> On 6 Apr 2021, at 17:28, Jason Gunthorpe <jgg@xxxxxxxxxx> wrote: > > On Tue, Apr 06, 2021 at 03:01:41PM +0000, Haakon Bugge wrote: > >> Running rdma_server (RC transport) on an IB device that supports >> abi_ver 4, I see two ib_modify_qp() calls to the INIT state with >> mask 0x39 (STATE, ACCESS_FLAGS, PKEY_INDEX, and PORT) and of course, >> state == 1 (IB_QPS_INIT). >> >> Presumable, from ucma_init_conn_qp(), then again in ucma_modify_qp_rtr(). >> >> Slip of the keyboard? > > The control flow is complicated here and I worry parts of this are ABI > now. Would it be OK to return from rdma_create_qp_ex() with an > incompletely initialized QP? I don't know. Two transitions to INIT with same mask and data is idempotent. Should be possible to remove the second for this case. Let me discuss internally with Mark and get back. Thxs, Håkon
