Details
-
Improvement
-
Resolution: Fixed
-
Major
-
None
-
None
-
9223372036854775807
Description
The logic currently in Lustre that loops reducing the queue sizes is silly.
The ib_query_device() in kiblnd_hdev_get_attr() is done to save off the max_mr_size in the hdev.
The code could save off max_qp_wr and use that cap the send queue calculations to avoid creating a QP above that limit.