On Fri, Nov 22, 2019 at 02:58:44AM +0000, Zengtao (B) wrote: > > From: linux-rdma-owner@xxxxxxxxxxxxxxx > > [mailto:linux-rdma-owner@xxxxxxxxxxxxxxx] On Behalf Of Weihang Li > > Sent: Thursday, November 21, 2019 9:19 AM > > To: jgg@xxxxxxxx; leon@xxxxxxxxxx > > Cc: dledford@xxxxxxxxxx; linux-rdma@xxxxxxxxxxxxxxx; Linuxarm > > Subject: [PATCH rdma-core 1/7] libhns: Fix calculation errors with ilog32() > > > > Current calculation results using ilog32() is larger than expected, which > > will lead to driver broken. The following is the log when QP creations > > fails: > > > > [ 81.294844] hns3 0000:7d:00.0 hns_0: check SQ size error! > > [ 81.294848] hns3 0000:7d:00.0 hns_0: check SQ size error! > > [ 81.300225] hns3 0000:7d:00.0 hns_0: Sanity check sq size failed > > [ 81.300227] hns3 0000:7d:00.0: hns_roce_set_user_sq_size error for > > create qp > > [ 81.305602] hns3 0000:7d:00.0 hns_0: Sanity check sq size failed > > [ 81.305603] hns3 0000:7d:00.0: hns_roce_set_user_sq_size error for > > create qp > > [ 81.311589] hns3 0000:7d:00.0 hns_0: Create RC QP 0x000000 > > failed(-22) > > [ 81.318603] hns3 0000:7d:00.0 hns_0: Create RC QP 0x000000 > > failed(-22) > > > > Fixes: b6cd213b276f ("libhns: Refactor for creating qp") > > Signed-off-by: Weihang Li <liweihang@xxxxxxxxxxxxx> > > providers/hns/hns_roce_u_verbs.c | 11 ++++++----- > > 1 file changed, 6 insertions(+), 5 deletions(-) > > > > diff --git a/providers/hns/hns_roce_u_verbs.c > > b/providers/hns/hns_roce_u_verbs.c > > index 9d222c0..bd5060d 100644 > > +++ b/providers/hns/hns_roce_u_verbs.c > > @@ -645,7 +645,8 @@ static int hns_roce_calc_qp_buff_size(struct > > ibv_pd *pd, struct ibv_qp_cap *cap, > > int page_size = to_hr_dev(pd->context->device)->page_size; > > > > if (to_hr_dev(pd->context->device)->hw_version == > > HNS_ROCE_HW_VER1) { > > - qp->rq.wqe_shift = ilog32(sizeof(struct hns_roce_rc_rq_wqe)); > > + qp->rq.wqe_shift = > > + ilog32(sizeof(struct hns_roce_rc_rq_wqe)) - 1; > > > > qp->buf_size = align((qp->sq.wqe_cnt << qp->sq.wqe_shift), > > page_size) + > > @@ -662,7 +663,7 @@ static int hns_roce_calc_qp_buff_size(struct > > ibv_pd *pd, struct ibv_qp_cap *cap, > > } else { > > unsigned int rqwqe_size = HNS_ROCE_SGE_SIZE * > > cap->max_recv_sge; > > > > - qp->rq.wqe_shift = ilog32(rqwqe_size); > > + qp->rq.wqe_shift = ilog32(rqwqe_size) - 1; > > > > if (qp->sq.max_gs > HNS_ROCE_SGE_IN_WQE || type == > > IBV_QPT_UD) > > qp->sge.sge_shift = HNS_ROCE_SGE_SHIFT; > > @@ -747,8 +748,8 @@ static void hns_roce_set_qp_params(struct > > ibv_pd *pd, > > qp->rq.wqe_cnt = > > roundup_pow_of_two(attr->cap.max_recv_wr); > > } > > > > - qp->sq.wqe_shift = ilog32(sizeof(struct hns_roce_rc_send_wqe)); > > - qp->sq.shift = ilog32(qp->sq.wqe_cnt); > > + qp->sq.wqe_shift = ilog32(sizeof(struct hns_roce_rc_send_wqe)) - 1; > > + qp->sq.shift = ilog32(qp->sq.wqe_cnt) - 1; > > One suggestion, it's better to introduce a new micro instead of ilog32(x) -1. Is the -1 even correct? I would have guessed something called shift wants to be: shift = ilog32(qp->sq.wqe_cnt - 1) Such that 1 << shift == wqe_cnt When wqe_cnt is a power of two. 1 << shift > wqe_cnt When wqe_cnt is not a power of two. Jason
