On Fri, Feb 05, 2021 at 03:23:12PM +0000, Saleem, Shiraz wrote:

quoted

Subject: Re: [PATCH 04/22] ice: Register auxiliary device to provide RDMA
 

quoted

@@ -1254,20 +1282,37 @@ int ice_init_peer_devices(struct ice_pf *pf)
 		 * |--> iidc_peer_obj
 		 * |--> *ice_peer_drv_int
 		 *
+		 * iidc_auxiliary_object (container_of parent for adev)
+		 * |--> auxiliary_device
+		 * |--> *iidc_peer_obj (pointer from internal struct)
+		 *
 		 * ice_peer_drv_int (internal only peer_drv struct)
 		 */
 		peer_obj_int = kzalloc(sizeof(*peer_obj_int), GFP_KERNEL);
-		if (!peer_obj_int)
+		if (!peer_obj_int) {
+			ida_simple_remove(&ice_peer_ida, id);
 			return -ENOMEM;
+		}

Why is this allocated memory with a lifetime different from the aux device?

This ice_peer_obj_int is the PCI driver internal only info about the peer_obj (not exposed externally)
like the state machine, per PF. But Dave is re-writing all of this with the feedback about getting rid
of state machine, and this peer_obj_int will likely be culled.

I think what we will end up with is an iidc_peer_obj per PF which is
exported to aux driver with lifetime as described below.

I wouldn't call it 'peer' anything, this object represents the
programming API of the PCI device. The object and the API should be
understandable from the header files

A good design will have netdev also sit on this programming API, even
if it doesn't have the aux device. mlx5 used mlx5_core as the name,
I'd suggest something similar.

Jason