Final GPUSync implementation.wip-gpu-rtss12

4 files changed, 49 insertions, 8 deletions

diff --git a/include/litmus/gpu_affinity.h b/include/litmus/gpu_affinity.h
index d4db2003ad86..6b3fb8b28745 100644
--- a/include/litmus/gpu_affinity.h
+++ b/include/litmus/gpu_affinity.h
@@ -43,10 +43,7 @@ static inline lt_t get_gpu_estimate(struct task_struct* t, gpu_migration_dist_t
                 val = _fp_to_integer(t->rt_param.gpu_migration_est[i].est);
         }
 
-        // minimum value is 1 (val is 0 if we haven't run with local affinity yet)
-        // TODO: pick a better default min-value.  1 is too small.  perhaps
-        // task execution time?
-        return ((val > 0) ? val : 1);
+        return ((val > 0) ? val : dist+1);
 }
 
-#endif
\ No newline at end of file
+#endif
diff --git a/include/litmus/nvidia_info.h b/include/litmus/nvidia_info.h
index 580728051d4e..97c9577141db 100644
--- a/include/litmus/nvidia_info.h
+++ b/include/litmus/nvidia_info.h
@@ -12,6 +12,7 @@
 #define NV_MAX_SIMULT_USERS CONFIG_NV_MAX_SIMULT_USERS
 
 int init_nvidia_info(void);
+void shutdown_nvidia_info(void);
 
 int is_nvidia_func(void* func_addr);
 
diff --git a/include/litmus/rt_param.h b/include/litmus/rt_param.h
index 4553521146cc..0198884eab86 100644
--- a/include/litmus/rt_param.h
+++ b/include/litmus/rt_param.h
@@ -26,6 +26,7 @@ static inline int lt_after_eq(lt_t a, lt_t b)
 typedef enum {
         RT_CLASS_HARD,
         RT_CLASS_SOFT,
+        RT_CLASS_SOFT_W_SLIP,
         RT_CLASS_BEST_EFFORT
 } task_class_t;
 
@@ -189,8 +190,8 @@ struct rt_param {
         long unsigned int                       held_gpus;  // bitmap of held GPUs.
 
 #ifdef CONFIG_LITMUS_AFFINITY_LOCKING
-        fp_t    gpu_fb_param_a;
-        fp_t    gpu_fb_param_b;
+        fp_t    gpu_fb_param_a[MIG_LAST+1];
+        fp_t    gpu_fb_param_b[MIG_LAST+1];
 
         gpu_migration_dist_t    gpu_migration;
         int                             last_gpu;
diff --git a/include/litmus/sched_trace.h b/include/litmus/sched_trace.h
index 232c7588d103..b1b71f6c5f0c 100644
--- a/include/litmus/sched_trace.h
+++ b/include/litmus/sched_trace.h
@@ -10,7 +10,8 @@ struct st_trace_header {
         u8      type;           /* Of what type is this record?  */
         u8      cpu;            /* On which CPU was it recorded? */
         u16     pid;            /* PID of the task.              */
-        u32     job;            /* The job sequence number.      */
+        u32 job:24;             /* The job sequence number.      */
+        u8  extra;
 } __attribute__((packed));
 
 #define ST_NAME_LEN 16
@@ -136,6 +137,22 @@ struct st_nv_interrupt_end_data {
         u32 serialNumber;
 } __attribute__((packed));
 
+struct st_prediction_err_data {
+        u64 distance;
+        u64 rel_err;
+} __attribute__((packed));
+
+struct st_migration_data {
+        u64 observed;
+        u64 estimated;
+} __attribute__((packed));
+
+struct migration_info {
+        u64 observed;
+        u64 estimated;
+        u8 distance;
+} __attribute__((packed));
+
 #define DATA(x) struct st_ ## x ## _data x;
 
 typedef enum {
@@ -160,6 +177,9 @@ typedef enum {
         ST_EFF_PRIO_CHANGE,
         ST_NV_INTERRUPT_BEGIN,
         ST_NV_INTERRUPT_END,
+
+        ST_PREDICTION_ERR,
+        ST_MIGRATION,
 } st_event_record_type_t;
 
 struct st_event_record {
@@ -187,6 +207,9 @@ struct st_event_record {
                 DATA(effective_priority_change);
                 DATA(nv_interrupt_begin);
                 DATA(nv_interrupt_end);
+
+                DATA(prediction_err);
+                DATA(migration);
         } data;
 } __attribute__((packed));
 
@@ -259,6 +282,19 @@ feather_callback void do_sched_trace_nv_interrupt_begin(unsigned long id,
 feather_callback void do_sched_trace_nv_interrupt_end(unsigned long id,
                                                                                                 unsigned long unused);
 
+feather_callback void do_sched_trace_prediction_err(unsigned long id,
+                                                                                                          struct task_struct* task,
+                                                                                                          gpu_migration_dist_t* distance,
+                                                                                                          fp_t* rel_err);
+
+
+
+
+
+feather_callback void do_sched_trace_migration(unsigned long id,
+                                                                                          struct task_struct* task,
+                                                                                          struct migration_info* mig_info);
+
 
 /* returns true if we're tracing an interrupt on current CPU */
 /* int is_interrupt_tracing_active(void); */
@@ -331,6 +367,12 @@ feather_callback void do_sched_trace_nv_interrupt_end(unsigned long id,
 #define sched_trace_nv_interrupt_end(d) \
         SCHED_TRACE(SCHED_TRACE_BASE_ID + 19, do_sched_trace_nv_interrupt_end, d)
 
+#define sched_trace_prediction_err(t, dist, rel_err) \
+        SCHED_TRACE3(SCHED_TRACE_BASE_ID + 20, do_sched_trace_prediction_err, t, dist, rel_err)
+
+#define sched_trace_migration(t, mig_info) \
+        SCHED_TRACE2(SCHED_TRACE_BASE_ID + 21, do_sched_trace_migration, t, mig_info)
+
 #define sched_trace_quantum_boundary() /* NOT IMPLEMENTED */
 
 #endif /* __KERNEL__ */