Cluster assignment of nv klmirqd threads.

5 files changed, 62 insertions, 7 deletions

diff --git a/include/litmus/sched_plugin.h b/include/litmus/sched_plugin.h
index a13d1a2992fe..cfa218504d75 100644
--- a/include/litmus/sched_plugin.h
+++ b/include/litmus/sched_plugin.h
@@ -102,6 +102,10 @@ typedef int (*__higher_prio_t)(struct task_struct* a, comparison_mode_t a_mod,
                                                           struct task_struct* b, comparison_mode_t b_mod);
 #endif
 
+#if defined(CONFIG_LITMUS_NVIDIA) && defined(CONFIG_LITMUS_SOFTIRQD)
+typedef int (*default_cpu_for_gpu_t)(int gpu, int num_gpus);
+#endif
+
 
 /********************* sys call backends  ********************/
 /* This function causes the caller to sleep until the next release */
@@ -165,6 +169,10 @@ struct sched_plugin {
         change_prio_pai_tasklet_t       change_prio_pai_tasklet;
         run_tasklets_t                          run_tasklets;
 #endif
+
+#if defined(CONFIG_LITMUS_NVIDIA) && defined(CONFIG_LITMUS_SOFTIRQD)
+        default_cpu_for_gpu_t           map_gpu_to_cpu;
+#endif
 } __attribute__ ((__aligned__(SMP_CACHE_BYTES)));
 
 
diff --git a/litmus/nvidia_info.c b/litmus/nvidia_info.c
index b29f4d3f0dac..0b39dcc84115 100644
--- a/litmus/nvidia_info.c
+++ b/litmus/nvidia_info.c
@@ -449,13 +449,16 @@ static int init_nv_device_reg(void)
                 INIT_BINHEAP_HANDLE(&NV_DEVICE_REG[i].owners, gpu_owner_max_priority_order);
 
 #ifdef CONFIG_LITMUS_SOFTIRQD
-                // TODO: Make thread spawning this a litmus plugin call.
-                NV_DEVICE_REG[i].callback.func = nvidia_klmirqd_cb;
-                NV_DEVICE_REG[i].callback.arg = (void*)(long long)(i);
-                mb();
+                {
+                        int default_cpu = litmus->map_gpu_to_cpu(i, NV_DEVICE_NUM);
+
+                        NV_DEVICE_REG[i].callback.func = nvidia_klmirqd_cb;
+                        NV_DEVICE_REG[i].callback.arg = (void*)(long long)(i);
+                        mb();
 
-                if(launch_klmirqd_thread(0, &NV_DEVICE_REG[i].callback) != 0) {
-                        TRACE("Failed to create klmirqd thread for GPU %d\n", i);
+                        if(launch_klmirqd_thread(default_cpu, &NV_DEVICE_REG[i].callback) != 0) {
+                                TRACE("Failed to create klmirqd thread for GPU %d\n", i);
+                        }
                 }
 #endif
         }
diff --git a/litmus/sched_cedf.c b/litmus/sched_cedf.c
index 84aafca78cde..35ea1544ce69 100644
--- a/litmus/sched_cedf.c
+++ b/litmus/sched_cedf.c
@@ -1652,6 +1652,25 @@ static void cleanup_cedf(void)
         }
 }
 
+#if defined(CONFIG_LITMUS_NVIDIA) && defined(CONFIG_LITMUS_SOFTIRQD)
+static int cedf_map_gpu_to_cpu(int gpu, int num_gpus)
+{
+        /* TODO: Some sort of smart clustering on the PCIe bus topology */
+        int num_gpu_clusters = num_clusters;
+        unsigned int gpu_cluster_size = num_gpus / num_gpu_clusters;
+        int cpu_cluster = gpu / gpu_cluster_size;
+        int default_cpu = cedf[cpu_cluster].cpus[0]->cpu;  // first CPU in given cluster
+
+        if(num_gpus % num_gpu_clusters != 0) {
+                TRACE("GPU clusters are of non-uniform size!\n");
+        }
+
+        TRACE("CPU %d is default for GPU %d interrupt threads.\n", default_cpu, gpu);
+
+        return default_cpu;
+}
+#endif
+
 static long cedf_activate_plugin(void)
 {
         int i, j, cpu, ccpu, cpu_count;
@@ -1826,6 +1845,9 @@ static struct sched_plugin cedf_plugin __cacheline_aligned_in_smp = {
         .change_prio_pai_tasklet = cedf_change_prio_pai_tasklet,
         .run_tasklets = cedf_run_tasklets,
 #endif
+#if defined(CONFIG_LITMUS_NVIDIA) && defined(CONFIG_LITMUS_SOFTIRQD)
+        .map_gpu_to_cpu = cedf_map_gpu_to_cpu,
+#endif
 };
 
 static struct proc_dir_entry *cluster_file = NULL, *cedf_dir = NULL;
diff --git a/litmus/sched_gsn_edf.c b/litmus/sched_gsn_edf.c
index f27c104ea027..1b5d8d73dc16 100644
--- a/litmus/sched_gsn_edf.c
+++ b/litmus/sched_gsn_edf.c
@@ -1835,7 +1835,12 @@ UNSUPPORTED_AFF_OBS:
 #endif
 
 
-
+#if defined(CONFIG_LITMUS_NVIDIA) && defined(CONFIG_LITMUS_SOFTIRQD)
+static int gsnedf_map_gpu_to_cpu(int gpu, int num_gpus)
+{
+        return 0;  // CPU_0 is default in all cases.
+}
+#endif
 
 
 static long gsnedf_activate_plugin(void)
@@ -1918,6 +1923,9 @@ static struct sched_plugin gsn_edf_plugin __cacheline_aligned_in_smp = {
         .change_prio_pai_tasklet = gsnedf_change_prio_pai_tasklet,
         .run_tasklets = gsnedf_run_tasklets,
 #endif
+#if defined(CONFIG_LITMUS_NVIDIA) && defined(CONFIG_LITMUS_SOFTIRQD)
+        .map_gpu_to_cpu = gsnedf_map_gpu_to_cpu,
+#endif
 };
 
 
diff --git a/litmus/sched_plugin.c b/litmus/sched_plugin.c
index 30c216fd6fdc..ea89f5fedcab 100644
--- a/litmus/sched_plugin.c
+++ b/litmus/sched_plugin.c
@@ -222,6 +222,13 @@ static long litmus_dummy_allocate_aff_obs(struct affinity_observer **aff_obs,
 }
 #endif
 
+#if defined(CONFIG_LITMUS_NVIDIA) && defined(CONFIG_LITMUS_SOFTIRQD)
+static int litmus_dummy_map_gpu_to_cpu(int gpu, int num_gpus)
+{
+        return 0;
+}
+#endif
+
 
 /* The default scheduler plugin. It doesn't do anything and lets Linux do its
  * job.
@@ -262,6 +269,9 @@ struct sched_plugin linux_sched_plugin = {
 #ifdef CONFIG_LITMUS_AFFINITY_LOCKING
         .allocate_aff_obs = litmus_dummy_allocate_aff_obs,
 #endif
+#if defined(CONFIG_LITMUS_NVIDIA) && defined(CONFIG_LITMUS_SOFTIRQD)
+        .map_gpu_to_cpu = litmus_dummy_map_gpu_to_cpu,
+#endif
 
         .admit_task = litmus_dummy_admit_task
 };
@@ -322,6 +332,10 @@ int register_sched_plugin(struct sched_plugin* plugin)
 #ifdef CONFIG_LITMUS_AFFINITY_LOCKING
         CHECK(allocate_aff_obs);
 #endif
+#if defined(CONFIG_LITMUS_NVIDIA) && defined(CONFIG_LITMUS_SOFTIRQD)
+        CHECK(map_gpu_to_cpu);
+#endif
+
         CHECK(admit_task);
 
         if (!plugin->release_at)