[ofa-general] Re: IPoIB path caching

[Or Gerlitz]

Or Gerlitz wrote:
> Michael S. Tsirkin wrote:

>>> And ARP table aging gives a way to recover
>>> from stale cached data, eventually at least.

>> Does it?

>> $ grep path_list drivers/infiniband/ulp/ipoib/*c
>> drivers/infiniband/ulp/ipoib/ipoib_main.c: list_add_tail(&path->list, &priv->path_list);
>> drivers/infiniband/ulp/ipoib/ipoib_main.c: list_splice(&priv->path_list, &remove_list);
>> drivers/infiniband/ulp/ipoib/ipoib_main.c: INIT_LIST_HEAD(&priv->path_list);
>> drivers/infiniband/ulp/ipoib/ipoib_main.c: INIT_LIST_HEAD(&priv->path_list);

>> In other words we add paths to ipoib specific cache, but we never seem
>> to *remove* individual paths from cache - we only know how to do
>> full cache invalidates on events such as port state change.

> this seems like a bug, if the stack decided to delete OR change a 
> neighbour, the path associated with it must not be re-used to create the 
> address handle or to establish the connection, same for multicast 
> neighbours.

Roland,

Can you provide your take here?

Do you agree that using cached IB L2 info where the net stack wants to 
renew its IPoIB L2 (which is IB L3 && L4) info is a bug?

Or.