Move mutex unlock after the cudaEventRecord() in DeviceFree() (#240)

This commit fixes a race for the CUDA event between DeviceFree() and
DeviceAllocate(). If the mutex is unlocked before the
cudaEventRecord(), there is a short period of time when
- the memory block is already in the free list (cached_blocks), and
- the CUDA event status is not yet cudaErrorNotReady
and the DeviceAllocate() may consider that memory block to be free to
be used for another CUDA stream.

Author: makortel

HeterogeneousCore/CUDAServices/src/CachingDeviceAllocator.h

@@ -586,9 +586,6 @@ struct CachingDeviceAllocator
             }
         }
 
-        // Unlock
-        mutex.Unlock();
-
         // First set to specified device (entrypoint may not be set)
         if (device != entrypoint_device)
         {
@@ -601,7 +598,11 @@ struct CachingDeviceAllocator
             // Insert the ready event in the associated stream (must have current device set properly)
             if (CubDebug(error = cudaEventRecord(search_key.ready_event, search_key.associated_stream))) return error;
         }
-        else
+
+        // Unlock
+        mutex.Unlock();
+
+        if (!recached)
         {
             // Free the allocation from the runtime and cleanup the event.
             if (CubDebug(error = cudaFree(d_ptr))) return error;