[openib-general] Re: CMA deadlock

[Michael S. Tsirkin]

Quoting r. Sean Hefty <sean.hefty at intel.com>:
> >We seem to be calling rdma_destroy_id, which seems to be calling
> >ib_destroy_cm_id directly. No?
> 
> We're setting cm_id.ib to NULL, which prevents rdma_destroy_id() from calling
> ib_destroy_cm_id().  The handler will return a non-zero value to destroy
> cm_id.ib.

I see. I would be happier with a variant of rdma_destroy_id with an explicit
flag destroy_cm_id.

The current approach seems racy: what if e.g. listen call is in progress?  It
looks at id_priv->cm_id.ib which seemingly might change under its feet.



-- 
Michael S. Tsirkin
Staff Engineer, Mellanox Technologies