                // ignore it?
+//                              fq_of_new_on_fq = fq;
+                        }
                 }
                 else {
                         fq_of_new_on_fq = fq;
@@ -1569,7 +1581,7 @@ int ikglp_unlock(struct litmus_lock* l)
                 }
         }
 
-
+wake_kludge:
         if(waitqueue_active(&fq->wait))
         {
                 wait_queue_t *wait = list_entry(fq->wait.task_list.next, wait_queue_t, task_list);
@@ -1672,6 +1684,16 @@ int ikglp_unlock(struct litmus_lock* l)
                 // wake up the new resource holder!
                 wake_up_process(next);
         }
+        if(fq_of_new_on_fq && fq_of_new_on_fq != fq && fq_of_new_on_fq->count == 1) {
+                // The guy we promoted when to an empty FQ. (Why didn't stealing pick this up?)
+                // Wake up the new guy too.
+
+                BUG_ON(fq_of_new_on_fq->owner != NULL);
+
+                fq = fq_of_new_on_fq;
+                fq_of_new_on_fq = NULL;
+                goto wake_kludge;
+        }
 
         unlock_fine_irqrestore(&sem->lock, flags);
         unlock_global_irqrestore(dgl_lock, flags);
@@ -1917,7 +1939,7 @@ static struct affinity_observer* ikglp_aff_obs_new(struct affinity_observer_ops*
         if(aff_args.nr_simult_users > NV_MAX_SIMULT_USERS) {
                 TRACE_CUR("System does not support #simult_users > %d. %d requested.\n",
                                   NV_MAX_SIMULT_USERS, aff_args.nr_simult_users);
-                return(NULL);
+//              return(NULL);
         }
 
         ikglp_aff = kmalloc(sizeof(*ikglp_aff), GFP_KERNEL);
@@ -2600,8 +2622,8 @@ void gpu_ikglp_notify_acquired(struct ikglp_affinity* aff,
 
         tsk_rt(t)->gpu_migration = gpu_migration_distance(tsk_rt(t)->last_gpu, gpu);  // record the type of migration
 
-        TRACE_CUR("%s/%d acquired gpu %d.  migration type = %d\n",
-                          t->comm, t->pid, gpu, tsk_rt(t)->gpu_migration);
+        TRACE_CUR("%s/%d acquired gpu %d (prev = %d).  migration type = %d\n",
+                          t->comm, t->pid, gpu, tsk_rt(t)->last_gpu, tsk_rt(t)->gpu_migration);
 
         // count the number or resource holders
         ++(*(aff->q_info[replica].nr_cur_users));
@@ -2626,8 +2648,6 @@ void gpu_ikglp_notify_freed(struct ikglp_affinity* aff,
 
         est_time = get_gpu_estimate(t, gpu_migration_distance(tsk_rt(t)->last_gpu, gpu));
 
-        tsk_rt(t)->last_gpu = gpu;
-
         // count the number or resource holders
         --(*(aff->q_info[replica].nr_cur_users));
 
@@ -2636,12 +2656,15 @@ void gpu_ikglp_notify_freed(struct ikglp_affinity* aff,
         // update estimates
         update_gpu_estimate(t, get_gpu_time(t));
 
-        TRACE_CUR("%s/%d freed gpu %d.  actual time was %llu.  "
+        TRACE_CUR("%s/%d freed gpu %d (prev = %d).  mig type = %d.  actual time was %llu.  "
                           "estimated was %llu.  diff is %d\n",
-                          t->comm, t->pid, gpu,
+                          t->comm, t->pid, gpu, tsk_rt(t)->last_gpu,
+                          tsk_rt(t)->gpu_migration,
                           get_gpu_time(t),
                           est_time,
                           (long long)get_gpu_time(t) - (long long)est_time);
+
+        tsk_rt(t)->last_gpu = gpu;
 }
 
 struct ikglp_affinity_ops gpu_ikglp_affinity =