prov/efa: Fix error handling in efa_rdm_cq_poll_ibv_cq

[shijin-aws]

When ibv_start_poll or ibv_next_poll return error, it may read a cqe from a destroyed EP (QP). In this case we shouldn't call efa_rdm_cq_get_prov_errno, which may use a freed EP from the pkt_entry (wr_id) and cause segmentation fault.

[darrylabbate]

Is it possible to capture this in a condition inside efa_rdm_cq_get_prov_errno() instead?

[shijin-aws]

Is it possible to capture this in a condition inside efa_rdm_cq_get_prov_errno() instead?

Unless we want to make the return code (say err) of ibv_start_poll as an input for efa_rdm_cq_get_prov_errno(). When err is EINVAL, that means the polled pkt entry is from a destroyed QP.

There is no good way to tell that from pkt_entry (wr_id) itself.

So I prefer to just call ibv_wc_read_vendor_err directly to avoid these troubles

[shijin-aws]

bot:aws:retest