Memory reclaiming bug fix: while reclaiming memory, we should not hold the lock of the header describing the entire state of the handle while other handles are inspected to find available memory.

src/datawizard/data_request.c

@@ -304,8 +304,11 @@ static int starpu_handle_data_request(starpu_data_request_t r, unsigned may_allo
 	STARPU_ASSERT(!(r_mode & STARPU_R) || src_replicate->allocated);
 	STARPU_ASSERT(!(r_mode & STARPU_R) || src_replicate->refcnt);
 
+	_starpu_spin_unlock(&r->lock);
+
 	/* perform the transfer */
 	/* the header of the data must be locked by the worker that submitted the request */
+
 	r->retval = _starpu_driver_copy_data_1_to_1(handle, src_replicate,
 			dst_replicate, !(r_mode & STARPU_R), r, may_alloc);
 
@@ -313,9 +316,7 @@ static int starpu_handle_data_request(starpu_data_request_t r, unsigned may_allo
 	{
 		/* If there was not enough memory, we will try to redo the
 		 * request later. */
-		_starpu_spin_unlock(&r->lock);
 		_starpu_spin_unlock(&handle->header_lock);
-
 		return -ENOMEM;
 	}
 
@@ -325,7 +326,6 @@ static int starpu_handle_data_request(starpu_data_request_t r, unsigned may_allo
 		 * immediatly. We will handle the completion of the request
 		 * asynchronously. The request is put in the list of "pending"
 		 * requests in the meantime. */
-		_starpu_spin_unlock(&r->lock);
 		_starpu_spin_unlock(&handle->header_lock);
 
 		PTHREAD_MUTEX_LOCK(&data_requests_pending_list_mutex[r->handling_node]);

src/datawizard/memalloc.c

@@ -693,10 +693,19 @@ static ssize_t _starpu_allocate_interface(starpu_data_handle handle, struct star
 		allocated_memory = handle->ops->allocate_data_on_node(replicate->interface, dst_node);
 		STARPU_TRACE_END_ALLOC(dst_node);
 
-		if (allocated_memory == -ENOMEM) {
+		if (allocated_memory == -ENOMEM)
+		{
+			replicate->refcnt++;
+			_starpu_spin_unlock(&handle->header_lock);
+
 			STARPU_TRACE_START_MEMRECLAIM(dst_node);
 			reclaim_memory_generic(dst_node, 0);
 			STARPU_TRACE_END_MEMRECLAIM(dst_node);
+
+		        while (_starpu_spin_trylock(&handle->header_lock))
+		                _starpu_datawizard_progress(_starpu_get_local_memory_node(), 0);
+		
+			replicate->refcnt--;
 		}
 		
 	} while((allocated_memory == -ENOMEM) && attempts++ < 2);