Adding a check to prevent writeback cache with allocation-on-fill.

[git-p4: depot-paths = "//depot/gpgpu_sim_research/fermi/distribution/": change = 15277]

2 files changed, 19 insertions, 0 deletions

diff --git a/CHANGES b/CHANGES
index 7b80e58..4bba86b 100644
--- a/CHANGES
+++ b/CHANGES
@@ -11,6 +11,12 @@ Version 3.2.0+edits (development branch) versus 3.2.0
   have the same warp_id but different dynamic_warp_ids.
 - Added extensions to the cache class hierarchy to allow for the use of custom tag_array objects.
 - Added some additional const-correctness
+- Added a check in cache_config to prevent configuration that specifies a
+  writeback cache with allocation-on-fill policy.  The current implementation
+  of the allocation-on-fill policy assumes a non-writeback cache and never
+  generates any writeback traffic.  Even if the writeback traffic is generated,
+  the configuration (writeback cache + allocation-on-fill) will inevitably lead
+  to deadlock.  
 - Bug Fixes:
     - Applied patch from Kito Cheng to update libopencl for checking NULL error
       code pointer. 
diff --git a/src/gpgpu-sim/gpu-cache.h b/src/gpgpu-sim/gpu-cache.h
index fb59c01..6dce6b4 100644
--- a/src/gpgpu-sim/gpu-cache.h
+++ b/src/gpgpu-sim/gpu-cache.h
@@ -179,6 +179,19 @@ public:
         default: exit_parse_error();
         }
 
+        // detect invalid configuration 
+        if (m_alloc_policy == ON_FILL and m_write_policy == WRITE_BACK) {
+            // A writeback cache with allocate-on-fill policy will inevitably lead to deadlock:  
+            // The deadlock happens when an incoming cache-fill evicts a dirty
+            // line, generating a writeback request.  If the memory subsystem
+            // is congested, the interconnection network may not have
+            // sufficient buffer for the writeback request.  This stalls the
+            // incoming cache-fill.  The stall may propagate through the memory
+            // subsystem back to the output port of the same core, creating a
+            // deadlock where the wrtieback request and the incoming cache-fill
+            // are stalling each other.  
+            assert(0 && "Invalid cache configuration: Writeback cache cannot allocate new line on fill. "); 
+        }
     }
     bool disabled() const { return m_disabled;}
     unsigned get_line_sz() const